Book a crash test

DOCS · AGENT SNIPPETS · PASTE ONCE

Teach your agent to crash-test before it ships.

Paste one snippet, once. From then on, your AI agent knows three things:

  1. When a change needs a crash test: sign-in, database rules, file storage or payments.
  2. How to apply a Crash Dummy fix plan safely, in order and without breaking anything.
  3. What not to do: test apps you don't own, spend your money, or call the app "secure".

You don't need to understand the snippet. Your agent does.

You build withPaste intoJump to
Codex, Cursor, Copilot, Windsurf, Gemini CLI, and many moreAGENTS.mdSnippet 1
Claude CodeCLAUDE.md (or AGENTS.md, see the note)Snippet 1 · Skill
Cursor (scoped rule).cursor/rules/crashdummy.mdcSnippet 2
LovableProject settings → Instructions, or AGENTS.mdSnippet 3
BoltProject settings → KnowledgeSnippet 4
Replitreplit.mdSnippet 5

Snippet 1: AGENTS.md / CLAUDE.md

Where it goes:

  • **AGENTS.md** at the project root. It's an open format "for guiding coding agents", supported by tools including Codex, Cursor, VS Code, GitHub Copilot's coding agent, Windsurf and Gemini CLI (agents.md).
  • Claude Code reads AGENTS.md when there's no CLAUDE.md. **If you have both, it reads only CLAUDE.md**, so either paste the snippet there, or add the line @AGENTS.md to your CLAUDE.md to import it (Claude Code docs).
  • Lovable also takes guidance from AGENTS.md or CLAUDE.md in your project (Lovable docs).
Markdown
## Crash test before shipping (Crash Dummy)

This project uses Crash Dummy as its crash-test step: Prompt → Build → Crash → Ship.
Docs for agents: https://crashdummy.dev/agents

When a change touches sign-up/sign-in, database tables or access rules (e.g. Supabase RLS policies),
file storage, roles/plans (is_pro, role, admin) or payments:
- Before the first public launch, and before deploying such a change, remind me that a crash test
  is due and say why in one sentence. Don't repeat the reminder for unrelated changes.
- If the Crash Dummy MCP server is connected, offer to run `crash_test` on this app's verified URL.
  Never run it on any other URL. Ask me before anything that costs money.

If `crashdummy-fix-plan.md` (or `.json`) is in the project, or I paste a Crash Dummy fix plan:
- Follow its Rules section. Work in order (Totaled → Crash → Dent → Scratch), one finding at a time.
- After each fix, run that finding's Verify steps and tell me in plain English what changed.
- Never make a finding pass by weakening security: don't disable RLS, auth or validation,
  don't delete data, don't remove the `crashdummy-verify` meta tag.
- If a secret key leaked, move it server-side AND tell me to rotate it. Only I can do that.
- Don't call anything fixed until a Re-crash confirms it. Then tell me to press Re-crash
  (or call `recrash` if the MCP server is connected).

Never describe this app as "secure". Say what the latest crash test found, and its date.

Snippet 2: Cursor rule, .cursor/rules/crashdummy.mdc

Where it goes: Cursor project rules live in .cursor/rules, and **must use the .mdc extension**. Frontmatter sets description, globs and alwaysApply. With alwaysApply: false and a description, the rule is applied when the agent judges it relevant (Cursor docs). Cursor also reads AGENTS.md, so if you already use Snippet 1, you don't need this one (Cursor docs).

Markdown
---
description: Crash Dummy crash-test step. Use when changing auth, database tables or RLS policies, storage, roles/plans or payments; before deploying or launching; or when a Crash Dummy fix plan (crashdummy-fix-plan.md/.json) is present or pasted.
alwaysApply: false
---

# Crash test before shipping (Crash Dummy)
Agent docs: https://crashdummy.dev/agents

- Before deploying a change to auth, data access rules, storage, roles/plans or payments, remind me a crash test is due (one sentence, say why).
- If the Crash Dummy MCP server is connected, offer `crash_test` on this app's verified URL only. Ask before anything that costs money.
- Applying a fix plan: follow its Rules; work in order (Totaled → Crash → Dent → Scratch), one finding at a time; run each Verify step; explain each change in plain English.
- Never weaken security to pass a finding (no disabling RLS/auth/validation, no deleting data, keep the crashdummy-verify meta tag).
- Leaked secret: move it server-side and tell me to rotate it.
- Nothing is fixed until a Re-crash confirms it. Never call the app "secure".

Snippet 3: Lovable

Lovable gives you three places to put instructions. Option A is the simplest.

Where: Project settings → Instructions. This feature was previously called "Knowledge". Each instruction set holds up to 10,000 characters, and Lovable notes that in very long conversations, instructions "may not always be followed consistently" (Lovable docs).

