skill-intent-contract

name	skill-intent-contract
description	Use when starting a complex or ambiguous task that risks scope drift
disable-model-invocation	true

Host: Codex CLI — This skill was designed for Claude Code and adapted for Codex. Cross-reference commands use installed skill names in Codex rather than /octo:* slash commands. Use the active Codex shell and subagent tools. Do not claim a provider, model, or host subagent is available until the current session exposes it. For host tool equivalents, see skills/blocks/codex-host-adapter.md.

Intent Contract System

Purpose

The intent contract creates a persistent record of user intent that:

• Captures what the user is trying to accomplish
• Defines success criteria upfront
• Establishes boundaries and constraints
• Travels through the entire workflow
• Validates final outputs against original intent

This closes the loop between intention and delivery.

Intent Contract Structure

The intent contract is stored in .claude/session-intent.md and follows this format:

# Intent Contract

**Created**: [ISO timestamp]
**Workflow**: [discover/embrace/review/etc.]
**Status**: [active/validating/completed]

## Job Statement
What the user is trying to accomplish (JTBD framework).

[User's goal in plain language]

## Success Criteria

### Good Enough
- [Minimum viable success criterion 1]
- [Minimum viable success criterion 2]

### Exceptional
- [Excellence criterion 1]
- [Excellence criterion 2]

## Boundaries
What this should NOT be:
- [Boundary 1: What to avoid]
- [Boundary 2: What's out of scope]

## Context & Constraints

**Stakeholders**: [Who needs this to work for them]
**Existing Assets**: [What to build on]
**Timeline**: [Time constraints if any]
**Technical Constraints**: [Platform, language, dependencies]

## Clarifying Context
[Any answers from the 3-question pattern]

## Task Allocation
**Risk**: [low | intermediate | high]
**Initiative**: [human | AI | shared] — who starts and proposes
**Control**: [human | AI | shared] — who oversees execution as it runs
**Decision rights**: [human | AI] — who has final say on the outcome
**AI role**: [none | executor | collaborator | challenger]
**Execution disposition**: [AI-assisted | human-only | pending-user-decision]
**Escalation decision**: [not-needed | pending | user's recorded resolution]
**Resolved AUTONOMY_MODE**: [supervised | semi-autonomous | loop-until-approved | autonomous | not-applicable (contract-only sentinel)]

## Validation Checklist
- [ ] Meets "good enough" criteria
- [ ] Respects all boundaries
- [ ] Works for all stakeholders
- [ ] Builds on existing assets appropriately
- [ ] Allocation still fits what the task turned out to be

Implementation Instructions

When to Create Intent Contract

Create an intent contract when:

• User invokes a major workflow (/octo:embrace, /octo:discover, /octo:plan)
• User explicitly asks to "plan" or "set goals" for a task
• A workflow requires multiple phases and validation

Do NOT create for:

• Quick, single-action commands
• Simple file reads or searches
• Conversational questions

Step 0: Allocate the work before scoping it

Run this before capturing intent. The question is not how to run the task across agents but whether it should be delegated at all, and if so, which parts of the authority go where. Framework: Afroogh, Varshney & D'Cruz (2025), A Task-Driven Human-AI Collaboration (arXiv:2505.18422).

Classify risk. Complexity is already scored elsewhere — defer to estimate_complexity and classify_cynefin in scripts/lib/routing.sh rather than re-deriving it. Risk is a separate axis that nothing in the codebase measures, so judge it here on three questions:

• Irreversibility — can the effect be undone, and at what cost?
• Consequence — is anything material at stake: safety, money, data, users?
• Accountability — is a specific person expected to answer for the outcome?

Risk	Reading
Low	Reversible, no material consequence, no named accountability.
Intermediate	Reversible only at real cost, or consequence is unclear.
High	Irreversible, materially consequential, or someone must answer for it.

Allocate the three dimensions separately. They are independent, and treating them as one axis is the mistake this step exists to prevent. People readily hand AI the initiative on unfamiliar work while keeping control and decision rights — an allocation a single autonomy slider cannot express.

• Initiative — who starts, proposes, drafts.
• Control — who oversees execution while it runs.
• Decision rights — who has final say on the result.

