Install once, then turn it on
Sharpen needs Claude Code with plugin support and Node.js 18 or later for its hooks. In Claude Code, run each command as a separate prompt:
- Restart Claude Code. Plugins load at the start of a session, so the current session will not pick up the installation.
- Enable Sharpen: send
/sharpen on. - Try a real request. Sharpen is off by default and remembers its on/off setting across sessions.
From rough request to reviewed prompt
Sharpen sits before Claude starts the task. It gates the request, gathers useful context, engineers a prompt for that intent, checks it, and asks you what to do.
- Gate the request.Clear requests keep their wording. One obvious interpretation gets missing context filled in. If different interpretations would lead to different work, Sharpen asks which one you mean.
- Gather context.It checks the conversation first, makes a few focused code searches, and reads the tools and hook summary available in your Claude Code setup.
- Engineer and check.It states the target, relevant files, constraints, installed tools, finish line, and evidence. It checks for common prompt problems such as vague scope or a missing success condition.
- Show the card.The prompt is presented for your decision. Nothing in the task starts until you choose an option.
Sharpen classifies the request silently; it does not expose private chain-of-thought. You see the useful prompt and its concrete constraints.
See how the gate handles different requests
A short bug report with project context
You type: “fix the clipboard so it clears after 30 seconds”
In a documented real run, Sharpen found that the configured duration was 60 seconds and connected the implementation to the matching hint, test, product requirement, checklist, and project hook.
Change the clipboard clear for copied keys from 60 to 30 seconds. Where: apps/admin/src/lib/clipboard.ts:2 (HOLD_MS = 60_000 → 30_000) Also update: generate.tsx:181, clipboard.test.ts, PRD FR-19, checklist US-020. Keep: the "leave it alone if something else was copied" rule. Hooks: dead-code-check runs after every edit. Check: the clipboard test passes.
Those file references are from the example project in Sharpen's README. In your project, Sharpen should use paths it actually finds.
“Fix the login thing”
If the conversation or code shows two live login problems, such as a PIN lockout and an expired member session, those would lead to different fixes. Sharpen asks which one you mean and gives distinct choices, plus Run as typed.
“Rename getUser to fetchUser in src/api.ts”
There is one action and a real location. Sharpen keeps the requested rename intact, then can add relevant callers and a check command. It should not turn a clear request into a different task.
“Set the plan price to the right amount”
If the project and conversation do not establish the price, Sharpen marks it [ASK: …]. It does not invent a number or present a guess as a decision.
What Sharpen uses—and what it adds
It looks in a narrow order so it can use facts already established instead of asking you to repeat them:
- The conversation: decisions, paths, errors, earlier attempts, and constraints you have already supplied.
- The codebase: up to three quick searches for files, symbols, and checks; up to six for an ambiguous prompt when the first search has not found enough.
- Your Claude Code setup: listed skills, agents, MCP tools, and plugin commands. A session-start summary supplies hook names and enabled plugins that Claude otherwise may not see.
The prompt shape changes with the work. A bug brief includes symptom, location, reproduction, fix, and check. A feature brief adds the user need, integration point, existing pattern, done condition, and out-of-scope boundary. Refactors state what must not change; performance work names a metric and baseline; reviews and questions can be kept read-only.
You ask: “Add CSV export to reports.”
The brief can pin down who uses it, where the export belongs, the existing download pattern, what makes the file correct, and which reports are out of scope.
You ask: “Move date parsing into dateUtils.”
The prompt can preserve timezone behavior explicitly, keep callers in scope, and require the existing date tests to stay green.
You ask: “What does fetchSession return?”
The request can stay read-only: explain the returned value and point to the actual source, without changing code.
You ask: “Deploy the current main build.”
The brief should identify the target and exact action. If a step is irreversible, it should stop for your decision before that step.
Decided:Earlier choices that still constrain the work.
Tools:Relevant installed skills, agents, and tools, in a useful order.
Hooks:Hooks that may affect the work. Sharpen reads their names and summaries; it does not call them.
Check:A real test, build, or other observable completion condition when one can be found.
The review card is the approval point
Read the proposed prompt before Claude acts. The card gives you a few different ways to proceed:
Send it
Recommended when the engineered prompt matches your intent. Claude works on that prompt.
Run my original
Use the words you typed, unchanged.
Type your changes
Clarify a detail, such as “make it 15 seconds”. Sharpen updates the prompt and shows you another card.
Esc, then paste
The engineered prompt is copied to the clipboard. Press Esc, paste it with Ctrl+V (⌘V on macOS), edit if you want, and send it yourself.
Claude Code plugins cannot replace your input, fill the input box, or send a message as you. That is why Sharpen shows a card and uses the clipboard instead.
Ten focused workflows, on or off
These slash commands are available whether automatic Sharpen is on or off. Some hand back text, some coach or plan, and some write a file only after you review what will be saved.
/sharpen:sharpen-promptWrite a prompt for another reader
For prompts you will give ChatGPT, another agent, a ticket, or a teammate. It adds the context a reader outside this Claude Code session needs, then hands back the text without doing the work.
/sharpen:sharpen-prompt fix the empty state — for a ticket/sharpen:sharpen-decompileUnderstand or adapt a prompt
Analyze a prompt you already have, name its weak spots, adapt it, simplify it, or split it. The prompt is treated as material to analyze and is never run.
/sharpen:sharpen-decompile make the dashboard better/sharpen:sharpen-handoffMove work into a fresh session
Prepare a starting prompt with the goal, decisions, completed work, failed attempts, next steps, and checks. Useful before /clear or when handing work to a teammate.
/sharpen:sharpen-handoff/sharpen:sharpen-splitBreak a large request into waves
Separate a request into outcomes with their own finish lines, order prompts by dependency, and group independent work into waves. Sharpen writes the plan; you run the prompts.