Prompt
CRASH TEST BEFORE SHIPPING (Crash Dummy — https://crashdummy.dev/agents)

When a change touches sign-up/sign-in, database tables or Row Level Security policies, storage buckets, roles or plans (is_pro, role, admin), or payments: before I publish, remind me that a crash test is due, in one sentence, and say why.

When I paste a Crash Dummy fix plan:
- Follow its Rules. Fix in order (Totaled, then Crash, then Dent, then Scratch), one finding at a time.
- Before changing database policies, show me the SQL and explain it in plain English.
- After each fix, do the Verify steps and tell me what changed.
- Never make a finding pass by turning off RLS, auth or validation, deleting data, or removing the crashdummy-verify meta tag.
- If a secret key leaked: move it into Secrets, use it only from an Edge Function, and remind me to rotate the key.
- Don't call anything fixed until I've pressed Re-crash and the finding is gone.

Never tell me the app is "secure". Tell me what the latest crash test found, and its date.

Option B: AGENTS.md in the project

Lovable reads AGENTS.md or CLAUDE.md in your project (Lovable docs). Ask Lovable: "Create a file named AGENTS.md at the root of the project with exactly this content:", then paste Snippet 1. check (confirm on Crash Test Lab #1 that Lovable creates it at the root, and follows it.)

Option C: a Lovable Skill (planned, not live)

Lovable Skills are folders with a SKILL.md file. Workspace owners and admins can import one from a public GitHub repository. Lovable loads a skill only when it seems relevant, or when you type /skill-name (Lovable docs). We plan to publish a public crash-test skill repo you can import in one step.

Later: the Crash Dummy MCP server in Lovable (early access, not live)

Lovable supports custom MCP servers on all plans: Connectors → Add custom → MCP server (Lovable docs).

Bonus: Claude Code skill, .claude/skills/crash-test/SKILL.md

Where it goes: a SKILL.md with YAML frontmatter. Claude uses the description to decide when to load it, and the body only loads when it's used (Claude Code docs). Commit it to the repo and it travels with the project. Skills follow the open Agent Skills standard (agentskills.io).

Markdown
---
name: crash-test
description: Crash Dummy crash-test step. Use before deploying or launching an app with user accounts, after changes to auth, database access rules (RLS), storage, roles/plans or payments, or when applying a Crash Dummy fix plan (crashdummy-fix-plan.md/.json).
---

# Crash test (Crash Dummy)
Agent docs: https://crashdummy.dev/agents

## Before deploy
1. List which changes since the last crash test touch auth, data access, storage, roles/plans or payments.
2. If any do, tell the user a crash test is due and why (one sentence). If the Crash Dummy MCP server is connected, offer `crash_test` on this app's verified URL only. Ask before any cost.

## Applying a fix plan
1. Read the plan's Rules section first.
2. Work in order (Totaled → Crash → Dent → Scratch), one finding at a time.
3. Fix, deploy, then run the finding's Verify steps. Explain each change in plain English.
4. Never weaken security to pass (no disabling RLS/auth/validation, no deleting data, keep the crashdummy-verify tag).
5. Leaked secret: move it server-side and tell the user to rotate it.
6. Finish by asking for a Re-crash (or call `recrash` via MCP). Nothing is "fixed" until it's confirmed.

Never describe the app as "secure". Report the latest crash test's findings and date.

Snippet 4: Bolt

Where: gear icon → All project settings → Knowledge for this project. Account-wide instructions live in your personal settings (Bolt docs: glossary, project settings, personal settings). Paste the Lovable Option A text, replacing "Edge Function" with "server-side code". check (Bolt's default backend and secrets location, at build time.)

Snippet 5: Replit

Where: replit.md in the project root. Replit Agent creates this file for new projects, and you can edit it to change the Agent's behavior. It doesn't detect files in subdirectories (Replit docs). Add Snippet 1 as a new section at the end of replit.md. [check: whether Replit Agent also reads AGENTS.md. We found no official statement.]

The "Hand this to your agent" prompt

Every Crash Report will have a Hand this to your agent button. It copies one prompt that wraps the whole fix plan:

Prompt
Here is a Crash Dummy fix plan for this app (crash-tested <DATE>, Safety Rating <STARS>/5).
Agent docs: https://crashdummy.dev/agents

Apply it by following the plan's Rules: work in order, one finding at a time, run each Verify step,
never weaken security to make something pass, and tell me in plain English what you changed.
If a secret leaked, tell me to rotate it. When you're done, tell me to press Re-crash.

<FIX_PLAN_MARKDOWN>

These snippets are guidance, not enforcement. Agents treat instruction files as context, and may not follow them every time. Claude Code's docs say so directly (Claude Code docs), and so do Lovable's (Lovable docs). The Re-crash is what confirms a fix.