prompt
prompt — Claude-first prompt design and review
Modes
Detect mode from the request. Default to DESIGN.
| Mode | Signals | Loads | Output |
|---|---|---|---|
| DESIGN (default) | write, create, draft, fix, improve, migrate, "prompt for X" | references/design.md in this skill directory + ≤2 topic modules |
The prompt + ≤3-line rationale |
| REVIEW | review, audit, analyze, critique, evaluate, "what's wrong with" | references/review.md in this skill directory + ≤2 topic modules |
Verdict table + ranked fixes |
REVIEW framing rule. Frame every review as a search for defects ("find what would fail"), never as confirmation ("check this is correct" / "verify this works"). When dispatching a fresh-context reviewer, pass the artifact with zero authorship or intent framing — no "I wrote this", no "we think it's solid". Confirmatory framing collapses defect detection (evidence in review.md).
Core principles
1. Smallest set of high-signal tokens. Treat Claude as a brilliant new employee: clear, explicit instructions plus the context to generalize — then stop. Golden rule: if a colleague with minimal context would be confused by your prompt, Claude will be too. When trimming, run the removal test (review.md) instead of counting against a quota.
2. Explain why; say what to do, not what to avoid. Motivation generalizes: "the output is read aloud by a TTS engine, so never use ellipses" outperforms "NEVER use ellipses". Positive framing: "write flowing prose paragraphs" beats "do not use markdown".