AgentDirective

Walkthrough

From an idea to a clear build brief.

See how an idea becomes a brief for your coding agent, then a record of what it reports building. Work at your own pace and leave unanswered decisions open.

The whole flow

This illustrated example is not a result for your project. Use Next or choose a step; the animation is optional.

  1. Step 1

    Say what you are building

    Write it the way you would explain it to a colleague. No account, repository, source code, or special vocabulary. Prefer not to write? Pick what you are building gives you a short menu.

    Click Review my build

    We check the local library first. The optional Cloudflare AI check runs only when it finds nothing.

    agentdirective.org

    What are you asking your AI to build?

    A support tool that reads customer emails and automatically issues refunds.
    Review my build →Pick what you are building

  2. Step 2

    Check what it found

    The local check uses versioned rules to suggest project traits. Remove anything that does not apply and add anything we missed. Your confirmed choices determine which safeguards match.

    Click Show decisions

    Correct anything here — your corrections drive which safeguards apply.

    Confirm understanding

    We found 4 traits

    • ✓AI component reads customer content
    • ✓Untrusted input from customers
    • ✓Money movement — refunds
    • ✓Customer data
    Show decisions →

  3. Step 3

    See which safeguards apply, and why

    Your confirmed choices select the safeguards through published rules, not a model. Each one shows why it applies.

    The same confirmed traits, matching rules, and pack versions produce the same matches.

    Safeguards that match
    AUT-001 · Bounded AutonomyMatches because an automated action moves money.
    AI-001 · Untrusted Content and Tool BoundariesMatches because a model reads content it did not author.
    OBS-001 · Traceable Consequential ActionsMatches because a consequential action runs without a person present.

  4. Step 4

    Answer only what needs you

    Set limits, approvals, and who may do what. Answer what you know. Leaving one open is valid—the directive tells your agent to ask you before choosing a value.

    Click Review my answers

    An unanswered decision is never recorded as confirmed.

    Decisions · 1 of 6

    What is the maximum refund an automation may issue without human approval?

    • Any amount
    • Up to a set limit — $500
    • Not decided yet
    Review my answers →

  5. Step 5

    Take the result to your agent

    Copy the prompt or give your agent the complete bundle. Open decisions are marked as questions it must ask you. If no safeguards match, you get a no-match assessment, not an empty directive.

    Click Copy agent prompt

    Copying sends nothing new — the prompt is built from the compiled result in your browser.

    Draft hand-off · 2 decisions open
    Build rulesagent-directive.jsonReadable briefAGENTDIRECTIVE.mdBuild sequenceimplementation-plan.mdProgress recordcontrol-status.json
    Copy agent promptDownload bundle

  6. Step 6

    Your agent builds it, and reports back

    Paste the prompt into your coding agent. Ask it to confirm the boundaries, resolve open decisions with you, and record its work. Check that it follows these instructions; the directive cannot enforce them.

    The optional review checks the supplied report for evidence gaps and inconsistencies. Code review and runtime tests remain separate.

    Your coding agent
    “Before I start: refunds are capped at $500 without approval, and every refund must be logged with the acting account. Two decisions are still open — who may approve above the cap, and how long logs are kept. How do you want those?”
    implemented · 7partial · 2unresolved · 2

Ready when you are

Try it with your own description.

Start with a high-level idea. Review the matches, answer what you know, and keep the remaining decisions open.

Start a review Not sure how to describe it?