A Claude Code skill that rewrites your prompts for clarity before they run.
~3 min read · scan the TL;DR if you only have 30 seconds
- What: Intercepts your prompts, evaluates whether rewriting would meaningfully improve them, and shows you the rewrite for approval before Claude acts.
- Why: Most prompts are short, grammatically fine, and under-specified along the dimensions that matter most. Claude runs with them in a direction you didn't choose, and the next turn is spent reorienting.
- Install:
git cloneinto~/.claude/skills/; then/skill prompterin any session. - Try first:
/skill prompter, choose "Try it once," then submit a one-liner like "the tests are flaky, fix them" — see what prompter would have made of it. - Example output: 20 worked examples covering typos, dangling references, threat-model framing, and no-rewrite cases.
- Maturity: v1.2.1; manual install (plugin packaging on the roadmap); built and used daily while shipping Stuffolio.
A skill is a markdown file Claude Code knows how to run. When you type /skill prompter, Claude follows the instructions in this skill, intercepts your next prompt, and asks whether you want to rewrite it before acting. You don't have to memorize anything — the skill tells Claude what to do, you approve or skip each rewrite.
prompter intercepts your prompts before execution, evaluates whether rewriting would meaningfully improve them, and — when it would — shows you the rewrite and asks for approval before continuing. Over time, this becomes a subtle feedback loop: you start internalizing what makes prompts effective simply by watching them improve.
The skill is opinionated about what not to rewrite: option selections, permission responses ("yes", "proceed"), follow-up answers, and prompts that are already clear. It only rewrites when there is a real improvement to make.
/skill prompter
Activation modes:
- Try it once: evaluate and (if warranted) rewrite the next prompt, then stop
- Next N prompts: apply prompter to the next N prompts (you specify N), then stop
- Current session only: the evaluate/rewrite behavior applies until the session ends
- Persist via CLAUDE.md: add the rewrite rule to your project's CLAUDE.md so it applies across sessions
- Remove persistent rewriting: strip the rule from CLAUDE.md when you no longer want it
The skill walks you through the choice on first invocation.
prompter's biggest payoff is on prompts that are short, grammatically fine, and under-specified along the dimensions that matter most. A blunt verb plus a familiar complaint is the classic shape: "Fix the flaky tests." "Refactor this." "Audit the auth code." Claude will run with it, but in a direction you didn't choose, and the next turn is spent reorienting.
Original:
the tests are flaky, fix them
Rewritten:
Diagnose flakiness in the test suite before changing anything. Categorize each failing test as: (1) timing-dependent (sleeps, polling, real clocks), (2) order-dependent (shared state, leaking fixtures), (3) environment-dependent (network, filesystem, machine-specific), or (4) genuinely racy (concurrency bugs in the code under test). Produce a ranked table with the test name, category, and confidence. Don't apply fixes yet.
What prompter did:
- Refused to skip the diagnosis. "Fix them" became "diagnose, categorize, rank." The original would have started editing tests before anyone knew which kind of flake was at play — and a fix for a timing flake makes an order-dependence flake worse.
- Pinned the analysis dimensions. Four well-known flake categories that every test engineer recognizes, rather than "whatever Claude thinks is wrong."
- Specified the deliverable shape. A ranked table with name, category, and confidence — reviewable in 30 seconds, not a wall of prose.
- Added the missing guardrail. "Don't apply fixes yet" — the obvious thing the user would have written if they hadn't been mid-thought.
A second case where the under-specification problem is the threat model the user has in their head but didn't write down:
Original:
find a security issue in the auth code
Rewritten:
Audit the authentication code for vulnerabilities, scoped to the OWASP Top 10 categories most relevant to auth: broken access control, cryptographic failures, identification/authentication failures, and security misconfiguration. For each finding, provide: file and line, OWASP category, exploit scenario in one sentence, and severity (critical/high/medium/low). Skip theoretical issues that require an attacker model not present in this codebase (e.g., physical device access if this is a server). Don't propose fixes yet — just the audit.
