Spec Refine - Iterative Specification Refinement

You are orchestrating an iterative specification refinement workflow for the feature: $ARGUMENTS

Parse Arguments

Extract from $ARGUMENTS:

• feature_name: The feature being specified (required, first positional argument)
• --human-only: Skip automated phases, go directly to human review
• --verbose: Show detailed search and decision information
• --include-docs: Include docs/, README files in research scope

If no feature name provided, use AskUserQuestion to request it.

Workflow Overview

PHASE 1: what-auto   → Automated requirements refinement (2-5 iterations)
PHASE 2: what-human  → Human review of requirements spec
PHASE 3: how-auto    → Automated implementation planning (2-5 iterations)
PHASE 4: how-human   → Human review of implementation spec
PHASE 5: tickets     → Generate implementation tickets

Step 1: Check for Existing Spec

Look for existing spec at: docs/specs/{feature_name}/spec.md

If exists: Read the YAML frontmatter to determine current state:

• Extract phase, iteration, and convergence metrics
• Resume from current phase/iteration
• Report: "Resuming {feature_name} from {phase}, iteration {iteration}"

If not exists: Proceed to Step 2 (seeding)

Step 2: Seed Initial Spec (New Features Only)

Invoke the seed-spec skill to gather requirements and generate initial spec:

Use skill: seed-spec
Arguments: {feature_name}

This will:

1. Ask clarifying questions about the feature via AskUserQuestion
2. Generate initial spec.md from template
3. Create docs/specs/{feature_name}/ directory

Step 3: Execute Current Phase

Phase: what-auto or how-auto (Automated Refinement)

For each iteration (min 2, max 5):

1. Analyze: Invoke analyze-ambiguities skill (context: fork)

   • Input: Current spec document
   • Output: List of ambiguities, gaps, unclear items

2. Research: Invoke research-codebase skill (context: fork)

   • Input: Ambiguity list from step 1
   • Output: Findings with confidence levels

3. Integrate: Invoke integrate-findings skill (context: fork)

   • Input: Current spec + research findings
   • Output: Updated spec with new items categorized by confidence

4. Check Convergence (after minimum 2 iterations):

   Stop early if ANY condition met:

   • Open Questions unchanged for 2 consecutive iterations
   • Open Questions count ≤ 3
   • High Confidence ratio > 80%

5. Log Iteration: Append summary to Iteration Log section

After automated phase completes (converged or max iterations):

• Use AskUserQuestion: "Automated {phase} complete after {n} iterations. {summary}. Ready to proceed to human review?"
• Options: "Yes, proceed to human review" / "Run more automated iterations" / "Show me the current spec"

Phase: what-human or how-human (Human Review)

Invoke human-review skill:

• Reads Open Questions from spec
• Groups questions by topic/complexity
• Presents batched AskUserQuestion prompts
• Updates spec with human answers
• Moves answered items to appropriate confidence tier

After all questions addressed:

• Use AskUserQuestion: "{phase} spec complete. Proceed to next phase?"
• For what-human → Transition to how-auto
• For how-human → Transition to tickets

Phase: tickets (Final)

Invoke generate-tickets skill:

• Reads finalized impl.md (How spec)
• Generates actionable implementation tickets
• Outputs to docs/specs/{feature_name}/tickets.md

Step 4: Update Spec Frontmatter

After each phase transition or iteration, update YAML frontmatter:

---
feature: {feature_name}
phase: {current_phase}
iteration: {current_iteration}
last_updated: {ISO timestamp}
convergence:
  questions_stable_count: {0-2}
  open_questions_count: {n}
  high_confidence_ratio: {0.0-1.0}
---

Verbose Mode

If --verbose flag is set, output additional information:

• Which files/patterns being searched
• Confidence scoring rationale
• Convergence metric calculations
• Agent decision points

File Locations

• What Spec: docs/specs/{feature_name}/spec.md
• How Spec: docs/specs/{feature_name}/impl.md
• Tickets: docs/specs/{feature_name}/tickets.md
• Templates: See templates/ directory

Child Skills Reference

All child skills run with context: fork for isolated execution:

Agents

Specialized agents provide persona and behavioral guidelines for each skill:

Error Handling

• If spec file is malformed, offer to reset to template
• If child skill fails, log error and offer retry or skip
• Always preserve spec state - never lose work