# g4b_ai

A small collection of personal [Claude Code](https://docs.claude.com/en/docs/claude-code) skills, packaged as a plugin and served via the [reliquary](../../README.md) marketplace.

## Agents

All agents are tool-restricted and **cannot spawn sub-agents**.

- **scout** — Lightweight lookup agent (haiku). Extracts small pieces of information from code, files, or a single web page. Read-only, no file writing. Use for quick fact extraction.
- **search** — Read-only search agent (sonnet) for locating code, files, and symbols. Cannot modify files. Use for targeted lookups, symbol tracing, and "where is X" questions.
- **researcher** — Deep research agent (opus, medium effort). Can read, synthesize, and write research documents. Use for complex topics requiring multi-source analysis and written reports.

## Skills

- **workpad** — Decides *where* and *when* to put working documents (reports, plans, brainstorms, designs, research notes, specs). Steers other skills toward a single predictable location for thinking-in-markdown instead of scattering files across `/tmp`, the repo root, or per-skill scratch directories.
- **research** — Governs how research is conducted. Prevents recursive agent spawning (research agents must not spawn sub-agents), caps web fetching (download pages as markdown and scan locally, never crawl), and ensures the orchestrator controls breadth while individual agents control depth. Composes with research-tree and workpad.
- **research-tree** — Orchestrates multi-topic research as a dependency tree of parallel phases. Within a phase agents run in parallel; between phases later agents inherit earlier reports. Use it for "research X, Y and Z" / "do a deep dive on…" prompts where a single agent would skim.
- **implementation-orchestration** — How to drive non-trivial implementation work across one or more agents: when to delegate vs. work inline, how to sequence agents, and how to handle test/compile failures without thrashing. Defaults to sequential execution; parallel only when work is genuinely independent.
- **markdown-embedded-svg** — Rules and defaults for putting SVG into Markdown. Different renderers (GitHub, Gitea, Obsidian, VS Code) sanitize inline SVG very differently, so the default is to save SVGs as separate files and reference them via image-link syntax; inline SVG is treated as a special case.

## Install as a plugin

The reliquary marketplace exposes this plugin as `g4b_ai@reliquary`.

### From `git.g4b.org`

Interactive:

```
/plugin marketplace add git@git.g4b.org:dirigence/reliquary.git
/plugin install g4b_ai@reliquary
```

Or declaratively in a project's `.claude/settings.json`:

```json
{
  "extraKnownMarketplaces": {
    "reliquary": {
      "source": {
        "source": "git",
        "url": "git@git.g4b.org:dirigence/reliquary.git"
      }
    }
  },
  "enabledPlugins": {
    "g4b_ai@reliquary": true
  }
}
```

### From a local clone

```
/plugin marketplace add /path/to/reliquary
/plugin install g4b_ai@reliquary
```

Or in `settings.json`:

```json
{
  "extraKnownMarketplaces": {
    "reliquary": {
      "source": { "source": "local", "path": "/path/to/reliquary" }
    }
  },
  "enabledPlugins": { "g4b_ai@reliquary": true }
}
```

## Fetch individual skills

If you only want one or two skills (no plugin, no marketplace), copy them straight into a target project's or user-level skills directory.

User-level (available in every project):

```sh
git clone git@git.g4b.org:dirigence/reliquary.git /tmp/reliquary
mkdir -p ~/.claude/skills
cp -r /tmp/reliquary/plugins/g4b_ai/skills/workpad ~/.claude/skills/
cp -r /tmp/reliquary/plugins/g4b_ai/skills/research-tree ~/.claude/skills/
```

Project-level (only the current project):

```sh
mkdir -p .claude/skills
cp -r /tmp/reliquary/plugins/g4b_ai/skills/workpad .claude/skills/
```

Or symlink instead of copying so the skill stays in sync with the upstream clone:

```sh
ln -s /tmp/reliquary/plugins/g4b_ai/skills/workpad ~/.claude/skills/workpad
```

A skill is just a directory containing a `SKILL.md` — no further wiring is needed.