n8n AI Agent prompts for routing, summaries and triage

Turn this call transcript into a CRM note of at most 160 words. Transcript: [PASTE TRANSCRIPT]. Allowed CRM note fields: [ALLOWED NOTE FIELDS]. Output only those fields as a JSON object. Quote the customer once, under 25 words, using their wording. If a field in [ALLOWED NOTE FIELDS] has no support in the transcript, set it to [VERIFY]. Do not invent a close date, amount or competitor. Do not add keys beyond the allowed list, because extra keys break the later Set node.

Summarise this email thread for the next node. Thread: [PASTE EMAIL THREAD]. Audience of the digest: [INTERNAL TEAM]. Return five bullets: ask, decisions, open questions, dates named, and people named. Keep the whole digest under 140 words. Use names and dates only if they appear in the thread. If a date is referred to but not written, mark it [VERIFY]. Never invent a deadline or a commitment the sender did not make.

Extract action items from [MEETING TITLE] held on [MEETING DATE]. Notes: [PASTE NOTES]. Output a table with columns Owner, Task, Due date and Source line. Owners must match names in the notes. If a task has no owner or no due date, write [VERIFY] in that cell. Do not invent a teammate who was not named. Limit the table to 12 rows. After the table, list any task that is a question rather than work, in one line each.

Write a weekly ops digest from these n8n execution notes for week [WEEK DATES]. Notes: [PASTE EXECUTION NOTES]. Cover only workflows I named: [WORKFLOW NAMES]. Output four short sections: Ran as expected, Failed and why (quote the error text), Skipped, and Ask for a human. If an error text is missing, write [VERIFY] in Failed and why. Do not invent a success count, duration or node name I did not supply. Keep the digest under 180 words.

Draft a customer recap email of about 110 words to [CUSTOMER NAME] after [EVENT]. Facts you may state: [FACTS I APPROVED]. Tone: plain and warm, no exclamation marks. Do not invent a delivery date, feature or refund. If the facts skip a date the recap would need, write [VERIFY] in that sentence. Close with one question that uses only a topic from [FACTS I APPROVED]. Output subject line, then body.

Classify this support ticket into one category from my list and nothing else. Ticket: [PASTE TICKET TEXT]. Allowed categories: [CATEGORY LIST]. Return JSON with keys category, evidence_quote (under 30 words from the ticket), and confidence (high, medium, low). If the ticket fits none of the categories, set category to [VERIFY] and explain in one sentence. Do not invent a category, product version or account id. Do not call a write tool until category is one of the allowed values.

Set urgency using only words that appear in the ticket and in my SLA glossary. Ticket: [PASTE TICKET]. SLA glossary: [SLA GLOSSARY]. Output a line: Urgency = P1, P2 or P3, then a 25 word reason that quotes the glossary term you used. If the ticket never uses a glossary term, write [VERIFY] for Urgency. Do not invent an outage, a number of users affected, or a contractual deadline. After the reason, name the next node that should run: notify, queue, or stop.

Draft a first support reply of about 100 words to [CUSTOMER DISPLAY NAME] about [ISSUE SUMMARY]. Known facts: [KNOWN FACTS]. Macros I allow you to echo: [ALLOWED MACRO NAMES]. Structure: acknowledge, what we know, what we need, when we will write again using [REPLY BY WINDOW]. If [REPLY BY WINDOW] is blank, write [VERIFY] there. Do not invent a fix, a workaround, or a status. Do not promise a result I did not list in [KNOWN FACTS].

Build an escalation packet for [ENGINEERING CHANNEL] about ticket [TICKET ID]. Include four labelled blocks: Symptom (from [PASTE SYMPTOM]), Steps already tried ([STEPS TRIED]), Impact as the customer stated it ([IMPACT QUOTE]), and Ask of engineering in one sentence. If a block has no source text, write [VERIFY] in that block. Do not invent a stack trace, a commit, or a severity. End with the exact tool name from [CONNECTED TOOL NAMES] that should post the packet, or write none if I listed none.

Decide whether ticket A is a duplicate of tickets I already retrieved. Ticket A: [TICKET A TEXT]. Retrieved tickets: [RETRIEVED TICKET SUMMARIES]. Compare product area, error text and account name only. Output a table with columns Other ticket, Shared tokens and Duplicate (yes, no, unsure). If account name is missing on A, write [VERIFY] under Duplicate for every row. Do not invent a ticket id. Recommend merge only when Duplicate is yes on one row and the shared tokens include the error text.

Write a System Message of at most 180 words for my n8n AI Agent. Purpose of the agent: [AGENT PURPOSE]. Tools actually connected on this node: [CONNECTED TOOL NAMES]. For each tool, write one line: when to call it, and what it must receive. State that the agent may call only those tools. If I listed no tools, write [VERIFY] in the tools section and tell the agent to return text only. Do not invent a tool, credential or API path. End with: if a required input is missing, stop and return [VERIFY].

Write a System Message that forces this output shape and no extra keys: [REQUIRED JSON KEYS]. Each key's meaning: [KEY MEANINGS]. Downstream node that will read the object: [DOWNSTREAM NODE]. Instruct the model to fill a key with [VERIFY] when the user payload lacks that value, and never to invent a key. Include one valid example object that uses placeholder strings, not real customer data. Keep the message under 160 words. Mention that extra keys will break [DOWNSTREAM NODE].

Write a System Message for an n8n AI Agent that must refuse work outside [ALLOWED TASKS]. If the user message asks for anything else, the agent returns a JSON object with keys status (refused) and reason, and calls no tools. Allowed tasks: [ALLOWED TASKS]. User message examples that must be refused: [REFUSAL EXAMPLES]. Do not invent a policy, a legal basis or a penalty. If [ALLOWED TASKS] is blank, put [VERIFY] in the allowed list and tell the agent to refuse everything until I fill it. Keep the message under 150 words.

