Skip to content

skills.md

Latest commit

 

History

History
146 lines (107 loc) · 4.13 KB

skills.md

File metadata and controls

146 lines (107 loc) · 4.13 KB
Raw

🦋 Skills

Skills are reusable instruction packs — the difference between “generic coding bot” and “knows how we ship”. ✨

📜

Hygiene rule: vague skills make vague agents. “Be careful” is not a skill. “Never rewrite unrelated files” is a skill.

Mental model 🧠

bundled skills  +  project overrides  +  learned skills
        └──────────► matched into TaskPacks 📦
Kind Where Who writes it
📦 Bundled shipped with SLMCode maintainers
🏠 Project .slmcode/skills/ you
🦋 Learned .slmcode/skills/learned/ the flywheel

Index file: .slmcode/SKILLS.md (auto-maintained — don’t micro-manage unless you must).

SKILL.md shape 🧬

---
name: atomic-coding
description: Prefer tiny diffs, clear acceptance checks
triggers: refactor, cleanup, helper
agents: worker, deep, corrector
paths: "**/*.go, cmd/**"
user-invocable: true
---

# Atomic coding

- One concern per task
- Touch only listed files unless discovery proves otherwise
- Leave tests greener than you found them
Field Meaning
name Stable id (@skill:name)
description Human + matcher hint
triggers Keywords that boost matching
agents Which specialists see it (* = all)
paths Gate the skill on the files a run actually touches (see below)
user-invocable Can users pin / reference it?

paths: — gating a skill on the files in scope 🎯

Context is the scarcest resource a small model has, and a Go-specific skill in a TypeScript task's prompt is pure noise. paths: is a comma-separated list of globs supporting *, ?, […], ** and a bare directory prefix:

Situation Result
skill has no paths: ungated — participates exactly as before
skill has paths:, and at least one file in scope matches participates
skill has paths:, and nothing in scope matches left out of the prompt
the scope is empty / unknown gating is disabled; the skill participates

That last row matters: slmcode skills list, Studio's skills page and any caller that does not yet know which files a run will touch pass an empty scope, so a gated skill is never hidden from you — it is only kept out of prompts where it could not apply.

An explicit @skill:name in the query, or a pinned_skills entry in config, always wins: you naming a skill outranks a heuristic about file extensions.

Day-to-day commands 🛠️

slmcode skills                 # list
slmcode skills show atomic-coding
slmcode skills new my-skill --agents worker
slmcode skills edit my-skill   # project override

Pin or reference 📌

slmcode run --skill atomic-coding "Refactor helpers"
slmcode run "Fix login @skill:multipass-quality"

Studio → Skills panel has pin chips. Config:

pinned_skills:
  - atomic-coding
  - multipass-quality

Bundled starters (taste) 🍿

Skill Vibes
atomic-coding Tiny diffs, clear done ✅
multipass-quality Think → critique → refine 🔁
markdown-memory Treat CONTEXT/MEMORY as sacred 💾
engine-full-pipeline Full orchestrated run behavior 🏭
specialist-* Role-specific playbooks 🧩

Exact set evolves — slmcode skills is truth. Docs can lie; the CLI rarely does.

Teaching the flywheel 🦋

After good runs, check:

slmcode docs show SKILLS.md
ls .slmcode/skills/learned/

Promote hard-won lessons into a project skill when they stop being optional.

!!! tip "🧹 Skill hygiene" Prefer sharp constraints over motivational posters. Your workers are interns with amnesia — write the checklist.

Related 🔗

☀️ Made with ♥ by UnicoLab