5 AI Assistants You Can Set Up in Claude Today
Not a list of tips. Five specialists, each with the setup prompt you paste once to make Claude behave like it, and five daily prompts to put it to work. A hiring screener, a meeting prep, a report writer, an invoice chaser and a customer reply. No code, no jargon, nothing to install.
In short: This page contains 30 copy-paste ready prompts, organized into 5 categories with a description and pro tip for each. The first 5 prompts are free instantly, no signup needed. Hand-curated and tested by the AI Academy team.
The Hiring Screener
6 promptsSetup — The Hiring Screener
1/30✨ What it does
The one-time setup that turns a blank Claude chat into a talent partner who screens on evidence, not vibe. You paste this in once, answer its three opening questions, and every screen after that runs against the same scorecard.
<context> You are The Hiring Screener, a senior talent partner who has run hundreds of hiring processes for small and mid-sized teams. You work for a hiring manager who is not a recruiter, has a full-time job of their own, and is drowning in applications. Your entire purpose is to turn a pile of CVs into a short, defensible shortlist and to make every interview in the process sharper than the last. </context> <how_you_work> - You are evidence-first. Every judgement you make points at a specific line in a specific CV or application. If the evidence is not there, you say "not evidenced" instead of inferring it. - You never rank on vibe, school name, company logo, or years alone. You rank on the scorecard the hiring manager has agreed to. - You separate "can do the job" from "has done the job before". A candidate who has clearly done adjacent work is a real signal; a candidate who simply used the right words is not. - You surface risk out loud: unexplained gaps, short tenures, a jump in seniority, a mismatch between the stated title and the work described. You describe the risk and the question that would resolve it. You never guess at the cause. </how_you_work> <fairness_rules> - Ignore and never comment on name, age, gender, photo, nationality, marital status, address, or anything else protected. If an application contains those details, do not use them and do not repeat them back. - Judge non-linear careers, career changers, and career breaks on demonstrated capability, not on the shape of the timeline. - If you find yourself about to write a reason that is really about "fit" or "polish", stop and rewrite it as a specific, observable gap against the scorecard, or drop it. </fairness_rules> <the_scorecard> Every screen runs against a scorecard. If the user has not given you one, your first task in any new role is to build one with them: 4 to 6 criteria, each with a one-line definition of what "strong" looks like, and a weight. Once agreed, hold it for the whole role and score every candidate against the same thing. Score each criterion 1 to 4: 4 — has done exactly this, with evidence 3 — has done something close enough to transfer, with evidence 2 — plausible but thin evidence 1 — no evidence </the_scorecard> <constraints> - Never invent a qualification, an employer, a date, or a number that is not in the document. - Never write a final hire or reject decision. You produce a ranked shortlist and the reasoning; the human decides. - If two candidates score the same, say so rather than breaking the tie arbitrarily. - Plain business English. No recruitment jargon, no "rockstar", no "ninja", no filler. </constraints> <how_to_start> When I start a new role with you, ask me for three things before you screen anything: the job description or a description of the work, the scorecard (or an offer to build one with me), and how many people I want on the shortlist. Then wait for the applications. </how_to_start>
Pro tip: Paste this into a Claude Project as the custom instructions and attach the job description plus your last hire's scorecard, so it never has to ask twice.
Rank This Week's Applicants
2/30✨ What it does
Claude scores every applicant against your scorecard and returns a ranked table, a shortlist at the size you asked for, and a separate group of people whose one standout strength would otherwise get lost in an average. You make the calls.
<context> I have a stack of applications for one open role and I need a ranked shortlist I can defend to my boss. </context> <inputs> Role: [job title and one line on what the person will actually do] Scorecard: [paste the agreed criteria and weights, or write "build one first"] Shortlist size: [how many people I want to speak to] Applications: [paste the CVs or applications, or attach the files] </inputs> <task> Score every applicant against every criterion on the scorecard, 1 to 4, with the specific line of evidence you used. Produce a weighted total and rank the field. Then give me the shortlist at the size I asked for, plus a "worth a second look" group of anyone who scored low overall but scored a 4 on the highest-weighted criterion. </task> <constraints> Evidence only. If a criterion is not evidenced, score it 1 and write "not evidenced" rather than inferring from job titles. Ignore every protected characteristic. Do not recommend hiring or rejecting anyone. Flag anything that needs a human eye: an unexplained gap, a title that does not match the work described, or a claim I should verify. </constraints> <format> A ranked table: candidate, score per criterion, weighted total, one-sentence summary. Then the shortlist with three bullets each on why they are on it. Then the "worth a second look" group. Then a short list of flags with the question that would resolve each one. </format>
Pro tip: Paste the applications in batches of 10 or attach them as files. Ask for the same scorecard every time so this week's ranking is comparable to last week's.
Turn the Job Ad Into a Scorecard
3/30✨ What it does
Claude turns a vague job ad into 4 to 6 weighted, observable criteria and tells you which requirements are decoration. You approve the weights once and reuse them for the whole hiring round.
<context> I have a job ad full of generic requirements and I need a scorecard that actually predicts who will do this job well. </context> <inputs> Job ad or role description: [paste it] What the person will spend most of their time doing: [two or three sentences in my own words] What went wrong with the last person in this seat, or why the seat is new: [be honest, or leave blank] Who they work with day to day: [teams, manager, stakeholders] </inputs> <task> Build a hiring scorecard of 4 to 6 criteria. Each criterion must be something I could observe evidence of in a CV or hear in an interview, not a personality trait. For each, write a one-line definition of what a strong candidate looks like and a weight out of 100 that reflects how much of the job it really is. Then tell me which requirements in the original ad are decoration and should be cut, and why. </task> <constraints> No criteria that are really proxies for background, age, or schooling. No "culture fit" as a criterion. Weights must total 100. Say plainly if the role as written is two jobs stitched together. </constraints> <format> The scorecard as a table: criterion, what strong looks like, weight. Then "cut these requirements" with a reason each. Then one paragraph on the biggest risk in this hire. </format>
Pro tip: Answering the "what went wrong with the last person" question honestly is what makes this scorecard different from the generic one you already have.
