aithos-rr/aithos-selection/librarian
Claude Code and agent skill by aithos-rr/aithos-selection
- community
Checking that the SKILL.md file is still in the repository.
Process _inbox/ items by classifying them as prompts, skills, agents, subagents, MCP configs, tools, stack notes, or references. Propose destinations with frontmatter or manifest. Bookmark external URLs into references/ (process-reference). Run the auto-pilot evening routine that combines inbox processing and GitHub stars delta-sync into one commit (nightly-sync). Suggest workflows that could integrate a new atom. Propose coherent tags for a file. Use when the user mentions "process inbox", "process this URL", "bookmark this", "import my GitHub stars", "run nightly sync", "evening routine", "where does this go", "tag this file", "suggest workflows for", or when files are detected in _inbox/.
Install
npx vskill@latest install aithos-rr/aithos-selection/librarianSKILL.md
The same SKILL.md was published earlier as aithos-rr/aithos-selection/librarian.
Librarian
1. Purpose
The librarian skill keeps the Aithos Selection repository tidy. It removes the cognitive cost of "where does this go?" by triaging the _inbox/ quick dump zone, generating well-formed frontmatter and manifests for new items, proposing coherent tags from a controlled taxonomy, and surfacing existing workflows that a newly added atom could plug into. The skill is advisory and plan-first: it always shows what it intends to do before touching a file, and it never auto-applies destructive changes.
2. When to invoke
Invoke this skill in any of the following situations:
- The user types one of the trigger phrases: "process inbox", "process the
inbox", "process _inbox/", "triage inbox", "where does this go", "tag this file", "suggest tags", "suggest workflows for <path>", "what workflows could use <path>", "process this URL", "bookmark this", "import my GitHub stars", "process _inbox/links-dump.md", "run nightly sync", "nightly sync", "evening sync", "evening routine", "sync everything", "process inbox and sync stars" (Italian: "sync serale", "controlla l'inbox", "fai la routine serale", "sincronizza tutto").
- The user asks a generic "help me classify/organize/move this" question
about a file in _inbox/ or about a new atom.
- You (Claude Code) detect that
_inbox/contains files other than
.gitkeep and README.md at the start of a session, and the user has not yet asked about them — surface the count and offer to run the process-inbox sub-flow.
- The user pastes a single URL or a Markdown list of URLs and asks the
librarian to bookmark them — run the process-reference sub-flow.
- The user has just added a new atom (prompt, MCP config, tool, stack note,
skill) and asks which agents or workflows might benefit from it.
Do not invoke this skill for changes that the user is making in a deliberate, hand-driven flow (e.g. they are editing an existing prompt's frontmatter themselves). The librarian is for triage, not authoring.
3. Sub-flow: process-inbox
Input
Everything currently inside _inbox/, recursively. Ignore .gitkeep and README.md — those are the folder's permanent scaffolding.
Algorithm
- Enumerate. List every file under
_inbox/recursively. Treat any
subfolder as a single candidate (a folder in the inbox almost always represents a composite — agent, workflow, or skill — that ships as a bundle).
- Classify. For each candidate, pick exactly one type from the table
below. Consult references/classification-heuristics.md for the precise rules.
| Type | Destination shape | |
|---|---|---|
prompt | prompts/library/<id>.md | |
template | prompts/templates/<id>.md (when the file contains {{...}} placeholders) | |
mcp-config | mcp-servers/<name>.json (+ optional companion <name>.md) | |
stack-note | stack/<tool>.md | |
tool-script | tools/<name>.py (or the appropriate extension) | |
skill | skills/<name>/SKILL.md | |
agent | agents/<name>/ (requires agent.md, manifest.yaml, README.md) | |
subagent | subagents/<name>/ (requires the entrypoint .md + manifest.yaml) | |
workflow | workflows/<name>/ (requires flow.md, manifest.yaml, README.md) | |
n8n-workflow | n8n-workflows/<name>.json or n8n-workflows/<name>/ | |
unknown | keep in _inbox/, flag for the user |
Skill vs agent vs subagent — the three-way distinction. All three are agent-shaped primitives but live in different folders:
- `skill` — Anthropic skill format. A folder with
SKILL.mdwhose
frontmatter only needs name and description. No tools or mcpServers declared at the skill level. Often ships with references/, scripts/, assets/.
- `agent` — Aithos-internal declarative agent. A folder with
agent.md (system prompt), manifest.yaml (declares system_prompt, uses.prompts, uses.mcp_servers, uses.tools), and README.md. Used to compose existing atoms into a curated agent we author ourselves.
- `subagent` — Claude Code subagent (Anthropic primitive,
invoked as /<name> and deployed to .claude/agents/). A folder with an entrypoint .md file whose frontmatter declares tools, mcpServers, optional skills dependencies and memory. The Aithos-side manifest.yaml mirrors these for indexing. Often ships with supporting material: BUILD-BRIEF.md, ARCHITECTURE.md, PROGRESS.md, references/, discovery/, research/, test-fixtures/.
