AgentDirective

Before you build

Set the boundaries before your AI builds.

Describe your project. Decide who can do what, which limits apply, and what happens when something fails. Get a clear brief to give your coding agent.

Watch how it works Not sure how to describe it?

Already have a working project?

Review and improve your project

Try an example

Start with one click.

Each example starts a browser-local review. Nothing is compiled until you choose to create the files.

Keep it high-level. Do not paste secrets, customer records, credentials, or source code. We check the local library first, then show you what matched.

Optional settingsWork context and optional AI suggestions

About you (optional)

Choose the setting closest to your work. It does not affect the review and is counted only as a group.

No account or repository access. Matching stays in your browser. Creating files sends your description and answers to Cloudflare for this request; they are not stored. Optional AI has separate consent.

What you get

Ready for your agent
  • Build instructions.agentdirective/agent-directive.jsonStructured rules your coding agent can follow.
  • Readable project brief.agentdirective/AGENTDIRECTIVE.mdA plain-language version for people and agents.
  • Implementation progress.agentdirective/control-status.jsonWhat your agent reports as built, incomplete, or still open.
Clear instructions, not a vague score.

Your agent gets the boundaries to implement, available validation scenarios, and a structured way to report progress.

Why clear rules matter

A working refund app can still break an approval rule. Explore an illustrated comparison of an incomplete brief and a tested boundary.

Read the refund safeguard

The film · under a minute

Watch an idea become a hand-off.

One build idea, from a plain sentence to a brief your coding agent follows and reports back against.

A short film showing one build idea move through AgentDirective. A text version follows.
Read the film as text
  1. Describe. Type the build idea in plain words, such as an AI support tool that reads customer emails and automatically issues refunds. No account, repository, or source code.
  2. Detect. Published rules find what the system does: an AI component, outside input, customer data, and money movement. You correct anything that is wrong.
  3. Match. Your confirmed traits select versioned Control Packs, such as AUT-001, AI-001, and OBS-001, and each one says why it applies.
  4. Decide. You answer the questions that matter, such as a $500 refund limit before a person must approve. Leaving a decision open is a valid answer, and your agent must ask rather than guess.
  5. Hand off. Your agent gets the boundaries, the open questions, and the files it reports into. It restates what it must follow and asks about anything still open.
  6. Report back. Your agent records what it implemented, what is partial, and what is still unresolved. Reported is not verified, and the report says so.
Try it with your own idea

How it works

From an idea to a clear hand-off.

Six steps, in order. Scroll to move through them.

  1. 01 / Describe

    Tell us what you want to build.

    No account, repository, or source code. Just the description you were going to give your coding agent.

    What are you asking your AI to build?

    Build a support tool that reads customer emails and automatically issues refunds.

    Review my build Local matching; optional AI is separate.
  2. 02 / Detect

    We identify what the system does.

    Published rules look for known traits. You see the list and can correct it before continuing.

    Traits found in your description
    AI componentOutside inputMoney movementCustomer data
    Every trait is yours to correct or remove.
  3. 03 / Match

    Your confirmed traits select the safeguards.

    Published rules decide what applies, and each pack explains why it was selected.

    Safeguards that apply
    • AUT-001
      Bounded AutonomyBecause money can move without a person.
    • AI-001
      Untrusted ContentBecause a model reads text you did not write.
    • OBS-001
      Traceable ActionsBecause someone will have to answer for a refund.
  4. 04 / Decide

    You answer the questions that matter.

    Limits, approvals, and scope stay open until you confirm them. Your agent must ask about any value you leave open.

    AUT-001-Q1 · Threshold

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

    $500Any amountLeave open — a valid answer
    An open decision is recorded as open, and your agent is told to ask.
  5. 05 / Hand off

    Your agent gets clear instructions.

    A directive it can implement, plus the files it uses to report what it built — and what it could not.

    .agentdirective/
    • agent-directive.jsonThe boundaries, as structured rules.
    • AGENTDIRECTIVE.mdThe same thing, for a person to read.
    • control-status.jsonWhere your agent reports what it built.
  6. 06 / Report back

    Your agent records what it built.

    It reports what it implemented, what it could not, and the evidence behind each claim. Reassess the same project later against its directive.

    control-status.json
    • implemented
      APPROVAL_THRESHOLD2 tests named, evidence pointed at.
    • partial
      ACTION_TRAILLimits and remaining work recorded.
    • unresolved
      EXPORT_SCOPEStill yours to decide.
    Reported is not verified, and the report says so.

See the full walkthrough →

Why it helps

Make the important decisions yourself.

A coding agent may make assumptions when a prompt leaves something out. Review the choices that affect authority, data, and failure handling before implementation.

Where to start

Start where you are.

No account, no repository access, no source code. Pick whichever matches where you are.

Describe a build Paste the prompt you were about to give your coding agent and review it before anything is built. Go to the description box Pick what you are building Choose a project type and its main actions. Then review and correct the suggested traits. Open the guided start Review a project you already have Create an inspection prompt for your agent, or bring a previous directive to review what changed. Review an existing project

What this is not

Design guidance, not a security guarantee.

AgentDirective prepares instructions and checks reported metadata. You still need code review and testing to find out whether the implementation follows those instructions.