Write a System Message that redacts secrets before any tool call. Secret patterns I care about: [SECRET PATTERNS]. Replacement token: [REDACTION TOKEN]. The agent must scan the user payload [PAYLOAD DESCRIPTION], replace matches with [REDACTION TOKEN], then call tools from [CONNECTED TOOL NAMES] only. If a pattern is too vague to match, write [VERIFY] beside it in the message and skip matching that pattern. Do not invent extra secret types such as passport numbers unless I listed them. Keep the message under 170 words.

Write a System Message that caps the agent at [MAX STEPS] tool calls. After that many calls, it must return whatever it has, with a key named stop_reason. If it still lacks [REQUIRED OUTPUT KEYS], those keys must be [VERIFY], not guesses. Purpose: [AGENT PURPOSE]. Connected tools: [CONNECTED TOOL NAMES]. Do not invent a loop, a memory store or a second agent. Tell it that repeating the same tool with the same arguments is wasted work and must stop. Keep the message under 140 words.

Write the instruction text for an HTTP Request tool used by my n8n AI Agent. Allowed methods and paths: [ALLOWED METHODS AND PATHS]. Base host: [BASE HOST]. The agent may fill query or body keys only from [ALLOWED PARAMETER NAMES]. If it needs a parameter I did not list, it must not call the tool and must return [VERIFY] instead. Do not invent a header, a path segment or a status code meaning. Include one example request using placeholders, then one line on what to do when the response is empty.

Write the tool description for a Call n8n Workflow tool. Workflows the agent may start: [WORKFLOW NAMES]. Input fields each workflow expects: [EXPECTED INPUT FIELDS]. Instruct the agent to pass only those fields, using names exactly as written. If a required field is absent from the user payload, return [VERIFY] and do not start the workflow. Do not invent a workflow id, a wait node or an extra input. Add one sentence on what the parent agent should do with the child workflow's returned JSON.

Write instructions for a Google Sheets or similar write tool. Spreadsheet: [SPREADSHEET NAME]. Tab: [TAB NAME]. Columns the agent may write, in order: [COLUMN NAMES]. The agent must emit one row as an array matching that order. Empty cells it cannot fill from [SOURCE PAYLOAD DESCRIPTION] must be the text [VERIFY]. Do not invent a column, a formula or a row number. If the payload has extra keys, drop them and list the dropped keys in a sidecar field named unused_keys.

Write the instruction that sits with my structured output parser. Required keys: [REQUIRED KEYS]. Optional keys: [OPTIONAL KEYS]. Types I expect (string, number, boolean, or array of strings): [KEY TYPES]. Tell the model to output JSON only. Required keys with no source in [USER PAYLOAD] must be the string [VERIFY], not null and not an invented value. Optional keys with no source must be omitted. Do not invent a nested object I did not define. Include one valid and one invalid example, both using placeholders.

Write retry rules for tools on this agent. Tools: [CONNECTED TOOL NAMES]. Retry only when the tool returns [RETRYABLE SIGNALS]. Maximum retries: [MAX RETRIES]. After that, return JSON with keys failed_tool, last_error_text and next_human_step. Copy last_error_text from the tool; if it is empty, write [VERIFY]. Do not invent an HTTP status, a timeout or a queue name. Do not retry on [DO NOT RETRY SIGNALS]. Keep the whole instruction under 160 words so it fits in the tool description field.

Dry-run this System Message against one sample without calling tools. System Message: [PASTE SYSTEM MESSAGE]. Sample user payload: [SAMPLE PAYLOAD]. Output a table with columns Step, Tool it would call, Arguments it would pass and Risk. If the message would require a tool I did not connect ([CONNECTED TOOL NAMES]), mark that row as blocked. If an argument is not in the sample, write [VERIFY] in Arguments. Do not invent a tool result. End with a one-sentence go or no-go for production.

Audit this agent output for keys that do not exist on the input. Input: [PASTE INPUT]. Output: [PASTE AGENT OUTPUT]. Allowed output keys: [ALLOWED KEYS]. List every output key. Mark each as copied, derived with evidence, or invented. For invented keys, quote the value and replace it with [VERIFY] in a rewritten object. Do not add new keys while rewriting. If a derived value has no evidence in the input, treat it as invented. Keep the rewritten object to the allowed keys only.

Decide whether this agent result is safe to send to a write node. Result: [PASTE RESULT]. Write node: [WRITE NODE NAME]. Require these keys to be present and not [VERIFY]: [MUST BE PRESENT KEYS]. If any required key is missing, tagged, or empty, output action = hold and list the gaps. If all required keys are present, output action = write and a 20 word summary of what will be written. Do not invent a default for a missing key. Do not lower the bar because the rest of the object looks complete.

Turn these production examples into a golden case table I can replay. Examples: [PASTE EXAMPLES]. For each example write a row: Input summary (under 20 words), Expected category or owner from [EXPECTED LABELS], and Notes. If an example has no expected label, write [VERIFY] in that cell. Do not invent an expected label. Limit the table to 10 rows. After the table, name two cases that would catch a regression if the System Message drifts, and say which field to watch.

Compare the agent on the same inputs before and after my prompt edit. Old System Message: [OLD MESSAGE]. New System Message: [NEW MESSAGE]. Shared test inputs: [TEST INPUTS]. For each input, write a row: What stayed the same, What changed, and Whether the change is intended given [INTENDED CHANGE]. If an output field appears in the new run but not in the input, mark it [VERIFY] as a possible invention. Do not invent a pass or fail score. End with keep, revert, or edit again, in one sentence.