junnv93/generate-prompts

generate-prompts

Scan → Verify → Generate harness prompts for example-prompts.md.

Critical rule: Never trust 1st-pass scan results. Always verify with Read/Grep before writing a prompt.

Workflow

1. Load existing prompts

Read .claude/skills/harness/references/example-prompts.md. Note completed (~~strikethrough~~), in-progress, and pending items.

2. Parallel scan (3 Agents)

Launch 3 Agents in parallel with prompts from references/scan-prompts.md:

• Agent A: Backend (TODO/FIXME, security gaps, performance, dead code)
• Agent B: Frontend (TODO/FIXME, hardcoded strings, a11y, missing error.tsx)
• Agent C: Infra/packages (unused enums, missing indexes, CI optimization)

Each agent must report findings with file:line.

3. Verify each finding (CRITICAL — never skip)

For every finding, use Read/Grep to confirm it exists in current code:

"X is missing" → Read the file, confirm it's actually missing
"Decorator missing" → Read the full controller
"Index missing" → Read the schema's index section

Classify each:

• Confirmed → write prompt
• False positive → exclude, record in archive with reason
• Needs user input → ask user (domain decisions like FK policy, permission removal)

4. Write prompts

Format per finding (confirmed only):

### {title} (Mode {0|1|2})

\`\`\`
{problem — what's wrong, verified locations with file:line}
{action — what to do}
검증: {success criteria}
\`\`\`

Mode: 0 (≤3 files) / 1 (4-15 files) / 2 (15+ or DB change)

Priority: 🔴 CRITICAL (security) → 🟠 HIGH (perf/gaps) → 🟡 MEDIUM (quality) → 🟢 LOW (polish)

5. Archive completed items

Move to <details> archive section:

• ~~Strikethrough~~ items
• Items with merged PRs (check git log)
• Items confirmed as already implemented

False positives → separate <details> section with lessons learned.

6. Update example-prompts.md

Write final prompts to .claude/skills/harness/references/example-prompts.md:

• Update header date
• Add new items under priority sections
• Move completed/false-positive to archive

7. User review

Present summary and ask for confirmation:

| Priority | New | Archived | False Positive |
|----------|-----|----------|----------------|

### Items needing your decision
{domain questions}

Exceptions

1. Never delete completed items — archive them (lessons learned)
2. Never include unverified findings — Step 3 is mandatory
3. Never make domain decisions autonomously — ask user
4. Test file TODO/FIXME → classify as MEDIUM or lower

Related Files

| File | Purpose | |------|---------| | references/scan-prompts.md | Agent prompts for Step 2 | | .claude/skills/harness/references/example-prompts.md | Prompt storage (SSOT) | | .claude/skills/harness/SKILL.md | Harness skill (prompt consumer) |