Agent skills
Skills for Claude Code, Codex, pi, and other agents that read SKILL.md folders. Each skill works on its own. yesh-router picks a skill for each task and applies the engineering principles in references/principles.
Turn on powerball mode
Powerball mode sends yesh-router to the agent on every turn, so each task starts with a method. Set it up for each host.
Claude Code
-
Build the output style from the installed router:
python3 scripts/build-mode.py # writes ~/.claude/output-styles/powerball.md
It also installs the per-turn reminder: a UserPromptSubmit hook in ~/.claude/settings.json that runs turn-reminder.py and adds one line asking the agent to re-match each new ask. The router text loses weight deep into long threads; this line sits next to the newest message. Pass --no-hook to skip it. T3 Code runs the same claude binary, so it gets the hook too.
When your login shell is not bash, it also sets CLAUDE_CODE_SHELL to a bash 4+ (brew install bash on macOS). Agents write bash; under an interactive zsh their commands fail on unmatched globs (a ? in a URL), echo ====, unsplit $VAR, and .zshrc aliases like ls=eza. That was about 5% of Bash calls before the change. An existing CLAUDE_CODE_SHELL is left alone, and --no-shell skips this. It also warns when timeout is missing (brew install coreutils).
It also sets "attribution": {"commit": "", "pr": ""}, so agent commits carry no Co-Authored-By trailer and PR bodies no "Generated with" line. --no-attribution skips this.
-
Link the worker subagent, so delegated steps start with the router loaded:
ln -sfn "$PWD/config/agents/powerball-worker.md" ~/.claude/agents/powerball-worker.md
-
Add "outputStyle": "powerball" to ~/.claude/settings.json, then restart Claude Code.
The status line segment in config/statusline is optional. It shows which skills a session read.
pi
Install the pi-powerball extension (private repo), then restart pi:
pi install git:github.com/Yeshwanthyk/pi-powerball
It reads the router from ~/.pi/agent/skills/yesh-router/SKILL.md and is on in every new session. Toggle it with /powerball on or /powerball off. To confirm it found the router, run pi -p "/powerball status".
Per-turn reminder by host
| Host | Reminder |
|---|
| Claude Code, T3 Code | On, installed by build-mode.py |
| pi | Not set up. Port it into pi-powerball (run turn-reminder.py on each user message) only if the session audit shows the header fading in long pi threads too |
| Codex | Not set up; check for a per-prompt hook first |
Whether it helps is measured with python3 scripts/audit-sessions.py --days N, which splits header and read rates by turns with and without the reminder.
Skills
| Skill | Use it to |
|---|
architect | Pick a target architecture from the real contracts, call paths, and state ownership. |
benchmark-checklist | Check a performance number before you report it or act on it. |
blast-radius | Find what a change could break beyond what grep shows. |
correct | Stop a mistake agents keep repeating in a repo, with types, checks, or tests. |
deep-dive-explainer | Research a topic and write it up as a long offline article. |
diagnose | Reproduce a bug, find the cause, and prove the smallest fix. |
how | Explain how code works and why it was built that way. |
interactive-explainer | Build an interactive HTML model of behavior or state. |
interrogate | Run independent adversarial reviews and merge them into one judgment. |
|
Name a skill to run it: /how in Claude Code, /skill:how in pi, or "use how" in prose. Skills marked disable-model-invocation: true stay out of the agent's skill list, so reach them by name or through the router.
Optional tools
None of these are required. A few skills call them when they are installed.
| Tool | Used by |
|---|
gh | show-me PR workflow |
calldiff (Node 22+) | show-me, viz, meat-review |
Quint (JDK 17+ for quint verify) or TLA+ | invariants, verification-skill |
| Lean 4 | invariants, verification-skill |
| Kani, Verus, CBMC, or Dafny | invariants, verification-skill |
Changing a skill
- Edit the skill, then run
python3 scripts/check-skill-routing.py to check links and router coverage.
- Run
node tools/skill-map/build.mjs to rebuild the skill map.
- When you retire a skill, playbook, or principle, add it to
scripts/retired-names.json.
- Push, then run
gitgud update --global and gitgud apply --global. A repo edit does not reach installed copies until you do.
- If you changed
yesh-router, run python3 scripts/build-mode.py and restart Claude Code. pi picks up the change on its own.
config/AGENTS.md is the host instructions template. python3 scripts/audit-sessions.py reports how recent Claude Code and pi sessions used powerball.
Credits
research and writing-for-agents are adapted from mattpocock/skills (MIT). The show-me PR sections draw on its pr skill and on Dex Horthy's humanlayer/skills. The router and its playbooks draw on the Cursor pstack plugin (MIT), and the formal branch of verification-skill on its create-formal-verification-skill (0.15.27). The forward implementation first principle draws on Vuk97/forward-implementation-first.
