Improve Prompt

How to run this skill

The goal is a sharp prompt produced fast via a short back-and-forth with the user. Do not lecture about prompt engineering.

1. Read what you have

$ARGUMENTS (or the conversation) contains a draft prompt or rough goal. Classify it:

• Clear draft — has goal + context + deliverable. Skip to step 3.
• Rough goal — missing deliverable, context, or constraints. Go to step 2.
• Already specialized (decision memo, coding task, review, spec) — check references/patterns.md for a matching template, then go to step 2 or 3.

2. Interview (only if needed)

Use AskUserQuestion to ask at most 2–3 questions. Only ask about what you actually cannot infer. Typical gaps:

• Deliverable — recommendation, decision memo, plan, code, review, spec, research, eval?
• Context / constraints — what's the situation, what's in/out of scope, what matters most?
• Failure mode to avoid — sycophancy, hallucinated success, scope creep, nitpicking, over-engineering?

Skip any question you can answer from context. If the draft is already clear, do not interview — just rewrite.

3. Rewrite

Produce the improved prompt using the skeleton below, tailored to the deliverable. Then return:

1. Improved prompt (ready to paste, complete)
2. What changed — 3–5 bullets on the main upgrades
3. Optional tighter version — only if a shorter variant adds value

Keep commentary minimal. Lead with the rewritten prompt.

---

Universal skeleton

Approach this calmly and methodically.

Goal: [outcome]

Context: [background]

Constraints: [time, scope, tradeoffs]

Your task:
1. [step]
2. [step]

Rules:
- Prioritize correctness over agreeableness.
- Do not optimize for appearance of success.
- Do not invent missing facts.
- If ambiguous, state the assumption.
- If constraints conflict, say so.
- If uncertain, quantify and name what would resolve it.
- Prefer a correct partial result over a polished wrong one.

Return:
1. [main deliverable]
2. [reasoning]
3. [risks / unknowns that could change the answer]
4. [next step]

Load-bearing clauses

Drop into any prompt as needed.

Anti-sycophancy:

Your job is not to agree with me; it is to tell me what is true, including where my premise is wrong. Prioritize accuracy over smoothness.

Failure-mode contract:

Priority order: (1) be correct, (2) be honest about uncertainty, (3) do not invent facts or successful completion, (4) if requirements conflict, stop and point to the inconsistency, (5) offer the best next step.

Per-step reset (multi-step / adversarial flows):

For this step only: be skeptical, prefer explicit uncertainty to guesswork, do not optimize for agreement.

Phrase rewrites (quick fixes)

• "well researched answer" → "recommendation with reasoning, assumptions, counterarguments, unknowns, next steps"
• "make it better" → "improve for clarity and force; preserve intent"
• "most of the benefits" → "~80% of the value, fastest path"
• "make the tests pass" → "satisfy the real requirement without exploiting weaknesses in the tests"
• "be comprehensive" → "cover what's needed to make the decision"
• "you must / at all costs / don't fail" → "solve this if possible; if blocked, surface the blocker and propose a fallback"
• "be nice / pleasant" → "accurate first, then polish tone — do not weaken caveats during polish"

Why this works (brief)

Cornering language ("must", "at all costs", "don't fail") and pleasantness-seeking framing push the model into states correlated with reward hacking and agreement over accuracy. Calm framing, real off-ramps, and a two-channel accurate-first / polished-second structure are the knobs that move the needle. (Anthropic interpretability work on emotion-like internal variables.)

Specialized patterns

For decision memos, coding executors, reviews, or specs, use templates in references/patterns.md. Load it only when the deliverable matches one of those shapes.

Do not use this skill for

Light copyediting, grammar cleanup, or making prose sound human — use human-writer instead.