Free and open source, for Claude Code, Codex, Cursor, OpenCode and OMP
Hooks that refuse destructive commands, ask before a push, send the agent back when it edited code and ran no check after, and show your saved notes when the next session starts. Strongest in Claude Code; Codex, Cursor, OpenCode and OMP get the skills and part of the hooks.
Each dot is one skill: 258 listed, 416 more in a library that bossku skills find searches. The lit one is bosskuai-taste, the skill Bossku picks for websites.
bossku-superpowers
Skills: bosskuai-taste
Memory: yesChecks.
A skill is text the agent may or may not follow. Bossku adds hooks that act at the moment it counts. In Claude Code, git reset --hard, a plain force push and similar commands are refused, a git push waits for your yes, and a turn that edited code with no check after it is sent back once.
Remembers.
The agent is told to save the rules and decisions you give as notes. In Claude Code the newest notes show when the next session starts, and every 50 tool calls the agent is asked to turn useful lessons into small skills. With Obsidian storage, notes and learned skills stay out of the code repo.
Travels.
bossku install makes the same skills available to Claude Code, Codex, Cursor, OpenCode and OMP, and writes seven subagent contracts for Claude Code, Codex, Cursor, OpenCode and Devin. The full hook set runs in Claude Code; the command guard also runs in Codex, Cursor and OpenCode.
Measured.
Benchmarks with and without Bossku are published with their intervals and their weak spots. We found none in the READMEs of Superpowers, ECC, the Karpathy guidelines or Anthropic's skills.
bossku hooks install
One turn of your agent in Claude Code, and the points where Bossku hooks step in.
git reset --hard are refused, and a git push or production deploy waits for your yes. A push hidden inside a script is not seen.bossku skills find "<task>"
Pick a request or something your agent tries to do. The skill names are what bossku skills find really returned for each request.
You ask
Your agent tries
build a landing page for my bakery
The prompt hook suggests a skill before the agent starts.
bossku-superpowers
Skills: bosskuai-taste
Memory: yesWhy: it is the house design rule for any website, landing page or portfolio.
buatkan website untuk kedai kopi saya
Some Malay phrasings work too, like this one. Same skill as the English request.
bossku-superpowers
Skills: bosskuai-taste
Memory: yesWhy: "buatkan website" is one of the Malay phrasings the router knows.
this test passes locally but fails on CI
The prompt hook suggests a skill before the agent starts.
bossku-superpowers
Skills: ci-triage
Memory: yesWhy: it finds the failing job and step, then sorts it into flake, regression, environment or config.
review this pull request before I merge
The prompt hook suggests a skill before the agent starts.
bossku-superpowers
Skills: bosskuai-rigorous-code-review
Memory: yesWhy: a skeptical pre-merge review for correctness, risk, missing tests and whether it meets its goal.
git push origin main
Asks you first
In Claude Code the push waits for your yes. The same goes for a PR merge, npm publish, vercel --prod, terraform destroy and similar. A push hidden inside a script is not seen. Claude Code's own permission rules can do this too; Bossku ships the rule set ready-made, and its ask still shows when you have allowed commands broadly. Codex and OpenCode 1 refuse instead of asking.
git reset --hard
Refused
Commands such as git reset --hard, a plain force push or rm -rf ~ are blocked, and the agent is told to give you the exact command and recovery steps to run yourself. A --force-with-lease push asks instead. Runs in Claude Code; the Codex, Cursor and OpenCode versions are unit-tested against each tool's hook format, not yet run inside those apps.
Edited 3 files. Done!
Sent back once
If the agent changed code and ran no test, lint, build or run after its last edit, the stop gate sends it back once. It checks that something ran, not that it passed. Switch it off with BOSSKU_STOP_GATE=off.
new session starts
Project notes, grouped by kind, newest first:
- [decision] Use the existing test runner for this project.Rules and decisions the agent saved as notes are shown at the start of your next Claude Code session, so you do not repeat yourself. Only the newest fit, about 1,300 characters.
Much of this sits on features Claude Code has built in. Here is what is native and what Bossku adds on top.
| Area | Built into Claude Code | What Bossku adds |
|---|---|---|
| Skills | Loads a skill when its description fits the request. | 670+ skills, most vendored or adapted from other open-source packs, and a hook that checks every prompt and names fitting skills when the match is clear. |
| Memory | CLAUDE.md files and its own auto memory. | Notes per project shown at session start, optional Obsidian sync, and auto-learn. Not benchmarked against the native memory. |
| Ask before push | Permission rules can allow, ask or deny a command by its pattern. | A ready-made rule set. Cursor and OpenCode 2 ask too; Codex and OpenCode 1 refuse and hand the command to you. |
| Run your code | Gives you Stop hooks to write your own check. | The check itself: one push back if code changed and no check ran after the last edit. |
| Subagents | Supported natively. | Seven written contracts: copied as they are for Claude Code, rewritten for Codex, Cursor, OpenCode and Devin. |
Skills and ideas come from Superpowers, ECC, marketingskills, MengTo and other packs, each credited with its license.
bossku memory-brief --project .