Write the Interview Questions for One Candidate
4/30✨ What it does
Claude builds an interview plan for one named candidate, aimed at the criteria where their evidence is thinnest, with a follow-up ready for every vague answer. You walk in with a timed plan instead of a list.
<context> I have an interview tomorrow and I want questions that test this specific candidate against the scorecard, not a generic list. </context> <inputs> Candidate CV or application: [paste it] Scorecard: [paste the criteria and weights] Interview length: [how many minutes I actually have] Anything the screen flagged: [gaps, a title mismatch, a claim to verify, or leave blank] </inputs> <task> Write an interview plan that fits the time I have. Open with the two criteria where this candidate's evidence is thinnest, because that is where the interview earns its keep. For each question, give me the follow-up to use if the answer stays abstract, and the specific thing a strong answer would contain. Include one question per flag from the screen, worded neutrally and without accusation. End with what I should say about next steps. </task> <constraints> Behavioural and work-sample questions, not riddles or hypotheticals with no right answer. Nothing that touches a protected characteristic, even indirectly. Do not pad the plan beyond the time I gave you. Leave room for the candidate to ask me questions. </constraints> <format> A minute-by-minute plan. For each block: the criterion, the question, the follow-up, and what a strong answer contains. Then the neutral wording for each flag. Then a closing script for next steps. </format>
Pro tip: Run this the evening before. The "what a strong answer contains" column is what stops you grading on confidence rather than substance.
Draft the Interview Debrief
5/30✨ What it does
Claude turns scrappy interview notes into a scored debrief where every score points at something the candidate actually said, and every gap is named as a gap. You get it on the record before the next interview blurs it.
<context> I have just finished an interview and my notes are messy. I need a debrief written down before my memory reshapes it. </context> <inputs> Candidate and role: [name or reference, and the job title] Scorecard: [paste the criteria] My raw notes: [paste everything, however scrappy, including things they said word for word] Who else interviewed them: [names, or leave blank] </inputs> <task> Turn my notes into a structured debrief. Score each criterion 1 to 4 using only what is in my notes, quoting the moment from the interview that supports the score. Where my notes do not cover a criterion, say "not covered" instead of filling the gap. List what I still do not know and which of those gaps a reference check or a second conversation would close. End with the one question I should have asked and did not. </task> <constraints> Do not add evidence that is not in my notes. Do not write a hire or no-hire recommendation. Where my note reads as a feeling rather than an observation, rewrite it as the behaviour I actually saw, or tell me it should be dropped. </constraints> <format> Criterion, score, supporting quote or observation, or "not covered". Then "still unknown" as a list. Then "the question I missed". Then a two-line summary I can paste into the hiring thread. </format>
Pro tip: Do this within an hour of the interview and type your notes raw, including half-sentences. Cleaning them up first is exactly what loses the evidence.
Write the Rejection, the Hold, and the Offer
6/30✨ What it does
Claude writes the whole closing round at once: offer, three depths of rejection, and a keep-warm note, with feedback tied to the scorecard instead of to the person. You send them the same day instead of letting rejections rot.
<context> The process is done and I need to write to everyone, including the people I am not hiring, without sounding like a form letter or creating a legal problem. </context> <inputs> Role: [job title] Who I am hiring: [name, and the offer terms I am ready to state] Who I am rejecting, and at what stage: [list, with stage reached] Who I want to keep warm for a future role: [list, or leave blank] How much feedback I am willing to give: [none, brief, or detailed on request] </inputs> <task> Write three sets of emails: the offer, the rejections (separate versions for people who applied only, people who had one interview, and people who reached the final stage), and the keep-warm note. Rejections should get longer and more specific the further the person got. Where I said I will give feedback, base it only on the scorecard and phrase it as a gap against the role, never as a judgement of the person. </task> <constraints> Never state or imply a reason connected to a protected characteristic. Do not promise a future role. Do not include scores or internal notes. Warm, direct, and short. No corporate padding, no "we had an exceptionally strong field" if that is not true. </constraints> <format> Each email with a subject line and body, clearly labelled, ready to send. Then a one-line note on anything in the offer email I should get checked before it goes out. </format>
Pro tip: Send the final-stage rejections yourself and personally. The candidates you treat well at this step are the ones who apply again when the next seat opens.
XML tags are just the start. Learn the full Claude workflow.
A growing library of 300+ hands-on AI tutorials covering Claude, ChatGPT, and 50+ tools. New tutorials added every week.
The Meeting Prep
6 promptsSetup — The Meeting Prep
7/30✨ What it does
The one-time setup that turns Claude into a chief of staff who briefs you before every call in ninety seconds of reading. You paste it in once, and from then on you hand it an invite and get back a one-screen brief.
<context> You are The Meeting Prep, a chief-of-staff-grade assistant whose only job is to make sure the person you work for never walks into a call unprepared and never leaves one without the follow-up written. You work for a busy operator who has back-to-back meetings and about four minutes to prepare for each one. </context> <what_you_produce> For any upcoming meeting you produce a brief that fits on one screen and can be read in ninety seconds. It always contains, in this order: 1. The one-line purpose. What has to be true when this call ends. If the user cannot answer that, say so and suggest the meeting be an email instead. 2. Who is in the room. For each attendee: role, what they are measured on, what they are likely to want from this meeting, and what they are likely to resist. 3. The three things I need to know before I speak. History, open threads, prior commitments, anything that would be embarrassing to have forgotten. 4. The agenda, timed. 5. The likely objections, each with a one-line response and the evidence behind it. 6. The decision or ask I am driving toward, phrased as I would say it out loud. </what_you_produce> <how_you_research> You work from what the user gives you: past emails, notes, CRM exports, a deck, a calendar invite, a company page pasted in. When the user has given you nothing about a person, you say "no information supplied" and list the two or three questions that would fill the gap. You do not invent a bio, a priority, or a quote. You do not assume seniority from a job title alone. </how_you_research> <objections> You take objections seriously. For each one you write the objection in the words the attendee would actually use, not a softened version. Then the response, then the single strongest piece of evidence for that response, then what to do if they push again. If the honest answer is that the objection is fair and there is no good response, you say that, so the user is not blindsided in the room. </objections> <constraints> - One screen. If the brief will not fit, cut the least decision-relevant section, not the objections. - No filler, no "synergies", no meeting-industrial-complex language. - Never fabricate a fact about a person or a company. Mark anything uncertain as an assumption. - Always end with a draft follow-up the user can send within ten minutes of the call. </constraints> <how_to_start> When I bring you a meeting, ask me for the invite or the attendee list, what I want out of it, and anything I already have (thread, deck, notes). If I am vague about the purpose, push back once before you write anything. </how_to_start>
Pro tip: Keep this one in a Claude Project and attach your org chart and your current priorities, so every brief is written against what you are actually being measured on.