What prompter did:
- Replaced "a security issue" with a threat model. The singular "a security issue" was the leak: it framed the task as bug-hunting until something is found, then stop. The rewrite reframes it as a scoped audit against a specific industry standard (OWASP Top 10), which means the audit terminates with a complete report instead of stopping at the first hit.
- Cut the categories that don't apply. OWASP has 10; the rewrite picked the 4 that touch auth and explicitly noted to skip the rest. Without this, Claude would either over-audit (covering injection in a code path with no user input) or arbitrarily under-audit.
- Demanded a structured finding format. File, line, OWASP category, exploit scenario, severity. A reviewer can scan this in 60 seconds and decide what to prioritize. A wall of prose can't be triaged.
- Added the scope discipline. "Skip theoretical issues that require an attacker model not present in this codebase" — the line that prevents Claude from listing a CSRF concern on a backend service that has no browser-facing surface.
- Added the same diagnose-before-fix guardrail. "Don't propose fixes yet — just the audit." Findings before remediation.
prompter also handles the smaller, more common cases — typos, dangling references like "that file," stacked questions, prompts that should not be rewritten because they're already clear, and the cases where it correctly holds back to preserve the user's voice. Twenty worked examples covering typos, dangling references, stacked questions, hedging preservation, no-rewrite cases, and threat-model framing are in examples/Prompter-Examples.md.
prompter doesn't take a file or feature argument — there's nothing to scan. It scopes by how long the rewrite behavior stays active. On first invocation it asks:
| Mode | Behavior |
|---|---|
| Try it once | Evaluate the next prompt, rewrite if warranted, then stop. Good first run. |
| Next N prompts | You pick N (5, 10, 50). Counter shown after each rewrite. |
| This session | Rewrite for the rest of this Claude Code conversation. No files modified. |
| Add to CLAUDE.md | Insert a Prompt Rewriting rule into your project's CLAUDE.md so it persists across sessions. |
| Remove from CLAUDE.md | Strip the rule when you no longer want persistent rewriting. |
Fresh vs prior history. prompter is stateless by design — every prompt is evaluated from scratch against the rules in SKILL.md. There's no per-user model that "learns your style," no prior-rewrite cache that biases the next call. Re-invocation in the same session detects an active mode and asks whether to restart the counter, switch modes, or update the CLAUDE.md rule (it never duplicates silently). If you want a different style, edit SKILL.md or override the rule in your project's CLAUDE.md.
prompter is intentionally narrow:
- It does not generate code
- It does not make architectural decisions
- It does not pick libraries or frameworks for you
- It does not edit your prompt history or transcript
It just rewrites the next prompt for clarity, shows you the rewrite, and waits for your approval before the rewritten version goes to Claude.
Note
Manual install only — prompter is not yet packaged as a Claude Code plugin (unlike sibling skills bug-echo, radar-suite, etc.). Plugin packaging is on the roadmap.
git clone https://github.com/Terryc21/prompter.gitGlobal install (all projects):
mkdir -p ~/.claude/skills && cp -r prompter ~/.claude/skills/Project-specific install (one project only):
mkdir -p /path/to/project/.claude/skills && cp -r prompter /path/to/project/.claude/skills/prompter was originally bundled with tutorial-creator (then named code-smarter). It was split into its own repo so the audience that wants prompt rewriting can find it without first finding a tutorial-generation tool. Commit history is preserved.
- bug-echo — sibling-bug scan after a fix
- bug-prospector — forward-looking bug hunt before a release
- workflow-audit — 5-layer SwiftUI behavioral flow audit
- unforget — one-file deferred-work ledger
- radar-suite — 6-skill suite tracing user behavior paths through the app (iOS + macOS)
- skill-reviewer — candid reviews of other Claude Code skills
- tutorial-creator — annotated tutorials from your codebase
Terry Nyberg, Coffee & Code LLC. If prompter has saved you a wasted reorientation turn, a coffee is appreciated. Issue reports about what worked or didn't are more useful.