Record every outcome explicitly:

Risk / complexity	Initiative	Control	Decision rights	AI role	Execution disposition	Escalation decision	Mode
Low / low	AI	AI	AI	executor	AI-assisted	not-needed	autonomous
Low / high	shared	human	human	collaborator	AI-assisted	not-needed	loop-until-approved
High / low	human	human	human	executor	AI-assisted	not-needed	supervised
High / high	human	human	human	challenger	AI-assisted	not-needed	supervised

The High / high allocation is adversarial: the human leads while AI attacks the proposed decision as a deliberate counterweight to the human's own bias. In High / low work, AI may execute only the bounded actions the human directly approves.

The rule that inverts. For intermediate-risk work where uncertainty is highest, the cited evidence says avoid AI entirely — "neither as a gatekeeper nor as a second opinion". This contradicts the smooth intuition that middling risk implies middling involvement, and it also sits in tension with the same paper's broader claim that complete human autonomy is rarely justified. That tension is real and unresolved; surface it to the user and let them decide rather than quietly picking a side.

Resolve to a setting. The workflow engine reads one variable, AUTONOMY_MODE, with four values:

Allocation	AUTONOMY_MODE
Human holds control and decision rights, approving each phase	supervised
AI runs; human is pulled in on failures and quality gates	semi-autonomous
AI runs and iterates; human holds final decision rights	loop-until-approved
AI holds all three	autonomous

Record the three dimensions and the resolved mode. The mapping is lossy: one axis cannot represent three independent allocations, so a contract that stores only the mode loses the reason it was chosen. That record is what a later reviewer needs when the allocation turns out to have been wrong.

not-applicable is a persisted, contract-only sentinel for human-only work; it is not a fifth runtime value and must not be passed to the workflow engine.

For intermediate risk, record Execution disposition: pending-user-decision. Record Escalation decision: pending, then stop before execution. Ask the user to choose human-only handling or a specific documented AI allocation. Record their answer, rewrite the Task Allocation fields to match it, and change Escalation decision to the user's resolution before continuing. For a human-only resolution, record AI role: none and Resolved AUTONOMY_MODE: not-applicable; human-only not-applicable must not be passed to the workflow engine. Resolve to the supported human-only behavior and do not execute AI work in that state.

For a documented AI-assisted resolution, record every Task Allocation field:

• Initiative, Control, Decision rights, and AI role from the chosen allocation
• Execution disposition: AI-assisted
• Escalation decision: user's recorded resolution
• Resolved AUTONOMY_MODE: exactly one of supervised, semi-autonomous, loop-until-approved, or autonomous, mapped using the table above

Validate the selected runtime mode and only then execute. The runtime must reject not-applicable and every unknown or unsupported mode before workflow execution rather than defaulting to autonomous behavior.

Step 1: Capture Intent

After asking the 3 clarifying questions in a workflow, prompt the user to define:

AskUserQuestion({
  questions: [
    {
      question: "What are you ultimately trying to accomplish?",
      header: "Goal",
      multiSelect: false,
      options: [
        {label: "Let me describe it", description: "I'll write my own goal statement"},
        {label: "Make a decision", description: "Choose between options"},
        {label: "Create deliverable", description: "Build something specific"},
        {label: "Understand a problem", description: "Research and learn"}
      ]
    },
    {
      question: "What defines success for this?",
      header: "Success",
      multiSelect: true,
      options: [
        {label: "Clear recommendation", description: "Know what to do next"},
        {label: "Working implementation", description: "Code that functions"},
        {label: "Team alignment", description: "Everyone understands"},
        {label: "Problem solved", description: "Issue is resolved"}
      ]
    },
    {
      : ,
      : ,
      : ,
      : [
        {: , : },
        {: , : },
        {: , : },
        {: , : }
      ]
    }
  ]
})

If user selects "Let me describe it", follow up with a text prompt for their custom goal.