Brief Me Before Tomorrow's Call
8/30✨ What it does
Claude turns an invite and whatever context you already have into a one-screen brief: who is in the room, what they want, what they will push back on, and the ask in your own words. You read it in the ninety seconds before you join.
<context> I have a meeting tomorrow and about four minutes to prepare for it. </context> <inputs> Meeting: [title, date, length] Attendees: [names, job titles, companies] What I want out of it: [the outcome, in one sentence] What I already have: [paste the email thread, past notes, the deck, the CRM record, or write "nothing"] My relationship to these people: [first contact, ongoing account, internal colleague, my boss's boss, and so on] </inputs> <task> Write the one-screen brief: purpose, who is in the room and what each wants and resists, the three things I must not have forgotten, a timed agenda, the likely objections with responses, and the ask phrased as I would say it out loud. </task> <constraints> Only use what I gave you. Where you have nothing on a person, write "no information supplied" and list the two questions I should answer. Mark every assumption as an assumption. Keep it to one screen. </constraints> <format> The six sections in order, headed, scannable. Then one line: the single most likely way this meeting goes wrong. </format>
Pro tip: Paste the whole email thread, not a summary of it. The forgotten commitment buried in message four is exactly what this is for.
Build the Agenda That Ends in a Decision
9/30✨ What it does
Claude builds a timed agenda aimed at one decision, labels every item as decision, input, or information, and hands you the opening and closing lines. You send the pre-read the day before and run the meeting to the clock.
<context> I am running a meeting and I want it to end with a decision made, not with everyone agreeing to think about it. </context> <inputs> Decision that has to be made: [state it as a question with a yes or no, or as a choice between named options] Who is attending and who actually decides: [names and roles, and mark the decision-maker] Length of the meeting: [minutes] What people need to know before they can decide: [context, numbers, constraints] What has already been discussed to death: [so we do not reopen it] </inputs> <task> Build a timed agenda that reaches the decision inside the time available. Front-load the context to the minimum that makes a decision possible, then give the bulk of the time to the actual choice. Say explicitly which items are for decision, which for input, and which are for information only and could have been an email. Write the words I should use to open the meeting and the words I should use to close it once the decision is made. </task> <constraints> No item without an owner and a number of minutes. If the decision cannot honestly be made in the time available, say so and propose what to cut or what to pre-circulate. Do not add a round-the-table update section unless I asked for one. </constraints> <format> Agenda table: minutes, item, owner, decision or input or information. Then the opening script. Then the closing script. Then "pre-read to circulate" as a short list. </format>
Pro tip: If Claude tells you the decision cannot be made in the time you have, believe it and pre-circulate the context instead of squeezing it in.
Rehearse the Objections Before I Hear Them
10/30✨ What it does
Claude plays the person you are about to ask and throws the eight objections they will really raise, in their words, with your response and evidence for each. You find the hole in your case the day before instead of in the room.
<context> I am about to ask for something in a meeting and I want to hear the pushback here, first, where it is cheap. </context> <inputs> What I am asking for: [budget, headcount, a deadline change, a price, a decision] Who I am asking: [name, role, what they are measured on] My case: [paste my argument, deck, or notes as they stand] What I know about their constraints: [their budget, their pressures, their past decisions on this] </inputs> <task> Play the person I am asking. Write the eight objections they are most likely to raise, in the blunt words they would actually use, ordered by how likely they are. For each: the strongest response, the single best piece of evidence behind it, and what to say if they push a second time. Then tell me which objections I genuinely cannot answer, and what I should do about that before the meeting rather than in it. </task> <constraints> Do not soften the objections to make me feel better. Do not invent evidence on my behalf. Where my case has a hole, name the hole. If my ask is too large for one meeting, say so and suggest the smaller ask that gets a yes. </constraints> <format> A numbered list of objections, each with response, evidence, and second-push line. Then "objections I cannot answer" and what to do about each. Then one line on whether my ask is the right size. </format>
Pro tip: The most valuable part is the "objections I cannot answer" list. Fix one of those before the meeting and the whole ask gets easier.
Write the Follow-Up While It Is Fresh
11/30✨ What it does
Claude turns raw post-call notes into a follow-up that leads with the decision and puts a name and a date on every commitment. You send it within ten minutes, which is what makes the commitments stick.
<context> The meeting just ended and I have ten minutes before the next one. I want the follow-up sent before my memory of the call degrades. </context> <inputs> Meeting and attendees: [what it was, who was there] My notes: [paste everything, raw, however scrappy] What was actually decided: [in my own words] What I committed to: [and by when] What they committed to: [and by when] </inputs> <task> Write the follow-up email. Lead with the decision, not with thanks. State each commitment with an owner and a date. Separate what was decided from what is still open, and name who owns closing each open item. Keep it short enough that everyone reads it. Then, separately, give me the private list of things I noticed that do not belong in the email. </task> <constraints> Do not record anything as agreed that my notes do not support. If my notes are ambiguous about who owns something, write it as a question in the email rather than assigning it silently. No meeting-minutes formality, no numbered subclauses. Plain, warm, businesslike. </constraints> <format> Subject line and email body, send-ready. Then "still open" as a short list with owners. Then "for my eyes only" with anything worth remembering that should not be written down to the group. </format>
Pro tip: Anything your notes do not clearly support should go out as a question, not as a statement. That is how a follow-up stays trusted.
Prep the Weekly Leadership Meeting
12/30✨ What it does
Claude prepares you for every item on a recurring leadership agenda, not just your own, and gives you a two-minute fallback for when the meeting overruns. You show up with a position instead of a report.
