mods for your AI coding agent
Install • Skills • Before/After • Benchmarks • Contributing
15 Claude Code skills that change how your agent talks, thinks, debugs, and reviews code. One command install. Each skill is a SKILL.md file — drop it in, type a slash command, work differently.
Some save tokens. Some change the workflow. Some are just fun.
npx skills add repowise-dev/modpack # all 15 skills
npx skills add repowise-dev/modpack/caveman # just oneActivate any skill: /<skill-name>
Deactivate: /default or "normal mode"
That's it.
|
|
|
|
|
|
Same fix. Different brain.
| Skill | What it does | Activate |
|---|---|---|
| caveman | Drop articles, filler, preamble. Fragments. Saves ~84% output tokens. Inspired by JuliusBrussee/caveman. | /caveman |
| hemingway | Short sentences. Active voice. No adverbs. No throat-clearing. | /hemingway |
| intern | Ask max 1 question, then execute. Zero preamble, zero summary. Bias to action. | /intern |
| five-whys | Walk a root-cause chain (Why 1 → Why 2 → ... → Root) before proposing any fix. | /five-whys |
| monk | Enforce minimum viable complexity. Push back on over-engineering. "Do we need this abstraction yet?" | /monk |
| war-room | Every bug = production incident. Severity → Blast radius → Mitigation → Root cause → Comms draft. | /war-room |
| ducky | Only ask questions. Never give answers. User finds the bug themselves. | /ducky |
| archaeologist | Treat the codebase as a dig site. Every file gets historical context, era, tech debt surfaced. | /archaeologist |
| time-traveler | Review code as a dev from 2030. What aged well, what became debt, what didn't survive. | /time-traveler |
| Skill | What it does | Activate |
|---|---|---|
| senior-at-3am | Paged at 3am. Maximum terse. "Line 47. Null check. You're welcome." | /senior-at-3am |
| gordon-ramsay | Code review in Ramsay's voice. Loud, specific, always explains WHY. | /gordon-ramsay |
| passive-aggressive | Perfect work. Disappointed tone. "Fixed the null check. Again." | /passive-aggressive |
| fortune-teller | Predicts 3 things that could go wrong before you ship. "The cards reveal..." | /fortune-teller |
| noir | Debugging narrated as detective fiction. The bug is the criminal. | /noir |
| conspiracy | Every bug is intentional. Someone wanted this. "The git log says Dave. It's always Dave." | /conspiracy |
Real token counts from Gemini 3 Flash (reproduce it yourself). Each prompt runs 3 trials at temperature=0, we take the median.
| Skill | Baseline (tokens) | With skill (tokens) | Saved |
|---|---|---|---|
| caveman | 1,016 | 172 | 84% |
| hemingway | 643 | 131 | 78% |
| Skill | Score | What it measures |
|---|---|---|
| intern | 1.00 | Asks ≤1 clarifying question, then acts |
| five-whys | 1.00 | Walks ≥3 causal whys before fixing |
| monk | 1.00 | Flags unnecessary complexity |
| war-room | 1.00 | All 5 sections: severity, blast radius, mitigation, root cause, comms |
| ducky | 1.00 | Responds with questions only (unless override) |
Note:
time-travelerandarchaeologistare excluded from automated benchmarks — both need real codebase context that the harness can't synthesize.
pip install -r benchmarks/requirements.txt
export GEMINI_API_KEY=your-key
python benchmarks/run.py --dry-run # see the plan
python benchmarks/run.py # full run (~210 calls, ~10 min)
python benchmarks/judge.py # score behavioral skills
python benchmarks/run.py --update-readme # patch the table aboveThe runner checkpoints after every call. Safe to Ctrl-C. Reruns skip completed work.
The archaeologist skill reads more than code. With the Repowise MCP server connected, it pulls real provenance — why a function exists, what it replaced, which decisions it carries.
pip install repowiseWithout Repowise, archaeologist falls back to git history and pattern inference. It still works. It sees less.
See CONTRIBUTING.md for details.
The short version: create a folder, add a SKILL.md, open a PR. Good skills are specific — they name what they do and what they leave alone.
Inspired by caveman by Julius Brussee. The original mod.
MIT — do what you want.