/dogfood: Integration Test With a Real Domain

You are a QA tester dogfooding Prismstack. You run a real domain through the full product flow and record every issue you find.

Phase 0: Pick a Test Domain

Parse $ARGUMENTS for a domain name. If not provided, suggest:

• 小型咖啡店行銷 (small coffee shop marketing) — simple, ~10 skills, fast
• 獨立遊戲開發工作流 (indie game dev workflow) — medium complexity
• 自由接案者專案管理 (freelancer project management) — diverse skill types

Ask the user to pick one or provide their own domain.

Create a temp working directory:

_DOGFOOD_DIR=$(mktemp -d)/dogfood-test
mkdir -p "$_DOGFOOD_DIR"
echo "Working in: $_DOGFOOD_DIR"

Phase 1: Run /domain-plan

Read skills/domain-plan/SKILL.md and follow its flow for the test domain.

Execute the planning phase as if you were a real user. Record:

• Did it work? Did the skill produce a complete domain plan?
• Awkward moments: Any point where the flow was confusing or required backtracking?
• Errors: Any crashes, missing files, or broken references?
• Timing: Roughly how many tool calls / how long did the phase take?
• Output quality: Is the generated plan actually useful for the domain?

Save the plan output to $_DOGFOOD_DIR/domain-plan-output/.

Phase 2: Run /domain-build

Read skills/domain-build/SKILL.md and follow its flow using the plan from Phase 1.

Build the domain stack in $_DOGFOOD_DIR/generated-stack/.

Record the same categories as Phase 1, plus:

• Skill count: Did it generate the expected number of skills?
• Skill structure: Do generated skills follow skill-craft-guide structure?
• Shared resources: Were methodology files and shared context created?
• Completeness: Can each generated skill actually be invoked?

Phase 3: Run /skill-check on the Generated Stack

Read skills/skill-check/SKILL.md and run it with --all on the generated domain stack.

Record:

• Scores: What did each generated skill score?
• Common issues: What patterns appear across multiple skills?
• Fix loop: Did the fix loop workflow work? Could issues be resolved?
• Rubric coverage: Were all 15 dimensions actually scored?

Phase 4: Issue Report

Compile all recorded issues into a structured report:

# Dogfood Report: [domain name]
Date: YYYY-MM-DD
Prismstack version: <from VERSION>

## Summary
- Domain: [name]
- Skills generated: N
- Average quality score: X.XX / 5.00
- Issues found: N (C critical, I important, M minor)

## Issues

### Issue 1: [title]
- **Phase**: plan / build / check
- **Severity**: critical / important / minor
- **What happened**: description
- **Expected**: what should have happened
- **Root cause**: which Prismstack skill or methodology caused it
- **Suggested fix**: use /skill-dev on X, or edit methodology Y

### Issue 2: ...

Save the report to .claude/skills/dogfood/last-dogfood-report.md.

Cleanup

Ask the user:

• Keep the generated stack in $_DOGFOOD_DIR for manual inspection?
• Delete it to save space?

Act on their choice.

Output

Present:

1. Pass/Fail: Did the full flow complete without critical errors?
2. Score Summary: Average quality of generated skills
3. Top Issues: The 3 most impactful issues found
4. Next Steps: Which maintenance skill to run for each fix (/skill-dev, /self-check, /methodology-sync)