Step 2: Write Intent Contract File

Use the Write tool to create .claude/session-intent.md:

cat > .claude/session-intent.md <<EOF
# Intent Contract

**Created**: $(date -u +"%Y-%m-%dT%H:%M:%SZ")
**Workflow**: ${WORKFLOW_NAME}
**Status**: active

## Job Statement
${USER_GOAL}

## Success Criteria

### Good Enough
${MIN_SUCCESS_CRITERIA}

### Exceptional
${EXCEPTIONAL_CRITERIA}

## Boundaries
What this should NOT be:
${BOUNDARIES}

## Context & Constraints

**Stakeholders**: ${STAKEHOLDERS}
**Timeline**: ${TIMELINE}

## Clarifying Context
${THREE_QUESTION_ANSWERS}

## Validation Checklist
- [ ] Meets "good enough" criteria
- [ ] Respects all boundaries
- [ ] Works for all stakeholders
EOF

Step 3: Reference During Execution

Throughout the workflow, periodically read .claude/session-intent.md to:

• Stay aligned with user goals
• Make decisions consistent with boundaries
• Keep stakeholders in mind

At key decision points, explicitly say:

Checking against intent contract: [reference specific criterion]

Step 4: Validate at End

When the workflow completes, read .claude/session-intent.md and validate:

Validation Process:

1. Read the intent contract

2. Check each success criterion:

   • ✓ Met - explain how
   • ✗ Not met - explain why and what's needed
   • ~ Partially met - explain gaps

3. Check boundaries:

   • ✓ Respected - confirm
   • ✗ Violated - explain what happened

4. Generate validation report:

# Validation Report

## Success Criteria Check

### Good Enough Criteria
- [✓] Criterion 1: [How it was met]
- [✗] Criterion 2: [Why not met, what's needed]

### Exceptional Criteria
- [~] Criterion 1: [Partial progress explanation]

## Boundary Check
All boundaries respected: [Yes/No]
- Boundary 1: [✓/✗] [Explanation]

## Gaps & Next Steps
[If any criteria not met, list concrete next steps]

## Overall Assessment
[Summary: Does this fulfill the original intent?]

5. Present to user:
   • Show the validation report
   • Ask if they want to address any gaps
   • Update intent contract status to "completed" or "validating"

Step 5: Update Intent Contract Status

Update the Status field in .claude/session-intent.md:

• active → workflow in progress
• validating → checking against criteria
• completed → all criteria met, boundaries respected
• incomplete → some criteria not met, gaps identified

Integration with Workflows

Embrace Workflow

1. Ask 3 clarifying questions (scope, focus, autonomy)
2. Create intent contract
3. DISCOVER phase (reference intent)
4. DEFINE phase (reference intent)
5. DEVELOP phase (reference intent)
6. DELIVER phase (reference intent)
7. Validate against intent contract
8. Present validation report

Discover Workflow

1. Ask 3 clarifying questions (depth, focus, output)
2. Create intent contract
3. Execute multi-provider research
4. Synthesize findings
5. Validate against intent contract
6. Present validation report

Plan Workflow (Future)

1. Capture comprehensive intent
2. Create intent contract
3. Route to appropriate workflows
4. Execute custom sequence
5. Validate against intent contract
6. Present validation report

Example Intent Contract

# Intent Contract

**Created**: 2026-01-21T15:30:00Z
**Workflow**: embrace
**Status**: active

## Job Statement
Build a user authentication system that our team can implement and maintain.

## Success Criteria

### Good Enough
- Team understands what to build
- Clear technical approach selected
- Security considerations documented
- Implementation plan with steps

### Exceptional
- Multiple authentication methods evaluated
- Security audit performed
- Code examples provided
- Integration tests included

## Boundaries
What this should NOT be:
- Over-engineered with unnecessary features
- Disconnected from our existing Node.js/Express stack
- Experimental or unproven technologies

## Context & Constraints

**Stakeholders**: Development team (5 engineers), Product manager
**Existing Assets**: Express.js API, PostgreSQL database
**Timeline**: Need to start implementation next sprint
**Technical Constraints**: Must work with Express.js, PostgreSQL

## Clarifying Context

**Scope**: Medium feature (multiple components)
**Focus Areas**: Security, Architecture design
**Autonomy**: Supervised (review after each phase)

## Validation Checklist
- [ ] Meets "good enough" criteria
- [ ] Respects all boundaries
- [ ] Works for all stakeholders
- [ ] Builds on existing assets appropriately

Benefits

For Users:

• Clear expectations set upfront
• No forgotten requirements
• Validation against original goals
• Closed-loop accountability

For Workflows:

• Clear success criteria to optimize for
• Boundaries to constrain solutions
• Context for better decisions
• Validation framework built-in

Ready to use! Workflows can now create and validate against persistent intent contracts.