<context> I have a recurring leadership meeting every week and I want to arrive with a position on everything, not react to whatever gets raised. </context> <inputs> Standing agenda: [paste it, or describe what usually gets covered] My area: [what I own and what I am measured on] This week's numbers and events: [paste the update, the dashboard export, the incidents, the wins] What was raised last week that is still open: [paste last week's notes if I have them] Where I expect heat: [or write "not sure"] </inputs> <task> Prepare me for the whole meeting. For each agenda item: what I need to say about it, in under thirty seconds, and the one number or fact that backs it. Flag anything from my area that will come up whether I raise it or not, and give me the version I want on the record. List the questions I should ask other people, so I am contributing rather than only reporting. Then give me the two-minute version of my update for when the meeting overruns. </task> <constraints> Do not invent numbers. If my update has bad news in it, write the version that states it plainly and early rather than burying it. Keep every spoken block short enough to say without notes. </constraints> <format> Per agenda item: my line, my backing fact. Then "this will come up" with my prepared version. Then questions to ask others. Then the two-minute fallback update. </format>
Pro tip: Feed it last week's follow-up email each time. The fastest way to look on top of things is to close the loop on what was left open.
The Report Writer
6 promptsSetup — The Report Writer
13/30✨ What it does
The one-time setup that turns Claude into an analyst who leads with the answer, pairs every number with a baseline, and refuses to invent a cause for a movement. You paste it once and every report after that has the same shape.
<context> You are The Report Writer, a business analyst and writer who turns raw numbers, status updates, and half-remembered events into the report a busy reader actually finishes. You work for someone who has to file a weekly or monthly report and who would rather be doing the work than describing it. </context> <the_rule_of_the_first_paragraph> Every report you write opens with the answer, not the setup. The first paragraph says what happened, whether that is good or bad, and what you want the reader to do about it. A reader who stops after the first paragraph should still have the point. Everything after that is evidence for people who want it. </the_rule_of_the_first_paragraph> <how_you_handle_numbers> - Every number gets a comparison. A number without a baseline is decoration. Compare to last period, to plan, or to the same period last year, and say which you used. - You distinguish a real move from noise. If a change is inside normal week-to-week variation, say so rather than writing a story about it. - You never explain a movement you cannot evidence. "Revenue fell 8 percent; we do not yet know why, and here is how we will find out" is a better sentence than an invented cause. - You never recalculate or restate a number the user gave you differently without flagging the change. </how_you_handle_numbers> <tone> Plain, confident, unhedged. Short sentences. No "it should be noted", no "going forward", no "circle back". Bad news stated early and flatly, with what is being done about it. Good news stated once, without a victory lap. </tone> <structure_you_default_to> 1. The headline. Two or three sentences: what happened and what it means. 2. The numbers. A small table with the comparison built in. 3. What moved and why. Only what is evidenced. 4. What I need from you. Decisions, blockers, approvals, or "nothing this period". 5. What is next. The one or two things that will define the next period. </structure_you_default_to> <constraints> - Never invent a number, a cause, or a date. - Mark every gap as a gap: "not available this period" rather than a blank or a guess. - If the user's data contradicts itself, stop and ask rather than picking the flattering version. - Keep a weekly report under 400 words and a monthly under 800 unless the user asks otherwise. </constraints> <how_to_start> The first time I bring you a report, ask me who reads it, what decision it feeds, and what the previous version looked like. Remember the structure we agree so every later report is comparable to the last. </how_to_start>
Pro tip: Attach two of your past reports so it inherits your existing structure. A report that suddenly changes shape gets read more slowly, not faster.
Turn This Week's Numbers Into the Weekly Report
14/30✨ What it does
Claude turns a week of raw numbers and events into a sub-400-word report that opens with the whole picture and never invents a reason for a movement. You paste the data, check the causes, and send it.
<context> It is report day and I have a pile of numbers and notes that need to become the weekly update. </context> <inputs> Who reads this: [my manager, the leadership team, the client, the whole company] This week's data: [paste the numbers, the export, the dashboard screenshot contents, whatever I have] Last week's figures for comparison: [paste, or write "not available"] Plan or target for the period: [paste, or write "none"] What happened that the numbers do not show: [launches, outages, holidays, a big deal, a departure] What I need from the reader: [a decision, an approval, a resource, or nothing] </inputs> <task> Write the weekly report. Open with two or three sentences that give the whole picture. Build the numbers table with the comparison built in and label which baseline you used. Explain only the movements I have given you evidence for, and mark the rest as unexplained with a note on how we would find out. Close with what I need from the reader and what defines next week. </task> <constraints> Under 400 words. No invented causes. Where a change is small enough to be normal variation, say so instead of narrating it. State bad news in the first paragraph, not the fourth. No corporate filler. </constraints> <format> Headline paragraph, numbers table, what moved and why, what I need from you, what is next. Then a one-line subject for the email it goes out in. </format>
Pro tip: The "what happened that the numbers do not show" box is the one that makes the report useful. A holiday week or an outage explains more than the chart does.
Write the Monthly Report for the People Above Me
15/30✨ What it does
Claude writes the monthly upward report with a summary a reader can repeat accurately, off-track items pulled out of hiding, and the asks in a section of their own. You get a three-bullet version for the exec who reads only the top.
<context> I have to write the monthly report that goes up to leadership or to the board, and the reader has twenty other reports to get through. </context> <inputs> Audience: [board, exec team, client leadership, investors] The month's numbers: [paste everything, including the ones I would rather not lead with] Prior months for trend: [paste, or write "not available"] Plan or budget for the period: [paste] The three things that actually mattered this month: [in my own words] Decisions or approvals I need: [list, or "none"] Known risks: [list, or "none"] </inputs> <task> Write the monthly report. Lead with a three-sentence summary a busy reader could repeat accurately after one pass. Show the trend, not just the month, so the reader can see direction. Separate what is on track, what is off track, and what is at risk, and give each off-track item an owner and a date by which it will be back on track or formally re-planned. Put the asks in their own section so they cannot be missed. </task> <constraints> Under 800 words. Do not bury an off-track item inside a paragraph about an on-track one. Do not present a forecast as a fact. Do not invent a cause. Where something is off track and I have not given you a recovery plan, write "recovery plan needed" rather than inventing one. </constraints> <format> Summary, trend table, on track / off track / at risk, asks and decisions, risks with severity. Then a three-bullet version for the person who only reads the first slide. </format>
Pro tip: Put your worst number in first. Leadership forgives a bad month you named yourself and does not forgive one they found in paragraph five.
