skila

Files (10)

• SKILL.md
• SKILL.md.tmpl · assets/
• prompt-templates.md · assets/prompts/
• proposal-row.tmpl · assets/
• proposal-criteria.md · references/
• quality-bar.md · references/
• skill-creator-schema.md · references/
• lint.mjs · scripts/
• scan.mjs · scripts/
• validate.mjs · scripts/

SKILL.md (first lines)

## Overview

**skila** is a TypeScript Claude Code plugin that continuously evolves a
personal skill library from live session memory. It inverts every learner
gap identified in v1: no Python, no fire-and-forget distillation, no silent
overwrites, and no opaque quality gates.

---

## When `/skila` Fires

The `/skila` slash command is the **primary trigger**. Execute this procedure
only when:

1. The user explicitly invokes `/skila`, OR
2. The user says "capture this workflow", "save this as a skill", or
   "add this to my skill library".

If loaded by any other means, do nothing and return immediately.

---

## Distill Workflow

### Step 1 — Read the Proposal Criteria

Read `references/proposal-criteria.md`. Apply the three criteria as a
disjunction (any one criterion suffices for promotion):

- **Criterion 1**: Multi-step, multi-tool workflow that succeeded.
- **Criterion 2**: Reusable workflow likely to recur across sessions.
- **Criterion 3**: User correction or stated preference that should persist.

Extract candidates from the agent's **current in-context conversation memory
only** — do NOT read `~/.claude/projects/` or any JSONL transcript files.

### Step 2 — Inventory Scan

Run the inventory scanner to discover installed skills:

```
node skills/skila/scripts/scan.mjs
```

Capture the JSON list of `{name, description, scope, path}` objects.
If the script exits non-zero, surface stderr and abort.

### Step 3 — LLM Judge Pass

For each distilled candidate, classify as `[NEW]` or `[UPDATE→<name>]`
using the current inventory (Step 2) and the last 30 tool-trace entries
from the session. Emit `[UPDATE→X]` only when X exactly matches a `name`
in the inventory. Apply the Wave-1 soft cap of 8 rows; collapse excess into
`[+N more low-confidence candidates omitted]`.

Row format (see `assets/proposal-row.tmpl` for exact template):
```
[NEW@global]     <proposed-name>     <one-line reason>
[NEW@local]      <proposed-name>     <one-line reason>
[UPDATE→X]       X                   <one-line reason>
[UPDATE→X@staging]  X               <one-line reason> (in staging)
```

### Step 4 — Validate Gate (BLOCKING)

For each approved row, draft a SKILL.md using `assets/SKILL.md.tmpl`.
Run the validator:

```
node skills/skila/scripts/validate.mjs <draft-path>
```

If validate exits non-zero, surface all errors and **do not write** the
skill to `~/.claude/skills/`. Fix errors before proceeding.

### Step 5 — Lint Gate (Advisory)

Run the linter on each approved draft:

```
node skills/skila/scripts/lint.mjs <draft-path>
```

Lint warnings are advisory — they never block a proposal, but append
`(WARN: <reason>)` inline next to the row in the multi-select UI.

### Step 6 — Write to Staging

Validated skills are written to `~/.claude/skills/.draft-skila/<name>/`.
The frontmatter MUST include `skila.status: staging`. The web UI
(`skila serve` → port 7777) shows staged skills for review before promotion.

### Step 7 — User Approval + Promotion

Present the multi-select AskUserQuestion with all rows. For each approved
selection:

- `[NEW@global]` → write to `~/.claude/skills/<name>/`
- `[NEW@local]` → write to `<cwd>/.claude/skills/<name>/`
- `[UPDATE→X]` → bump `skila.version`, append to `skila.changelog`,
  set `skila.parentVersion` to prior version, write to published path.
- `[UPDATE→X@staging]` → write to staging tier; explicit `/skila promote`
  or web UI action required to replace the published version.

---

## Quality Bar

Read `references/quality-bar.md` before accepting any proposal. The three
structural rules are:

1. **Deterministic procedures → `scripts/`**: If a workflow step is a
   deterministic, repeatable Node script (FS walk, frontmatter parse, regex
   validation), ship it in `scripts/` and reference it by path in SKILL.md.
   Use `.mjs` wrappers that delegate to `node dist/cli.js`.

2. **Domain knowledge → `references/`**: Stable, enumerable knowledge
   (schema specs, decision criteria, error tables) goes in `references/`.
   SKILL.md body must direct the agent to read the file.

From luyao618/skila, MIT.

More from luyao618/skila