/prove-it

Garden · not yet reviewed Beta, from the WDAI Build Cycle pilot Version 0.1.0

Checks an AI result against something real before you trust it or pass it on, using the source, the actual number, or a rerun with one thing changed, and says what evidence would change the answer.

How to call it

Once it's installed, ask for it by name, for example "use prove-it on this" or type /prove-it . Your AI tool also starts it without being asked when your request matches its description.

When it starts: someone is about to send, publish, decide on or reuse an AI output, asks "is this right?", "can I trust this?", "check this", "verify", or when an output looks polished and contains numbers, names, quotes or claims that matter.

Connections

Practises

• Check the result

Concepts it uses

• Verification (garden)
• Hallucination (garden)
• Evals (garden)

Comes after

• /build-brief (garden)
• practice-coach (garden)
• /expert-council (garden)

Try next

• /make-it-repeatable (garden)
• discernment-nudge (garden)
• evals-skills (garden)

Used in blueprints

• Build a style guide so AI drafts sound like you (garden)
• Ask AI to find where your draft misleads (garden)
• Compare AI tools on a task you actually do (garden)
• Write a handover brief before you give an agent a real task (garden)
• Hire and onboard an AI copilot for one piece of work (garden)
• Teach your AI assistant your team's jargon, one confirmed term at a time (garden)
• Turn messy notes into a status draft you can check (garden)
• Write a one-page build plan before you ask AI to make an app (garden)

From the field, after this

• discernment-nudge (garden)
• ChatPRD plugin (Cursor, Codex, ChatGPT) (garden)
• evals-skills (garden)
• draft-review-kit (garden)
• awesome-pm-skills (garden)
• no-ai-slop (garden)
• human-review (garden)

The skill itself

This is the instruction text your AI tool reads, shown as written.

Prove it

The more finished something looks, the less people check it. Help the person run a quick check against something real.

How this skill behaves

They predict or point first; the check uses something outside the AI, never the AI grading itself. “Just do it” runs the checks and reports.

Steps

1. Point to the risk. Ask which claim, number or step would hurt most if it were wrong. If they’re not sure, ask them to predict what a correct version should say before looking closely.
2. Pick an outside check. For each risky claim, choose the cheapest real check:
   • the source: open the document, page or notes it came from and find the line;
   • the real thing: open the file, read the actual number, click the link, run the formula;
   • a rerun: change exactly one thing, run the same input again, and compare;
   • a person: someone who would know.
3. Run it and show the evidence. Report what was checked, what matched, and what didn’t. Label anything you couldn’t verify as unverified rather than smoothing it over.
4. What would change the answer? Name the evidence that would overturn the result, so they know what to watch for.
5. Decide what still needs a person. Ask what they’ll check themselves or hand to someone before it goes out.

Output

A short check table (claim → how it was checked → result), anything still unverified, and what needs a human decision.

Next

For a workflow that will run again and again, point to the field skill evals-skills (Hamel Husain) for proper testing. For a light check after any substantive answer, recommend Anthropic’s discernment-nudge.

---

Credit: 'Change one thing, rerun, compare' from WDAI's Build Cycle; informed by Anthropic's AI Fluency Index and pstack principle-prove-it-works (MIT, Lauren Tan).

Licence: Apache-2.0 (instructions) and CC BY 4.0 (teaching text). From AI Practice Commons by Women Defining AI; method from the WDAI Build Cycle (beta).. Plain Markdown for agents: /skills/prove-it.md