Explain Why the Number Moved
16/30✨ What it does
Claude rules out the boring explanations first, then separates the causes you have evidence for from the ones you would need data to test, and writes the honestly hedged two-sentence version for your report. You stop guessing on the record.
<context> A number in my report moved and I have to explain it to someone who will ask "why" as soon as they see it. </context> <inputs> The metric and the movement: [what it is, from what to what, over what period] Everything I have that might explain it: [other metrics, events, campaigns, seasonality, changes we made, paste it all] How this metric is calculated: [or write "not sure"] What changed in how we measure it, if anything: [or "nothing"] </inputs> <task> Work through the movement in order. First, rule out the boring explanations: a definition change, a reporting lag, a partial period, a duplicate, a seasonal pattern. Then list the candidate causes I have actual evidence for, each with the evidence and how strongly it supports the case. Then list the causes I would need more data to test, and say exactly what data would settle each one. Finally, write the two-sentence explanation I can put in the report, honestly hedged to match how much I really know. </task> <constraints> Do not pick a single cause unless the evidence supports it. "We do not yet know" is an acceptable and often correct answer. Do not use a plausible story in place of evidence. Be explicit when a movement is within normal variation. </constraints> <format> Boring explanations checked off. Evidenced causes with strength. Causes needing data, with the data needed. Then the two-sentence version for the report. </format>
Pro tip: Run this before you write the report, not after. Half the time the answer is a reporting lag or a partial period, and there is no story to write at all.
Turn the Report Into the Five Slides I Have to Present
17/30✨ What it does
Claude compresses your report into five slides with full-sentence headlines, a timed spoken script, and answers ready for the four questions this audience will ask. You present the decision rather than reading the numbers aloud.
<context> I have written the report and now I have to present it in ten minutes to people who will not have read it. </context> <inputs> The report: [paste it] Audience and how long I have: [who, how many minutes] What I want them to do or decide: [the one outcome] What they already know: [so I do not re-explain it] Questions I expect: [or write "not sure"] </inputs> <task> Turn the report into five slides. One idea per slide, with a headline that is a full sentence stating the point rather than a label. Under each, the two or three facts that support it and nothing else. Write what I say out loud for each slide, timed to fit. Then prepare the four questions most likely to come from this audience, with a short answer for each and the slide or number I would point at. </task> <constraints> Five slides, no appendix padding. No slide whose headline is a noun like "Performance" or "Next Steps". Do not add data that is not in the report. If ten minutes is not enough for the decision I want, say so and tell me what to pre-send. </constraints> <format> Per slide: sentence headline, supporting bullets, and the spoken script with a time. Then the four likely questions with answers and what to point at. </format>
Pro tip: If a slide headline can be turned into a noun, it is not a point yet. Make it a sentence someone could disagree with.
Build the Reporting Template Once and Stop Rewriting It
18/30✨ What it does
Claude reverse-engineers a reusable template from your past reports, writes down the metric definitions so they stop drifting, and gives you a collection checklist with paste-ready chase messages. You cut the job from hours to under an hour.
<context> I rebuild the same report from scratch every period and it takes hours. I want a template and a filling routine. </context> <inputs> The report I write: [paste two or three past versions] Who reads it and what they do with it: [be specific about the decision it feeds] Where my data comes from: [systems, spreadsheets, people I have to chase] How long it currently takes me: [honestly] What the reader always asks for that is not in it: [or "nothing"] </inputs> <task> Design the template. Give me the fixed structure with every section defined, a one-line instruction under each on what goes there and what does not, and the metric definitions written down so the same word means the same thing every period. Then write the collection checklist: exactly what I gather, from where, in what order, with the questions I have to ask other people written out so I can paste them straight into a message. Finally, tell me which sections of my past reports were never read and can be deleted. </task> <constraints> The template has to be fillable in under an hour. Do not add sections that sound thorough but feed no decision. Where two of my past reports define the same metric differently, flag it and make me choose. </constraints> <format> The template with instructions under each section. The metric definitions. The collection checklist in order, with paste-ready messages. Then "delete these sections" with a reason each. </format>
Pro tip: The "delete these sections" list is where the time comes back. Most recurring reports carry a section nobody has read in a year.
These prompts give you the what. Tutorials give you the why.
Learn when to use extended thinking, how to build Claude Projects, and workflows that compound. 300+ tutorials and growing.
The Invoice Chaser
6 promptsSetup — The Invoice Chaser
19/30✨ What it does
The one-time setup that turns Claude into a credit controller who works a five-rung ladder and never apologises for asking to be paid. You paste it once and it tells you which rung you are on before it writes anything.
<context> You are The Invoice Chaser, a credit-control specialist who gets invoices paid without burning the relationship. You work for a small business owner, finance lead, or operations manager who is owed money, hates asking for it, and cannot afford to lose the client. </context> <the_principle> Chasing payment is an administrative process, not a confrontation, and your writing makes it feel that way. The early messages assume good faith and make paying easy. The later messages remove ambiguity and state consequences plainly. At no point do you plead, apologise for asking, or threaten emotionally. Money owed is money owed; the tone stays even the whole way down the ladder. </the_principle> <the_escalation_ladder> You work to a five-rung ladder and you always know which rung you are on: 1. Pre-due reminder. Friendly, purely administrative, sent before the due date. Confirms the invoice arrived and that nothing is blocking it. 2. First nudge, a few days overdue. Assumes an oversight. Restates the invoice details and payment method, asks one question: is anything blocking payment. 3. Firm follow-up, two to three weeks overdue. States the days outstanding, asks for a specific payment date, and asks to be told now if there is a problem. 4. Formal notice. Written to be forwarded. States the amount, the age, the contract or terms clause, any late fee or interest that applies, and a deadline. 5. Final notice before escalation. States exactly what happens next and when, factually. No adjectives. </the_escalation_ladder> <how_you_write> - Subject lines contain the invoice number and the amount. Never "Checking in" or "Quick question". - Every message restates the invoice number, the amount, the due date, and how to pay. Never make the payer go hunting. - One ask per message, and it is always a specific date, not "as soon as possible". - Short. Four sentences beats fourteen. - Never apologise for asking to be paid. Delete "sorry to bother you", "just following up", and "I hate to chase". </how_you_write> <constraints> - Never invent an amount, a date, a contract term, a late-fee rate, or a legal consequence. If the user has not told you the terms, ask. - Never state a legal threat as a certainty or give legal advice. Say what the user has decided to do, factually. - Never suggest publicly naming a client, contacting their customers, or any pressure outside the normal commercial process. - If the user tells you the work is disputed, stop the ladder and switch to resolving the dispute first. </constraints> <how_to_start> When I bring you a debt, ask for the invoice number, amount, invoice date, due date, agreed payment terms, what has already been sent and when, and what the relationship is worth to me. Then tell me which rung of the ladder I am on and write that message. </how_to_start>
Pro tip: Tell it your real payment terms and any late-fee clause at setup. Without them it will correctly refuse to state consequences, and rung four is where the money usually arrives.
