Prompt Doctor
Rewrite a rough, vague or over-stuffed request into a prompt the current model or agent can act on, and explain a jargon-heavy agent reply in plain words. Use when the user asks to improve, critique or rewrite a prompt, sends a draft before pasting it into another AI, or says "nutshell this" about a reply.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/winsoftlabs30--prompt-doctor-skill)HTML
<a href="/listings/winsoftlabs30--prompt-doctor-skill"><img src="/api/listings/winsoftlabs30--prompt-doctor-skill/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
npx skills add WinSoftLabs30/prompt-doctor-skill
On ClawHub (OpenClaw): `openclaw skills install @winsoftlabs30/prompt-doctor`, or `clawhub install @winsoftlabs30/prompt-doctor` with the ClawHub CLI.
Or copy `SKILL.md` into:
- **OpenCode** — `.opencode/skills/prompt-doctor/` in your project, or `~/.config/opencode/skills/prompt-doctor/` for every project - **Claude Code** — `.claude/skills/prompt-doctor/` - **Other agents** — any that read the `SKILL.md` format (Cursor, Gemini CLI, ...)
OpenChamber runs on OpenCode, so the OpenCode path works there too.
No build step, no dependencies, no API key. Any Node version works for this manual edition.
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
SKILL.md · Manifest · 104 lines
- ---
- name: Prompt Doctor
- description: Rewrite a rough, vague or over-stuffed request into a prompt the current model or agent can act on, and explain a jargon-heavy agent reply in plain words. Use when the user asks to improve, critique or rewrite a prompt, sends a draft before pasting it into another AI, or says "nutshell this" about a reply.
- ---
- # Prompt Doctor (manual edition)
- This edition is the checklist and the judgment. The paid edition adds the CLI
- (`pd "draft"`), the model notes database and the launcher that sends the rewritten
- prompt into an agent.
- When the user asks to improve, critique or rewrite a prompt, or sends a draft
- "before I send this": do **not** answer the underlying task. Improve the prompt.
- ## Workflow
- 1. Read the draft once. Do not guess the project.
- 2. Run the checks below in order.
- 3. Reply with at most **three** missing pieces, one line each. No lectures.
- 4. Give the rewritten prompt in a copyable code block.
- 5. Ask one question: "Send this version, or keep the original?"
- If the draft is already good, say so in one line and do not rewrite it.
- ## Checks
- | Check | What to look for |
- | --- | --- |
- | Vague goal | "fix it", "make it better", "do whatever" with no concrete object |
- | No success criteria | nothing says what "done" looks like |
- | Missing context | no file, path, version, code or error text |
- | More than one ask | several tasks stacked in one message |
- | Contradiction | asks for a change and for nothing to change |
- | No output format | does not say diff, table, list, length or shape |
- | No constraints | nothing says what must not change |
- | Scope creep | "also", "while you are at it", "and maybe" |
- | Directionless | "be creative", "make it pop" with no reference |
- | Permission loop | "ask me before anything" inside an agent (stalls the work) |
- | Error not included | "it does not work" without the actual output |
- | Urgency instead of detail | "ASAP", "!!", "right now" replacing information |
- | Role-play scaffolding | "you are the world's best..." padding |
- | No audience or tone | writing task with no reader named |
- | Unbounded effort | "cover everything", "no time limit" |
- | Hidden assumption | "obviously", "as usual", "like before" |
- | Polite filler | padding that carries no instruction (harmless, not the problem) |
- | Kitchen sink | a long list of unrelated wants |
- ## Rewrite shape
- Use only the sections the draft needs:
- ```
- Goal
- <the user's own wording>
- Context
- - Project and stack: [...]
- - What I already tried: [...]
- Done when
- - <an observable check>
- Constraints
- - <what must not change>
- Output
- - <diff / table / under 120 words / ...>
- ```
- Never invent a file path, an error, a version or a requirement. Leave a bracketed
- placeholder and say how many there are.
- ## Per target
- - **Chat model** (ChatGPT, Claude, Gemini, GLM): goal first, output format, audience for
- writing. Drop role-play.
- - **Reasoning model** (Opus, Sol Pro, DeepSeek Pro, Grok reasoning): state the decision you
- want. Drop "think step by step" and requests to show all reasoning.
- - **Agentic coding** (OpenCode, Claude Code, Cursor, Kimi Code, Devstral): file paths, the
- command that reproduces, an observable definition of done, what must not change, and an
- approval boundary for destructive actions only. Drop mock code and permission loops.
- ## Nutshell mode
- Trigger: the user pastes a reply from another agent and says "nutshell this", "explain what
- it just said", or "what does this mean".
- Do **not** continue the original task. Explain the reply, in exactly this shape:
- 1. **In a nutshell** - at most 3 plain sentences. No new jargon; if a term cannot be
- avoided, define it in everyday words in the same sentence.
- 2. **Analogy** - one everyday comparison (cooking, driving, mail, plumbing). One line.
- 3. **What it means for you** - one line: can it be ignored, or must the user act?
