Claude Prompt Library

Claude MCP Prompts With Stop Lines Before Destructive Calls

15 copy-paste prompts

Brief MCP servers and tool use in official Claude: allowed tools, forbidden actions, and [STOP] before delete, pay, or send. Unknown scopes marked [VERIFY].

In short: This page contains 15 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.

Louis Corneloup
By Louis Corneloup · Founder, Techpresso
Last updated ·Hand-curated & tested by the AI Academy team

Brief the MCP server

3 prompts

Server capability card

1/15

✨ What it does

Claude writes a capability card for [NAME] from [DOCS PASTE] with risks per tool and [VERIFY] on unknowns. Read before you enable the server in Claude.

I connect MCP server [NAME] with tools [LIST]. Write a capability card: what each tool does, inputs, outputs, and risks. Use only [DOCS PASTE]. Unknown tools [VERIFY].

💡

Pro tip: Paste official tool docs. Claude should not guess tool behavior.

Allow and deny list

2/15

✨ What it does

Claude builds allow/deny tool lists for [JOB] with a stop line before writes. Configure Claude to match the deny list in your MCP setup.

For job [JOB], list MCP tools allowed, tools forbidden, and tools allowed read-only only. Include stop line before any write or delete. Repo facts from [NOTES] or [VERIFY].

💡

Pro tip: Deny lists beat please be careful every time.

Human approval gates

3/15

✨ What it does

Claude defines APPROVE [ACTION] gates for send, pay, delete, and deploy on [SYSTEM]. Use those gates in every session with that MCP server.

Write approval gates for MCP actions on [SYSTEM]: which actions need me to type APPROVE [ACTION] before you call the tool. Cover send, pay, delete, deploy.

💡

Pro tip: Approval tokens should be specific, not one generic yes.

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.

Start 7-Day Free Trial

Safe tool calls

3 prompts

Read-only investigation

4/15

✨ What it does

Claude plans read-only MCP tool calls for [QUESTION] and stops if a write is required. Approve write tools separately if investigation needs them.

Using MCP read tools only on [SYSTEM], investigate [QUESTION]. List tools you would call in order, expected outputs, and stop if a write tool is required.

💡

Pro tip: Investigation prompts should default read-only until proven otherwise.

Dry run JSON args

5/15

✨ What it does

Claude shows JSON args for [TOOL] without calling it and flags risky fields [VERIFY]. Confirm args before you allow the real call.

For tool [TOOL] on MCP server [NAME], show JSON arguments you would send for task [TASK] without calling it. Flag risky fields [VERIFY]. Ask me to confirm.

💡

Pro tip: Dry run JSON catches wrong IDs before damage.

Rollback plan before write

6/15

✨ What it does

Claude writes rollback steps before [TOOL] write and waits for APPROVE WRITE. Do not skip rollback on production resources.

Before MCP write tool [TOOL] on [RESOURCE], write rollback steps if the write fails or overwrites. Stop at [STOP] until I paste APPROVE WRITE.

💡

Pro tip: If you cannot rollback, the write should not run.

Multi-tool workflows

3 prompts

Three-step workflow map

7/15

✨ What it does

Claude maps a 3-step MCP workflow for [JOB] with failure handling and [VERIFY] on unknown shapes. Run step 1 alone before chaining all three.

Map a 3-step MCP workflow for [JOB]: tool per step, data passed between steps, failure handling. No extra tools. Unknown data shapes [VERIFY].

💡

Pro tip: Chain workflows only after each step succeeds once.

Idempotency check

8/15

✨ What it does

Claude marks retry-safe vs unsafe steps in [PASTE] using [DOCS]. Add waits before retry on non-idempotent steps.

For workflow [PASTE], mark steps that are safe to retry vs steps that are not idempotent. Add wait rules before retry. Official MCP semantics only from [DOCS].

💡

Pro tip: Payment and send tools are rarely idempotent. Treat them as once-only.

Log line template

9/15

✨ What it does

Claude writes a post-call log template without secrets. Paste each log into your ticket system so you keep an audit trail.

Write a log template I should paste after each MCP call: timestamp, tool, resource id, result summary, errors. No secrets in logs.

💡

Pro tip: Logs beat memory when MCP calls fail at night.

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.

Try AI Academy Free

Failure and abuse

3 prompts

Tool error decode

10/15

✨ What it does

Claude decodes [PASTE] error for [TOOL] using [DOCS] and suggests a safe read-only next step. Fix root cause before retry.

MCP tool [TOOL] returned error [PASTE]. Explain likely causes using [DOCS]. Suggest next safe read-only step. Do not retry destructive call automatically.

💡

Pro tip: Automatic retry on delete tools is how accidents repeat.

Scope creep refusal

11/15

✨ What it does

Claude refuses unrelated [UNRELATED TASK] with admin tools and offers read-only alternatives. Keep sessions scoped to [JOB].

Act as Claude with MCP [NAME]. I asked you to [UNRELATED TASK] using admin tools. Refuse if outside [JOB]. Offer read-only alternative or human handoff.

💡

Pro tip: Scope creep starts with innocent just this once tool calls.

Secret redaction in tool output

12/15

✨ What it does

Claude redacts secrets from [PASTE] and lists redactions with [VERIFY] uncertain spans. Rotate any key that leaked into chat.

Review MCP output [PASTE] for secrets before you show me. Redact tokens and keys. List redactions made. Uncertain substrings [VERIFY].

💡

Pro tip: Assume tool output can contain secrets until proven otherwise.

Stop before destructive

3 prompts

Delete confirmation stop

13/15

✨ What it does

Claude stops at [STOP] before delete, lists impact and backups, and waits for APPROVE DELETE. Type the phrase only if you accept impact.

I asked to delete [RESOURCE] via MCP. Stop at [STOP]. List what delete affects, backups to check, and exact APPROVE DELETE phrase I must type.

💡

Pro tip: Never paraphrase the approve phrase. Exact match only.

Payment send stop

14/15

✨ What it does

Claude drafts send/payment payload for review and refuses until APPROVE SEND. You send from the official UI if unsure.

I asked to send payment or email via MCP tool [TOOL]. Stop before call. Draft payload for review only. Refuse to send until APPROVE SEND.

💡

Pro tip: MCP send tools are high risk. Default stop every time.

Prod deploy stop

15/15

✨ What it does

Claude plans prod deploy preflight from [RUNBOOK] and stops at [STOP] before deploy tools. Complete [VERIFY] checks before any APPROVE DEPLOY.

Plan deploy to production via MCP [NAME]. Stop at [STOP] before any deploy tool. List preflight checks from [RUNBOOK]. Missing checks [VERIFY].

💡

Pro tip: Deploy prompts belong with runbooks, not improvisation.

Free tool

Prompt Optimizer

Turn a rough idea into a structured, professional AI prompt.

Try it free →

Frequently Asked Questions

No. These prompts stay inside Claude's official product. It stays inside official Claude MCP integrations. If a setting is missing on your plan, the prompt says so.
A current Claude account. Some features sit on paid plans. The prompts still work as checklists if a feature is off.
Every prompt tells it to use only what you paste and to tag gaps as [VERIFY]. You still check the MCP tool plan before you save it.
Yes to copy. Running them needs your Claude account. Long files fit better on a paid plan.
No. It is a draft of your MCP tool plan. You own the facts, the files, and whatever your job or policy requires.

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.