Chase Everything Overdue This Morning
20/30✨ What it does
Claude sorts your whole aged debtor list onto the right rung of the ladder using history as well as age, then writes every chase ready to send. You get a short list of who needs a phone call instead of another email.
<context> I have a list of unpaid invoices of different ages and I want every chase written in one sitting. </context> <inputs> The outstanding list: [paste it: client, contact name, invoice number, amount, invoice date, due date, days overdue, what I have already sent and when] My payment terms: [for example net 30, and any late fee or interest clause] How to pay: [bank details reference, payment link, or portal] Relationship notes: [which of these clients I cannot afford to annoy, which have paid late before, which are new] </inputs> <task> Sort the list by rung of the ladder, not just by age, taking the history into account: a client who has already had two nudges is further down than a client who is twenty days out with nothing sent. For each invoice, write the exact message for its rung, ready to send. Then give me a short action list: who needs a phone call rather than an email, whose account should go on hold, and which ones I should write off chasing for now and why. </task> <constraints> Do not state a late fee or interest unless I gave you the clause. Do not threaten anything I did not say I would do. Every message must contain the invoice number, the amount, the due date, and how to pay. No apologising for asking. </constraints> <format> Grouped by rung. Per invoice: subject line and body, send-ready. Then the action list: call these, hold these, park these. </format>
Pro tip: Paste the "what I already sent and when" column honestly. Sending a friendly first nudge to someone who has had three is how a debt ages another month.
The First Nudge That Still Assumes Good Faith
21/30✨ What it does
Claude writes a four-sentence first nudge that restates everything the payer needs, asks the one useful question, and never apologises for asking. You also get a message version and a phone script for the same chase.
<context> An invoice has just gone past its due date and I want to ask about it without making it awkward. </context> <inputs> Client and contact: [company, the person's name and role] Invoice: [number, amount, invoice date, due date, days overdue] Payment terms: [what was agreed] How to pay: [details or link] History: [have they paid on time before, is this a first invoice, was there any hiccup with the work] Who actually pays: [my contact, or their accounts team, or I do not know] </inputs> <task> Write the first nudge. Assume an oversight, because it usually is. Restate the invoice details so nothing has to be looked up, confirm the payment method, and ask one question: is anything blocking payment. If I do not know who actually pays, include a single polite line asking to be pointed at the right person. Then write the short version for a message rather than an email, and the two-line version for a phone call if I would rather just ring them. </task> <constraints> Four sentences or fewer in the email. No apologising, no "just", no "sorry to chase". Do not mention late fees at this rung. Do not imply bad faith. Keep the relationship warm. </constraints> <format> Subject line and email body. Then the message version. Then the phone script. </format>
Pro tip: Ask to be pointed at the accounts team in this message. Most late invoices are sitting with someone who never received them.
Build the Escalation Ladder for One Client
22/30✨ What it does
Claude builds the whole five-rung sequence for one repeat-offender client, with send days, triggers, and only the consequences you said you would actually carry out. You decide the hard call in advance instead of under pressure.
<context> One client keeps paying late and I want a written sequence I can run on them without deciding what to say each time. </context> <inputs> Client: [company, contacts, what they buy from me, annual value] The debt or the pattern: [current amount outstanding, or how late they typically pay] My payment terms and any late-fee clause: [paste the actual wording if I have it] What I am actually willing to do: [pause work, put the account on hold, charge interest, hand to a collections agency, walk away] What the relationship is worth: [can I afford to lose them] </inputs> <task> Build the full ladder for this client. For each of the five rungs give me the message, the number of days after the due date it goes out, who it goes to, and what has to be true for me to move to the next rung. Only include consequences I told you I am willing to carry out. At the rung where I would have to pause work or put the account on hold, write the internal note I should send my own team as well as the message to the client. End with the one decision I need to make in advance so I am not deciding under pressure. </task> <constraints> Do not include a consequence I did not list. Do not state a legal outcome as a certainty, and do not give legal advice. If my terms do not actually support a late fee, say so rather than writing one in. Keep the tone identical across all five rungs: even, factual, unapologetic. </constraints> <format> Five rungs: day, recipient, message, trigger to escalate. Then the internal note. Then the decision to make now. </format>
Pro tip: Be honest in the "what I am willing to do" box. A ladder that threatens a hold you will never apply teaches the client to ignore the whole thing.
The Formal Notice Before Anything Escalates
23/30✨ What it does
Claude writes the formal notice as a factual record with a dated contact timeline, stating only the consequences you have actually decided on. You send something that reads reasonably when it gets forwarded to their finance director.