/sharpen:sharpen-split rebuild settings, tests, and docs/sharpen:sharpen-interviewTurn a fuzzy feature into a spec
Answer one consequential question at a time. Once decisions are clear, Sharpen writes SPEC.md; it does not build the feature during the interview.
/sharpen:sharpen-interview add team invitations/sharpen:sharpen-planPlan a project in phases
For a new project or a big feature. Sharpen sizes the work first: a small change gets a normal prompt, a big one gets kickoff rules, a spec, a plan, a task list and a reusable build prompt. It shows every file and writes them only after you approve; it never builds the feature itself.
/sharpen:sharpen-plan team invitations with roles/sharpen:sharpen-reviewImprove how you prompt
Get coaching on recent prompts in this conversation: what was missing, what that cost, a sharper example, and one practical habit to try.
/sharpen:sharpen-review/sharpen:sharpen-saveSave a prompt you repeat
Turn a recurring task into a reusable project slash command. Sharpen generalizes the changing details, shows the proposed file, and asks before writing it.
/sharpen:sharpen-save check-release/sharpen:sharpen-auditFind project knowledge an agent cannot see
Check guidance, test commands, conventions, project boundaries, and available tools. Get ranked, paste-ready suggestions; the audit changes no files.
/sharpen:sharpen-audit/sharpen:sharpen-helpGet the quick reference
See the current on/off state, controls, review-card choices, companion skills, status badge, and file locations in one compact reference.
/sharpen:sharpen-helpChoose when the automatic gate runs
/sharpen onEnable prompt review. The choice persists between sessions./sharpen offDisable automatic review. Companion skills can still be called directly./sharpenShow whether automatic Sharpen is on or off, and how many sharpened prompts needed a follow-up correction.raw: <request>Skip automatic sharpening for this one request.sharpen: <request>Ask for a review on this request even if automatic Sharpen is off./sharpen:sharpen onThe slash-command spelling also works inside Claude Code.Other slash commands are not intercepted. A pasted sharpened prompt is recognized so it is not sharpened a second time.
Read the status badge
If there is no status line yet, turning Sharpen on can add its badge. If you already use one, Sharpen leaves it alone and offers to chain its badge after yours.
[SHARPEN - CLEAR - READY]The clear request's review card is waiting.[SHARPEN - FILL-IN - READY]One likely reading has been filled in; review it.[SHARPEN - AMBIGUOUS - READY]Different possible actions are waiting for your choice.[SHARPEN - … - SENT]You chose a card option and work is underway.[SHARPEN - SENT]You pasted and sent a sharpened prompt.[SHARPEN - RAW]You bypassed it with the raw: prefix.What stays on your machine
- No API calls or telemetry. Sharpen runs through the AI agent you already use; it does not send prompts to a Sharpen service.
- No prompt log. Hooks inspect the prompt to make the decision and exit; they do not keep a prompt history.
- Hook settings are summarized carefully. The session summary keeps hook event and script names, not hook arguments that could contain tokens. Settings are read, not rewritten as part of that summary.
- The clipboard is used. When the review card appears, the engineered prompt is copied there so you can paste it if you choose.
Local support files include ~/.config/sharpen/state.json, ~/.config/sharpen/last-sharpened.txt, ~/.config/sharpen/stats.json (two counts only: cards shown and follow-up corrections, no prompt text), and ~/.config/sharpen/sharpen-statusline.js. Sharpen only edits ~/.claude/settings.json to add its badge when you have no status line; it does not overwrite an existing one.
The plugin has zero dependencies. Its hook and skill files are plain JavaScript and Markdown.
Update, restart, and check common issues
Update the marketplace and plugin, then restart Claude Code so the new version loads in a fresh session. Run these as separate prompts:
No review card appears
The session may still have an older plugin version loaded. Restart Claude Code, check that Sharpen is on with /sharpen, and try again.
The status badge is missing
If a status line already existed, Sharpen did not replace it. Add node "~/.config/sharpen/sharpen-statusline.js" to the existing status-line command if you want the badge alongside it.
Copying fails on Linux
Sharpen uses an available clipboard tool: wl-copy, xclip, or xsel. Install one if none is present. macOS uses pbcopy; Windows uses clip.
Common questions
For a clear request, it keeps the ask intact and adds helpful context. If a rewrite could change what you meant, you can choose the original from the card.
There is one review decision per sharpened prompt. Use raw: for a request you want to send unchanged, or turn automatic review off with /sharpen off.
It uses some of your current agent session's context: the method and setup summary load once per session, with a short reminder on each prompt. There is no separate Sharpen API or service fee.
It uses what Claude Code lists for skills, agents, MCP tools, and plugin commands, plus the hook and enabled-plugin summary it reads at session start. It does not execute hooks while gathering context.
The current plugin is for Claude Code. The project describes support for other agents as future work; do not assume the automatic review flow is available elsewhere today.
From the Anthropic, OpenAI and Google prompting guides, open-source projects such as the DAIR.AI Prompt Engineering Guide, the OpenAI Cookbook and GitHub Spec Kit, published studies on prompt and agent failures, and 2025–2026 developer surveys. CREDITS.md lists every source and what it contributed, and model-specific settings carry the date they were checked.
No. Nothing leaves your machine. /sharpen shows two local counts, the review cards shown and the follow-ups that pushed back, so you can see whether sharpened prompts land the first time. No prompt text is stored.
No prompt tool can guarantee that. Sharpen makes the requested work, boundaries, and checks clearer, and lets you review the prompt before it starts. The agent still needs to do and report the work.
For the full method, real review-card screenshots, evaluation notes, and source details:
Read the complete plugin README ↗Ask a question or report an issue ↗