← Back to library

Require the finished work plus the evidence tape

Carnage's article stresses that agents should validate work with screen recordings and receipts, not summaries alone. Demand the deliverable together with proof, and drop any number the bot cannot show how it got. We keep human approval on outward sends.

  • OutcomeMake every agent delivery include the artifact, recordings or screenshots, linked sources, guesses called out, and skips with reasons.
  • SetupMinutes
  • LevelStarter
  • CreditCarnage (@0xCarnagee)
View original
Works with
Grok BotChatGPT agentClaudeClaude Cowork

Published Oct 23, 2026Updated Oct 23, 2026AI dry run on sample input: passed Oct 23, 2026Source verified Oct 24, 2026

When should you use it?

  • You are about to hand a bigger job to an agent
  • Numbers keep appearing without sources
  • You want to stop re-checking every click yourself

Not for: Accepting a polished summary with no artifacts; Letting the agent invent screenshots or links; Skipping the guesses list because the answer sounded sure

What do you need, and what do you get?

You need

  • The job and the done artifact you expect
  • Whether recordings or screenshots are available
  • Which metrics must be sourced

You get

  • The finished artifact
  • Evidence pack: recording or screenshots, source links, guesses, skips
  • A clear call-out of anything omitted for lack of proof

How does it work, step by step?

  1. Define doneName the artifact you will open, not just a status update.
  2. Demand receiptsRequire recording or screenshots, linked sources, separate guesses, and skips with why.
  3. Kill unsourced numbersIf it cannot show how a number was produced, the number stays out.
  4. Review the tapeWatch or skim evidence before you widen scope.

The prompt

Our own prompt, inspired by Carnage (@0xCarnagee) and written to work in Grok Bot, ChatGPT agent, Claude and others. Values in lime are placeholders.

Fill in the blanks (optional; the copied prompt uses your values)

Prompt2 variables
Complete this job and show your work.
Job: {job}.
Done artifact: {artifact}.

Return:
1) The finished artifact.
2) Evidence: a capture or still frames covering the steps you took.
3) Exact source link for every number.
4) Anything you guessed, listed separately.
5) Anything you bypassed, with the reason.

If a figure has no proof trail, omit the figure.
Do not invent evidence.

Inspired by Carnage (@0xCarnagee): X Article: The Grok Bot Team's Own Workflow (11 steps). The prompt above is written by playbooks.bot in our own words; nothing from the original is quoted.

What can go wrong?

  • A summary without evidence trains you to trust claims you cannot audit.
  • Guesses mixed into facts are worse than an explicit guesses list.
  • Skipped steps need reasons; silence looks like completeness.
  • Do not ask the agent to fabricate a recording it did not capture.

Our take

Trust collapses when agents deliver confident summaries without receipts. Carnage's tape rule is the fix: finished work plus screen recordings, linked sources, guesses and skips called out. Defining done as an artifact you can open, rather than a status update, is the first step. Then ask for proof: a screen capture or still images, source links, a separate list of guesses, and any skipped steps with the reason. Mixing guesses into facts is worse than listing them honestly, and silence about a skipped step looks like completeness. If the bot cannot show how a number was produced, leave the number out. Review the evidence before widening the agent's scope; that is how trust is earned rather than assumed. Never ask an agent to recreate a recording it did not capture, since that would be fabricated evidence. The extra minutes spent skimming a tape cost far less than acting on a polished summary that turns out to be wrong.

Where the time goes back: You review receipts instead of redoing the work, so bigger jobs can move earlier without blind trust.

FAQ

What counts as evidence?
Step captures or still frames, exact sources for numbers, a guesses list and a skips list. Not a paragraph saying it went well.
When can I stop watching every run?
After repeated clean tapes on the same job. Widen scope gradually; do not skip evidence on new jobs.
What if a number lacks a source?
Leave it out. The article's rule is that unciteable numbers do not get reported.

Source and credit

Inspired by Carnage (@0xCarnagee). We describe the idea in our own words, link to the original, and wrote the prompt ourselves; nothing is quoted.

Inspired by
X Article: The Grok Bot Team's Own Workflow (11 steps)
Creator
Carnage (@0xCarnagee)
Published on
X
Original date
Aug 19, 2026
License
Not openly licensed (the poster keeps all rights), so we quote nothing and link to the original
Originally for
Grok Bot
AI dry run
Oct 23, 2026: passed. Dry-run tested by our AI test agent on sample input. Evidence pack complete; unsourced number dropped.
AI dry run on sample data; not run with live connected accounts. The playbooks.bot test agent filled the prompt with synthetic inputs, produced the output and checked it against the page's stated outputs and guardrails.
Verified
Oct 24, 2026: source link resolves to the original, the credited creator published it, our excerpt or prompt matches it, and the license allows our use.

Are you the creator? Claim or correct this page.

More in Operations & Admin

All 203 Operations & Admin playbooks →

Related playbooks

Operations & AdminPrompt✓ AI dry run

Hand off one offline task before you sign out

Before you go offline, get up to three useful task options with a concrete done artifact named for each, pick one, and wake up to a review-ready deliverable the agent prepared after your choice.

Credit: Scotty Beam (@ScottyBeamIO)

Minutes
Operations & AdminPrompt✓ AI dry run

Aim the agent at a janky internal app after it maps the UI

Have the agent study a janky internal UI out loud, complete one end-to-end task with a recording, and stop for login walls instead of guessing credentials.

Credit: Carnage (@0xCarnagee)

Under an hour
Operations & AdminPrompt✓ AI dry run

Brief an agent with nine fields, not a how-to

Before you hand a job to an agent, fill in nine short fields that name the result, the sources it may use, the deadline, what it may do alone, what needs your say-so, what it must never do, how to prove the work is finished, what to do when it gets stuck, and where to put the deliverable, then add a line that tells it to pick the simplest reliable method.

Credit: Michael Fenech (@Michael_Fenech_)

Minutes
Operations & AdminWorkflow✓ AI dry run

Have a desktop agent learn your folder of work scripts

Point a desktop agent at the folder of small scripts you rely on and get a catalog it can use, so you can run them by asking in plain words.

Credit: Tomasz Tunguz (@ttunguz)

Under an hour
Operations & AdminWorkflow✓ AI dry run

Run an overnight research team with handoff notes

Split research work into five agent roles that read, reproduce, reanalyze, scout and write, with one shared brief and human approval for anything outward-facing.

Credit: flash (@flasssh_ai)

An afternoon+