intent-contract

name	intent-contract
description	Explicitly invoked skill for producing an implementation/evaluation/message contract. Hard rule: never output a final Intent Contract before the user's 3rd message in the skill interaction.
triggers	["$intent-contract","/intent-contract","intent-contract"]

Intent Contract

This skill turns a user request into a compact final contract/prompt with goal, scope, decisions, acceptance criteria, verification, and handoff.

Non-negotiable rule: hard 3-user-message minimum

After this skill is invoked, do not output a final # Intent Contract: until the user has sent at least 3 messages in this intent-contract interaction.

Only user messages count.

Do not count:

• assistant messages,
• tool calls,
• file reads,
• searches,
• shell commands,
• repository inspection,
• assistant summaries,
• assistant guesses,
• assistant-proposed criteria.

The minimum sequence is:

1. User message 1: initial $intent-contract ... request.
2. User message 2: user's clarification, answer, narrowing, correction, or added context.
3. User message 3: user's confirmation, revision, or final direction.

The final # Intent Contract: may be produced only after user message 3.

There are no implicit exceptions for:

• “enough information,”
• “simple request,”
• “obvious answer,”
• “tool inspection completed,”
• “the assistant can infer it,”
• “the user already gave good details in message 2.”

The only exception is when the user explicitly says:

skip clarification and write the contract now

or an equivalent direct instruction.

What to do before user message 3

Before user message 3, do not use the final contract template.

Use one of these lightweight formats instead:

After user message 1

• Restate the request in one sentence.
• Ask 1-3 material questions.
• Say Turn status: 1/3.

After user message 2

• Summarize the narrowed understanding.
• Optionally inspect relevant files/data.
• Propose candidate direction or draft text if useful.
• Ask for confirmation, correction, or final preference.
• Say Turn status: 2/3.
• Do not output # Intent Contract:.

After user message 3+

• Produce the final # Intent Contract: if no blocker remains.
• If blocked, produce a short blocked contract with Status: blocked.
• Say Turn status: 3/3 complete.

Scope

Use this skill only when the user explicitly invokes it.

If the user invokes it for a non-code task such as a message, email, explanation, or planning note, still obey the hard 3-user-message minimum. The final contract may be lightweight and may include a final draft/message.

If the user did not invoke this skill, do not force this workflow onto ordinary chat, writing, translation, debugging, QA, review, browser testing, shipping, documentation, or memory-maintenance tasks.

Workflow after user message 3

1. Collect decision-relevant context

   • Greenfield: goal, audience/user, scope, constraints, success criteria, risks.
   • Brownfield: README, package/config files, existing specs, user-named files, and obvious relevant files only.
   • Separate facts by source: [from-user], [from-code], [from-research], [assumption].

2. Ask material questions only

   • Prefer questions whose answers change architecture, API/protocol, ownership/SSOT, data model, migration, permissions, UX, rollout, evaluation threshold, tone, audience, or verification.
   • Mark premature/unknowable items as Decide later.

3. Score readiness when relevant

   • Greenfield ambiguity: 1 - (goal*0.40 + constraints*0.30 + success*0.30).
   • Brownfield ambiguity: 1 - (goal*0.35 + constraints*0.25 + success*0.25 + context*0.15).
   • Use 0.0-1.0 clarity scores. Target ambiguity <= 0.20.
   • Do not proceed if an unresolved decision would materially change the contract.

4. Produce final Intent Contract Include:

   • Goal
   • Scope and non-goals
   • User-message history
   • Context facts and assumptions
   • Decisions made
   • Open decisions blocking finalization
   • Decide-later items
   • Entities / domain model / ontology when useful
   • Constraints
   • Acceptance criteria
   • Verification plan
   • Final output/draft when relevant
   • Handoff
   • Clarity score / status

5. Exit cleanly

   • After producing Status: ready, Status: ready-with-deferrals, or Status: blocked, exit intent-contract mode unless the user explicitly asks to revise the contract.

Output format

Use references/intent-contract-template.md only after the hard 3-user-message minimum is satisfied or explicitly skipped.

Before user message 3, never print # Intent Contract:.

Validation checklist

Before final output, verify:

• User-message count is at least 3, or the user explicitly skipped clarification.
• The final output includes User-message status: 3-message-complete or explicitly-skipped.
• Every acceptance criterion is observable by test, inspection, metric, manual check, review, or user approval.
• Scope and non-goals prevent at least one plausible wrong output.
• Facts are distinguished from user decisions and assumptions.
• No open decision remains that would materially change the result unless Status: blocked.