Stay sharp while AI does the work. sharp is a plugin for Claude Code (with Codex packaging) that keeps people learning and judging instead of just approving AI output. It's built on learning research and tested with A/B experiments.
The biggest risk from AI might be human laziness. We get the answer, move on, and our understanding quietly gets shallower. sharp makes the AI work in a way that keeps you thinking, and adds as little to each answer as it can.
- Your judgment stays in charge. Irreversible actions (dropping data, force-pushing shared branches, deploying, publishing, emailing customers, signing) stop for a 6-line Decision Card: the context, what can't be undone, the options, a recommendation, and then your call. A deterministic gate backs it up, and a decision journal brings the call back 30 days later for a retrospective.
- You understand what changed and why. A 4-line Change Brief replaces the usual wall of summary: What & why · Who/where/when · Impact (Verified / Not verified) · Next time.
/sharp:walkthroughgives a 2-minute guided tour of any change. - You reason, not just receive. In coach and learn mode, or on topics you choose to grow in, one question at the right moment, then the loop closes with your rule for next time.
- Lessons come back when they're needed. Lessons are rules with their context ("when X, in context C, do Y, because Z, unless U"). When a similar situation appears, sharp asks: "You've seen this before: what did you learn, and is anything different here?"
- It stays out of the way. At most one question per reply, no coaching during incidents, one line for mechanical edits, and
#skipand#ship 3halways work.
| Scenario | Without sharp | With sharp |
|---|---|---|
| "Drop this table" / "delete the test accounts" | went ahead without asking, every time | stopped, explained, offered a backup, asked: 100% |
| Junior persona journeys: prompted to reason | 8% | 75% |
| …left with a reusable rule | 17% | 75% |
| …words per conversation | 543 | 432 |
| A similar bug to one you've fixed before | 0% invited recall | 100% (near transfer) |
| Student asks for the homework answer | answer handed over | hint-first coaching |
| Production incident | — | zero extra friction |
Also measured honestly: proportionality is 0.67 with sharp vs 0.83 without; far transfer across domains is 0% in every arm; and one research idea I expected to win (a semantic lesson index) lost. Everything is in docs/EXPERIMENTS.md.
/plugin marketplace add pipeshub-ai/sharp
/plugin install sharp@pipeshub-pluginsThat's it. sharp works out of the box with sensible defaults (mode light, detail brief, role general). After installing, Claude Code may say "5 userConfig options not yet set"; that's optional, and you can skip it.
To customize it later (optional), run /plugin configure sharp@pipeshub-plugins, open /config, or use /sharp:mode and /sharp:grow. See sharp/README.md for every setting, command and gate rule.
- docs/FRAMEWORK.md: the framework, and how sharp behaves in each situation and role
- docs/RESEARCH.md: the learning science behind each feature, with evidence ratings
- docs/EXPERIMENTS.md: every A/B round, including what failed
- docs/LIMITATIONS.md: known limitations and open problems. Help wanted!
- ROADMAP.md: what's next, including a human pilot
learning-opportunities pioneered optional learning exercises for Claude Code and Codex, and Anthropic's learning and explanatory output styles inspired the coaching modes. sharp adds the enforced one-way door gate and decision journal, transfer-first lessons with just-in-time recall, the Change Brief, and an experiment suite.
Everything stays on your machine (~/.sharp/). No telemetry. Prompt text is never stored.
sharp has ten open problems, from far transfer to a human pilot, each with a done when criterion. Start with docs/LIMITATIONS.md. The quickest wins are labelled gate cases and realistic scenarios or personas from your field. See CONTRIBUTING.md.
MIT © 2026 PipesHub