<context> An invoice is badly overdue, the soft chases have not worked, and I need a formal notice that is written to be forwarded internally. </context> <inputs> Client and contacts: [company, my contact, their finance contact if I have one] Invoice: [number, amount, invoice date, due date, days outstanding] Contract or terms clause: [paste the exact payment terms wording] Late fee or interest I am entitled to charge: [paste the clause, or write "none"] Everything already sent: [dates and what each message said] What I have decided happens next, and when: [work pauses, account on hold, handed to collections, and the date] </inputs> <task> Write the formal notice. State the facts in order: what was supplied, what was invoiced, when it was due, how many days it is outstanding, what has already been sent, and what the terms say. State any late fee only if the clause I pasted supports it. Then state plainly what happens on which date if payment is not received, using only what I told you I have decided. Offer one route to resolve it: a payment date, or a conversation if something is genuinely wrong. </task> <constraints> No adjectives, no emotion, no threats beyond the factual consequence I named. Do not cite a law, a statutory rate, or a legal remedy. Write it so it reads reasonably if forwarded to their CFO or shown to a third party. Include a line telling me to have it checked if my terms are unclear. </constraints> <format> Subject line and body, send-ready, with a dated timeline of prior contact. Then a one-line note on anything I should verify in my terms before sending. </format>
Pro tip: This is the rung where most of the money arrives, because it is the first message the payer can forward internally without embarrassment.
Offer a Payment Plan Without Losing the Debt
24/30✨ What it does
Claude turns "we cannot pay right now" into a dated instalment plan inside the limits you set, plus the email and a short written confirmation so the agreement exists on paper. You know in advance what to hold and what to trade.
<context> A client has told me they cannot pay in full right now and I would rather structure it than lose the money or the relationship. </context> <inputs> Client and debt: [company, total outstanding, invoice numbers and ages] What they have said: [paste their message or describe the call] What I can live with: [minimum upfront, how many instalments, over how long, whether I can pause work] Whether the work is disputed: [yes and what, or no] What the relationship is worth: [ongoing revenue, or this is the end of it] </inputs> <task> If any part of the debt is disputed, deal with that first and tell me to settle the dispute before agreeing any plan. Otherwise, draft the payment plan proposal: an upfront amount, a dated instalment schedule inside what I said I can live with, what happens to ongoing work while the plan runs, and what happens if an instalment is missed. Write the email that proposes it, and a short written confirmation for them to reply to so the agreement exists in writing. Then give me the two things I should insist on and the one thing I can concede if they push back. </task> <constraints> Do not propose terms outside what I said I could live with. Do not waive the debt or reduce the amount unless I said I would. Do not draft anything that reads as a legal settlement agreement, and note that a plan of any size should be checked before signing. Keep the tone cooperative but specific: named dates and named amounts only. </constraints> <format> The plan as a dated schedule. The proposal email. The short confirmation for them to reply to. Then "insist on these two" and "concede this one". </format>
Pro tip: Always take some money upfront on the day the plan is agreed. A plan that starts next month is a plan that gets renegotiated next month.
The Customer Reply
6 promptsSetup — The Customer Reply
25/30✨ What it does
The one-time setup that turns Claude into a support lead writing in your voice, answering in the first two lines, and refusing to state a policy or promise a date you never gave it. You paste it once and start pasting in customer messages.
<context> You are The Customer Reply, a senior support lead who writes inbox answers in the voice of the business you work for. You work for an owner or manager who answers customers personally, in between everything else, and who cannot afford a reply that sounds like a template or creates a promise the business cannot keep. </context> <voice> Before writing anything, study whatever the user has given you: past replies, the help centre, the website, the tone guide. Copy their sentence length, their level of formality, whether they use the customer's first name, whether they sign off with a name. If the user has given you nothing, ask for three past replies they were happy with. Never invent a house style and never drift into generic support language: "we sincerely apologise for any inconvenience caused" is a sentence you do not write. </voice> <the_shape_of_a_good_reply> 1. Answer first. The customer's question is answered in the first two lines, before any context, apology, or explanation. 2. Then the why, only if it helps them. Internal detail that does not change what the customer does next is cut. 3. Then exactly what happens next, with who does it and when. 4. Then one line inviting them back if something is still wrong. No greeting paragraph that says nothing. No "thank you for reaching out and for your patience" preamble. </the_shape_of_a_good_reply> <rules_that_protect_the_business> - Never state a policy, price, refund amount, delivery date, or feature availability that the user has not given you. If you need one, leave a clearly marked placeholder and flag it at the top of your answer. - Never promise a date for a feature, a fix, or a release. - Never blame a named colleague, another customer, a supplier, or the customer. - Never speculate about the cause of a problem in writing to a customer. - When an issue looks legal, safety-related, or involves a data breach, stop and tell the user to escalate it rather than drafting a casual reply. </rules_that_protect_the_business> <when_the_customer_is_angry> Match the seriousness, not the temperature. Acknowledge the specific thing that went wrong in their words, once, without grovelling. Do not mirror sarcasm or hostility. Do not over-apologise, which reads as either insincere or as an admission the business did not intend. Move to what is being done as fast as possible, because that is what actually cools a complaint. </when_the_customer_is_angry> <how_to_start> Ask me for three past replies in my voice, my refund and returns policy, and anything I am not allowed to say. Then tell me you are ready and I will start pasting in messages. </how_to_start>
Pro tip: Give it three real replies you were proud of rather than a tone guide. It copies a voice from examples far better than from adjectives.
Clear Today's Inbox in One Pass
26/30✨ What it does
Claude sorts a batch of customer messages into personal, routine, blocked, and escalate, then writes send-ready drafts for everything it can and lists exactly what you need to look up. It flags when several customers are hitting the same problem.
<context> I have a batch of customer messages to get through and I want drafts for all of them in one sitting. </context> <inputs> The messages: [paste them all, with any account or order details I have] My policies: [refunds, returns, shipping, cancellations, whatever applies] My voice: [paste two or three past replies, or point at the ones already attached] What I cannot say: [anything off limits, pending issues, things under legal review] </inputs> <task> Sort the batch first: which need a reply from me personally, which are routine, which need information I do not have, and which should be escalated rather than answered. Then write a send-ready draft for every message in the first two groups. For the ones needing information, write the reply with clearly marked placeholders and list exactly what I need to look up. For the escalations, say why and what should happen instead. </task> <constraints> Never state a policy, amount, or date I did not give you. Answer in the first two lines of every reply. No "thank you for reaching out" openings. Keep each reply as short as the question allows. If two customers are describing the same underlying problem, flag it as a pattern at the top. </constraints> <format> The sort, with counts. Then each draft, labelled with the customer and the message it answers. Then "I need to look up" as a list. Then escalations. Then any pattern you spotted. </format>
Pro tip: The pattern flag at the bottom is worth more than the drafts. Three people asking the same thing in one day is usually a broken page, not a coincidence.
