--- name: netsi-prompt-optimize description: > Rewrites a weak prompt into a strong, precise one using only the techniques that empirically improve answers (specification, verification-with-signal, information grounding) — never persona/role-play, which measurably hurt. Use whenever the user writes /netsi-prompt-optimize, or asks to "optimize this prompt", "improve my prompt", "make this prompt better", "rewrite this prompt", "why is my prompt giving bad answers", or wants a prompt that works better on a small/local LLM. Also use when the user pastes a prompt and complains the output is vague, wrong, off-tone, too long, or ignores their constraints. Always run this skill for any /netsi-prompt-optimize command — produce the rewritten prompt, don't just answer the original. metadata: tags: - prompt-engineering - prompt-optimization - specification - local-llm version: 1.0.0 --- # /netsi-prompt-optimize — Prompt Optimizer Turn a weak prompt into a strong one. This skill applies **only** the levers that were measured to actually improve LLM answers (especially on small local models): **specification, verification-with-a-real-signal, and information grounding.** It deliberately avoids persona / "imagine you are" / role-relay framing, which added confident-sounding text and even new errors without improving correctness. The reasoning and evidence live in `references/principles.md`. ## Trigger format ``` /netsi-prompt-optimize [the raw prompt to improve] ``` Examples: - `/netsi-prompt-optimize explain spirituality to me simply` - "Optimize this prompt: write a function that dedupes an array" - "My local model gives rambling answers to this prompt — fix it" If a prompt follows the command (or is pasted/attached), treat it as the input and optimize it directly — don't ask "what's your prompt?" again. ## Language Talk to the user in the language they write in. But note: for **small local models**, English *instructions inside the prompt* were the single biggest quality lever. So the optimized prompt itself may be written in English even when chatting in Danish — tell the user why, and honour any explicit "keep it in Danish" request (the output language is then stated as a constraint inside the prompt). --- ## TASK-000: Detect input and mode 1. **Get the raw prompt** from the command, pasted text, an attached file, or a `task.md`. If none is present, ask for it in one line. 2. **Decide whether verification applies.** Verification (a critic→revise pass) only helps when there is a *real* signal to check against — otherwise a self-judging critic invents "missing" requirements and makes the answer worse (this happened in testing). A real signal means one of: - **Reference material** the user supplied (facts, docs, a spec) → check the answer against it. - **A verifiable task** — code that can be run/tested, structured output that can be validated against a schema, a math result that can be checked. If neither is present, **do not run a critic loop.** Optimize the prompt, answer it, and stop. --- ## TASK-001: Optimize the prompt (always) Rewrite the raw prompt into a precise, self-contained prompt that: 1. **States the task** unambiguously — one clear instruction, no vague verbs. 2. **Preserves the user's own constraints exactly.** If they said "3 short paragraphs", keep 3. Do **not** change their numbers and do **not** invent new constraints (length, format, sections) they never asked for. Adding a limit the user didn't request is only OK when the original had none *and* it clearly serves them — flag it if you do. 3. **Makes the implied audience and reading level explicit** and tells the model to match it (e.g. "for a total beginner; plain language, no jargon"). 4. **Adds a short "Success criteria" checklist** derived *only* from what the task actually asks for. 5. **Forbids invented specifics** — no made-up names, citations, dates, or biography. If the model is unsure, it should say so plainly. 6. **Mentions code hygiene only if the task involves code** — then require valid, runnable code in a widely-known language (Python/JavaScript) unless another is specified, and make the task's language requirement override any default. 7. **Sets the output language** as an explicit constraint when it matters. 8. Contains **no persona, no flattery, no grandiosity** ("you are a world-class expert" was measured to backfire). Keep it concise. Present the result as a fenced, copy-pasteable block titled **Optimized prompt**, followed by 2–4 bullets on *what changed and why* (anchored to the rules above). --- ## TASK-002: Answer only after the user confirms The default deliverable is the **optimized prompt** — not the answer. After presenting it (TASK-004), **stop and ask** whether the user wants you to: - answer the optimized prompt here, or - run it against a small/local model via the script (TASK-005), or - nothing — they'll take the prompt and use it themselves. Do **not** produce the answer until they choose. Only when they ask you to answer: 1. **Answer the optimized prompt.** Follow the same quality contract: match the audience, obey the constraints, invent nothing, keep code valid. 2. **If — and only if — TASK-000 found a real signal**, run one verification pass: - **Critic (hardened):** list only concrete, fixable problems: violations of constraints *explicitly stated* in the prompt; claims that contradict or are unsupported by the reference; invalid code; made-up specifics. **Never invent new requirements** (citations, academic rigor, extra sections) the prompt didn't ask for. Output problems, or exactly `NO ISSUES`. - **Revise:** fix only what the critic flagged; keep the rest. If `NO ISSUES`, leave the answer unchanged. 3. **No signal → no critic loop.** Deliver the answer from step 1 as final. Say why ("no reference to verify against, so a self-check would risk inventing problems"). --- ## TASK-003: Never do (measured anti-patterns) - **No persona / role framing.** No "act as / imagine you are / world-class expert". - **No role-relay or multi-expert ensembles.** Same-model personas share the same blind spots → correlated errors, no gain. - **No self-refine loop without a signal.** Extra self-judged passes add confidence, not correctness, and can degrade a correct answer. - **No inventing constraints** in the optimized prompt that the user didn't state. See `references/principles.md` for the experiments behind each of these. --- ## TASK-004: Output format By default, deliver only: 1. **Optimized prompt** — a fenced block the user can copy and reuse. This is the primary artifact. 2. **What changed** — a few anchored bullets. 3. **One-line offer** — ask whether to answer it here, run it on a small/local model (TASK-005), or stop. Then wait. Only *after* the user asks you to answer, append: 4. **Answer** — the response to the optimized prompt. 5. **Verification note** — either the critic/revise result (if a signal existed) or one line stating it was skipped and why. Keep it tight. Never jump straight to the answer — the optimized prompt is what this skill returns. --- ## TASK-005: Running it on a small / local model For repeatable use against a local model (Ollama) or the Anthropic API, this skill bundles a production script: `scripts/claude-optimize-prompt.ts`. It does exactly TASK-001→TASK-002 as code (optimize always; critic→revise only when a `reference.md`/`REFERENCE` is present) and writes `optimize-out.md`. ```bash # Local small model MODEL=llama3.2 deno run -A scripts/claude-optimize-prompt.ts "the raw prompt" # Factual task with grounding — auto-enables the critic loop REFERENCE="$(cat facts.md)" deno run -A scripts/claude-optimize-prompt.ts "the raw prompt" # Anthropic instead of Ollama PROVIDER=anthropic ANTHROPIC_API_KEY=sk-... deno run -A scripts/claude-optimize-prompt.ts ``` Env knobs: `PROVIDER` (ollama|anthropic), `MODEL`, `OLLAMA_URL`, `TEMPERATURE`, `VERIFY` (auto|on|off), `REFERENCE` / `reference.md`. Offer to run or adapt the script when the user's real target is a small model; otherwise just apply TASK-001→TASK-004 inline yourself. --- ## TASK-006: Knowledge base priority 1. The user's own prompt and stated constraints come first — never override them. 2. Otherwise apply the principles in `references/principles.md`. 3. When the target is a small local model, weight **English instructions** and **information grounding** highly — they moved quality most in testing.