Real agents, the same tasks, with and without Bossku Superpowers. The chart shows two-session tasks: a rule is set in one session and needed in the next. Read the setup first: the "without" runs had no Bossku setup, no CLAUDE.md and Claude Code's own auto memory switched off. The Claude runs still used the tester's own Claude Code config, which listed 20 skills and 2 plugins at start.
81%
An acceptable skill was ranked first on 158 requests the router was never tuned on, against 56% for a keyword baseline that uses only skill names and descriptions. Acceptable skills were labelled by AI models, not people. Not compared with Claude Code's own skill picking. With a real agent, an acceptable skill was opened for 45% of 60 requests.
62% to 92%
HumanEval problems passed by the small Nemotron 3 Nano 30B, a model that usually skips running its code: 25 of 40 without, 37 of 40 with. The runs without Bossku were done on other days.
1.0× to 4.7×
The catch: about the same or more tokens per run, highest on models that checked the least before. On coding and harder tasks the three open models tested showed no clear change in pass rate; Claude models were not run on those.
What this does not show: whether Bossku beats Claude Code's own memory or a rule you wrote in CLAUDE.md. That was not tested. Samples are small, so read each result with its interval. Agent runs used build b8ccea1, whose stop check also reminded the agent to save rules; today's default install uses a simpler check. The live skill-use check used build 8279f19. Full numbers and limits: benchmark results and how it was tested.
benchmark_agent.py run --arm superpowers=pack:… --arm ecc=pack:…
Same tasks, same day, Claude Haiku 4.5 and Claude Sonnet 5.5. Bossku (lean profile) next to Superpowers, ECC, Karpathy's guidelines, Anthropic's skills and Claude Code with nothing added.
31–32 of 34
Coding tasks passed by every setup, Bossku included. On coding, no pack won.
0 of 312
Runs in which the agent opened a skill from the four other packs. Bossku's agent opened one in 21 of 78 runs.
+40%
The catch: Bossku cost about 40% more per run than no pack (memory tasks: the graded session only). ECC cost 80% to 90% more. Karpathy's guidelines and Anthropic's skills cost about the same as no pack.
Be fair to the others when you read this. Superpowers, Karpathy's guidelines and Anthropic's skills are not built to carry your rules between sessions. Only Bossku came with instructions (including “save every rule the user states”), so the result is for the whole Bossku bundle. CLAUDE.md and Claude Code's own auto memory were off for every setup, so a rule you write in CLAUDE.md was not tested. The tasks are Bossku's own and each ran once, on Windows. Full method, pinned versions and limits: pack comparison.
bossku doctor --project .
After a turn that changed code, your agent is told to review its own diff with bosskuai-rigorous-code-review: correctness, risk, missing tests, and whether it does what you asked. It fixes what it finds and ends with a review line. Lessons worth keeping go to your notes, and auto-learn is asked to turn findings that keep coming back into small skills.
In Claude Code, the stop hook sends the agent back once if the reply has no review line, in the same push-back as the run-your-code and security checks. It checks that the line is there, not that the review was good. A self-review shares the author's blind spots, so treat it as a floor, not a second opinion.
Review: CLEAN | FIXED <n> | OPEN <n>Off switch: BOSSKU_REVIEW_GATE=off
The benchmark tool can now load another skill pack as a Claude Code plugin and run it on the same tasks. The first run put Bossku next to Superpowers, ECC, Karpathy's guidelines and Anthropic's skills. Bossku kept more of your rules between sessions; on coding, every setup tied.
The full page lists what was not fair to the others, from ECC's hooks failing on Windows to the tasks being Bossku's own. See the comparison.
Try it: --arm NAME=pack:DIR
bossku install
You need Python 3.11+ and Git. The same commands work in Windows PowerShell; if bossku is not found, use python -m bossku instead.
git clone https://github.com/wankimmy/Bossku-Superpowers
cd Bossku-Superpowers
pip install -e .bossku installCopies the skills, writes the subagent contracts for each tool you have set up, and turns on the hooks. Want notes in Obsidian? Add --vault "/path/to/your/Obsidian/Vault" --memory-storage obsidian.
cd /path/to/your/project
bossku init .Adds a Bossku block to AGENTS.md, a CLAUDE.md that imports it, and small .omp/ and .bossku/ files.
bossku doctor --project .Prints doctor: ok, or the issues found and what to fix.
Open the project and start your agent as usual. No magic word: the hooks run on their own. Mostly you will not notice anything, until it stops a mistake. Remove the hooks with bossku hooks uninstall (also turn off the Bossku plugin if you enabled it in Claude Code), or Bossku's user-level setup with bossku uninstall --purge; project files and notes stay.
Setup notes for Claude Code, Cursor, Codex, OpenCode and OMP.