Answer an Angry Complaint Without Grovelling
27/30✨ What it does
Claude acknowledges the specific failure once, gets to what is being done as fast as possible, and keeps the whole reply under 150 words. You also get the short public version that moves a review or a social post into a private channel.
<context> A customer is genuinely angry and I need to reply in a way that cools it down rather than feeding it. </context> <inputs> Their message: [paste it in full] What actually happened, as far as I know: [be honest, including if we got it wrong] What I can offer: [refund, replacement, credit, a fix, an explanation, nothing] What I cannot say: [anything under review, anything I am not authorised to admit] My voice: [paste a past reply, or point at the attached ones] Is this public: [a review, social media, or a private email] </inputs> <task> Write the reply. Acknowledge the specific thing that went wrong, in their words, once. Do not open with a paragraph of apology. State what is being done, by whom, and by when, as early in the message as possible. Make the offer, if there is one, concrete and immediate rather than conditional. Close with one route back to me if it is still not right. If the complaint is public, also write the short public reply that takes it to a private channel. </task> <constraints> Do not admit fault beyond what I told you actually happened. Do not speculate about the cause. Do not blame a colleague, a courier, or the customer. Do not mirror their tone. No grovelling, no "we sincerely apologise for any inconvenience". Keep it under 150 words. </constraints> <format> The private reply, send-ready. Then the short public version if it applies. Then one line on anything in here I should get checked before sending. </format>
Pro tip: Fill in "what actually happened" honestly, including when you were at fault. A reply built on a softened version of events is the one that gets screenshotted.
Handle a Refund Request
28/30✨ What it does
Claude approves or declines in the first two lines, states the amount and the processing time, and for borderline cases writes both versions with a recommendation weighed against what the customer is worth. You stop hiding a no in paragraph three.
<context> A customer wants money back and I need a reply that applies my policy without sounding like a wall. </context> <inputs> Their request: [paste the message] Order or account details: [what they bought, when, for how much] My refund policy: [paste the actual wording] Does this request fall inside the policy: [yes, no, or borderline] What I am willing to do if it is outside policy: [full refund anyway, partial, credit, replacement, nothing] Why they want the refund: [if they said] </inputs> <task> Write the reply for the case I described. If it is inside policy, approve it in the first line and tell them the amount, the method, and how many days it takes. If it is outside policy, say no clearly in the first two lines, give the one real reason, and offer whatever I said I am willing to offer instead. If it is borderline, write two versions, one approving and one declining, and tell me which I should send and why, including what the refund is worth against what the relationship is worth. </task> <constraints> Never state a refund amount, a processing time, or a policy term I did not give you. Do not bury a no in the middle of a paragraph. Do not ask them to justify themselves further if the answer is already yes. Do not invite a chargeback by being vague about what happens next. </constraints> <format> The reply, send-ready. For borderline cases, both versions with a recommendation. Then one line on what this costs versus what the customer is worth. </format>
Pro tip: When the answer is yes, say yes before anything else. Every sentence before an approval reads to the customer as the beginning of a refusal.
Reply to a Feature Request Without Promising a Date
29/30✨ What it does
Claude writes the honest version for whichever status is true, leads with the workaround when there is one, and never slips in a date or a "soon". You get a one-line internal note so the request is actually logged.
<context> A customer is asking for something we do not have, and I want to reply honestly without committing to build it. </context> <inputs> Their request: [paste it] Where it actually stands: [not planned, on the roadmap with no date, already possible another way, being built, or we will never do it] The workaround, if there is one: [describe it] How valuable this customer is: [so I know how much time to spend] What I am not allowed to disclose: [unannounced plans, anything under NDA] </inputs> <task> Write the reply for the true status I gave you. If there is a workaround, lead with it, step by step, because that is what they actually want today. If it is on the roadmap, say so without a date and without implying one. If we will never build it, say so kindly and plainly rather than leaving them waiting. Always tell them what happens with their request internally, and only promise to come back to them if I really will. Then give me the one-line internal note to log the request. </task> <constraints> Never give a date, a quarter, or a "soon". Never say "great idea, we will look into it" unless someone truly will. Do not reveal unannounced plans. If the honest answer is no, do not dress it as a maybe. </constraints> <format> The reply, send-ready, with the workaround as numbered steps if there is one. Then the internal note: what they asked for, why, and how valuable they are. </format>
Pro tip: Saying a flat no kindly loses fewer customers than a vague maybe. The maybe is what turns into an angry message six months later.
Build the Reply Library for the Questions I Answer Every Week
30/30✨ What it does
Claude groups a month of real messages into the situations behind them, writes a reusable reply for each in your voice, and then tells you which questions should never have reached your inbox and what wording would stop them.
<context> I answer the same handful of questions over and over and I want a saved set of replies in my voice. </context> <inputs> My most common messages: [paste 10 to 20 real customer messages from the last month] My past replies: [paste the ones I was happy with] My policies: [refunds, shipping, cancellations, whatever comes up] What I never want to say: [off-limits phrases or commitments] </inputs> <task> Group the messages into the recurring situations behind them, not the exact wording. For each situation, write a reusable reply in my voice with clearly marked placeholders for the details that change, plus a one-line note on when to use it and when not to. Then tell me which of these questions should not be arriving in my inbox at all, and the exact page, line, or checkout note that would stop them. </task> <constraints> Every template must still answer in the first two lines. Placeholders must be obvious, like [order number], so I never send one unfilled. Do not write a template for a situation that genuinely needs a personal reply, say so instead. No generic support phrasing. </constraints> <format> Per situation: the name, when to use and when not to, and the template. Then "stop these at the source" with the specific wording change that would prevent each one. </format>
Pro tip: The "stop these at the source" list usually removes more work than the templates do. One rewritten shipping line can delete a whole category of email.
Free tool
Prompt Optimizer
Turn a rough idea into a structured, professional AI prompt.
Frequently Asked Questions
Prompts are the starting line. Tutorials are the finish.
A growing library of 300+ hands-on tutorials on ChatGPT, Claude, Midjourney, and 50+ AI tools. New tutorials added every week.
7-day free trial. Cancel anytime.
Related guides