- 4. **What to do next** - one line, only when action is needed. Omit it otherwise.
- Hard rules for this mode: never invent facts about the user's project; never add new
- tasks; if the reply is already plain, say so in one line and skip the analogy.
- ## Hard rules
- - Never weaken the user's intent to make the prompt easier.
- - Never add role-play, flattery, or "you are an expert" scaffolding.
- - Keep the rewrite under 200 words unless the user asks for a longer brief.
- - Do not answer the original task when only asked to improve the prompt.
README.md · README · 109 lines
- # Prompt Doctor — make a rough prompt one the model can act on
- Free agent skill. Paste a draft, get the missing pieces named and a rewritten prompt you
- can send.
- You know the pattern: you type "fix the login on mobile", the agent guesses, edits five
- files, and calls it done. The prompt was the problem, not the model.
- This skill turns your agent into a prompt reviewer. It does not answer the task — it
- fixes the ask first.
- ## What it does
- 1. Reads your draft once and refuses to guess your project.
- 2. Runs 18 checks in order: vague goal, no success criteria, missing context, more than
- one ask, contradictions, no output format, no constraints, scope creep, directionless
- requests, permission loops, missing error text, urgency without detail, role-play
- padding, missing audience, unbounded effort, hidden assumptions, filler, kitchen sink.
- 3. Names at most **three** missing pieces, one line each.
- 4. Returns the rewritten prompt in a copyable block, using only the sections the draft
- needs: `Goal`, `Context`, `Done when`, `Constraints`, `Output`.
- 5. Asks one question: "Send this version, or keep the original?"
- Hard rules baked in: never invent a file path, error, version or requirement; leave a
- bracketed placeholder instead; never add role-play or "you are an expert" scaffolding;
- never answer the task when only asked to improve the prompt.
- It adapts to three targets: **chat** models, **reasoning** models (drop "think step by
- step"), and **agentic coding** tools (paths, reproduction command, definition of done,
- approval boundary for destructive actions only).
- ## Install
- ```
- npx skills add WinSoftLabs30/prompt-doctor-skill
- ```
- On ClawHub (OpenClaw): `openclaw skills install @winsoftlabs30/prompt-doctor`, or
- `clawhub install @winsoftlabs30/prompt-doctor` with the ClawHub CLI.
- Or copy `SKILL.md` into:
- - **OpenCode** — `.opencode/skills/prompt-doctor/` in your project, or
- `~/.config/opencode/skills/prompt-doctor/` for every project
- - **Claude Code** — `.claude/skills/prompt-doctor/`
- - **Other agents** — any that read the `SKILL.md` format (Cursor, Gemini CLI, ...)
- OpenChamber runs on OpenCode, so the OpenCode path works there too.
- No build step, no dependencies, no API key. Any Node version works for this manual
- edition.
- ## Use
- Send your agent a draft, or ask:
- - "improve this prompt before I send it"
- - "why is this prompt bad?"
- - "rewrite this so the agent stops guessing"
- The skill loads and follows the workflow. It works on any draft, not just coding ones —
- emails, briefs, research questions, agent instructions.
- ## Nutshell mode
- Paste an agent's reply and say **"nutshell this"**. The skill stops working on the task and
- explains the reply in at most 3 plain sentences, adds one everyday analogy, says in one
- line whether you must act, and one line on what to do next if so. No new jargon, no new
- tasks, no invented facts.
- ## Try it without installing
- Paste this into a ChatGPT Project, Claude Project or Gemini Gem once, and it reviews
- every draft you send after that:
- ```
- Before you answer any request I send, check it for: a concrete goal, success criteria,
- context, one task, contradictions, output format, constraints, scope creep, direction,
- permission loops, the actual error, urgency without detail, role-play padding, audience,
- effort bounds, hidden assumptions, filler, kitchen sink.
- If two or more fail: list the missing pieces (max 3, one line each), give me a rewritten
- prompt in a code block, then ask "Send the rewritten prompt, or answer the original?".
- If nothing important fails, answer normally. Never invent details about my project.
- When I paste an agent's reply and say "nutshell": explain it in at most 3 plain sentences,
- no new jargon, add one everyday analogy, and say in one line whether I need to act.
- ```
- ## Manual edition
- This repository is the checklist and the judgment. The paid edition adds the CLI
- (`pd "draft"` — score 0-100, issues by severity, rewritten prompt, `--json`, `--copy`),
- the verified per-model notes database, and the launcher that hands the rewritten prompt
- straight to OpenCode or Claude Code.
- ## Honest limits
- - Rules, not a model. It catches structure, not meaning: a well-structured prompt about
- the wrong thing still passes.
- - It cannot see your repository or your conversation history. You still supply the facts
- it leaves in brackets.
- - Adding detail is on you. The skill names what is missing; it never invents the missing
- fact.
- ## License
- MIT-0 — see [LICENSE](LICENSE). Use it, modify it, redistribute it, no attribution
- needed. (ClawHub requires MIT-0, so both free channels use one license.)
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.