When a zip extracts to <name>/<name>.md with tools: / mcpServers: in its frontmatter, classify as subagent. When it extracts to <name>/SKILL.md, classify as skill. When it has agent.md, classify as agent.
- Generate metadata. For each classified candidate, draft:
- The proposed final path (kebab-case stem matching the
id/name
field).
- The proposed frontmatter (for atoms) or manifest (for
composites and recipes). Fill every required field from the file's content; use today's date for created and updated; default author: riccardo; default status: draft unless the file content explicitly says otherwise; pick version: 0.1.0 for new items. See references/frontmatter-examples.md and references/manifest-examples.md for the canonical shapes.
- 3–6 proposed tags following
references/taxonomy.md. At least one
category tag, one domain tag, one language tag. Flag any tag that is not in the taxonomy as new tag, confirm with user.
- Present the plan. Produce a single consolidated Markdown table:
```
| # | Source | → | Destination | Type | Tags | |
|---|---|---|---|---|---|---|
| 1 | _inbox/foo.md | → | prompts/library/foo.md | prompt | extraction, italian, legal | |
| 2 | _inbox/bar.json | → | mcp-servers/bar.json | mcp-config | filesystem | |
| 3 | _inbox/baz/ | → | agents/baz/ | agent | classification, italian |
```
Below the table, show the generated frontmatter or manifest for each row in sequence (or under collapsible sections if the renderer supports it). Make every proposed change inspectable before any file moves.
- Ask for approval. Prompt the user verbatim:
> Apply this plan? (yes / no / partial)
- yes — perform the moves, write the frontmatter/manifests, then
commit with chore(librarian): processed inbox (<N> items) where <N> is the number of items moved.
- partial — ask the user to list the row numbers (e.g.
1, 3) to
apply, then proceed with that subset only. Commit the same way with the actual count moved.
- no — stop without changes.
- Apply. When applying:
- Use
git mvso history follows the file when possible. - Write the frontmatter or manifest before staging, so the moved
file is already well-formed.
- For composites (
agent,workflow,skill), create every required
sibling file (agent.md, manifest.yaml, README.md, etc.) even if stubbed — the destination must satisfy check.py.
- Never delete
_inbox/itself or its.gitkeepandREADME.md. The
folder stays for future use.
- Report. After committing, summarize what was moved (counts by type)
and what remained in _inbox/ as unknown, so the user can decide next steps.
Failure modes to watch
- A candidate matches more than one type → see the tie-breaking rule at the
end of references/classification-heuristics.md.
- A candidate is a folder but missing required sibling files (e.g. an
agents/ candidate with no agent.md) → classify as unknown and ask the user whether to stub the missing pieces.
- A proposed destination already exists → never overwrite. Ask the user
whether to suffix (-v2), merge, or skip.
4. Sub-flow: process-reference
Purpose
Bookmark an external URL (or a batch of URLs) as a reference under references/. References are pure curation — pointers to GitHub repos, articles, or workflow templates the user wants to remember without ingesting their content. They are not dependencies: composites and recipes never use: a reference.
Input modes
The sub-flow accepts three input modes, picked from how the user invokes the skill:
- Single URL. The user passes one URL inline
(e.g. "bookmark https://example.com/foo").
- List of URLs. The user passes a Markdown file — by convention
_inbox/links-dump.md — containing one URL per line. Blank lines and Markdown comments are ignored; lines like - https://… or [label](https://…) are accepted, the URL is extracted.
- GitHub stars bulk. Triggered by the user phrase "import my GitHub
stars" (or a close variant). The skill fetches the complete list via gh api users/<username>/starred --paginate. The user must supply their GitHub username on first run.
Algorithm
For each URL:
- Classify subtype.
- URL matches
https://github.com/<owner>/<repo>with no further path
segments → repo.
- URL matches a known template-hosting pattern —
n8n.io/workflows/…,
a gist whose body looks like a workflow/skill/agent template, or similar — → template.
- Anything else →
article.
- Extract metadata.
- For
repo: callgh api repos/<owner>/<repo>and readname,
description, stargazers_count, language, topics, pushed_at. Populate the optional GitHub fields in the frontmatter (github_owner, github_repo, github_stars, github_language, github_topics, github_last_commit).
- For
articleandtemplate: a best-effort web fetch retrieves the
page <title> and the meta[name="description"] content. If the fetch fails (network error, 4xx/5xx, paywall), leave description empty and add flag for user review next to the row in the plan.
- Generate frontmatter with the schema-required fields from
/docs/schemas/reference.schema.yaml: id, name, type: reference, subtype, url, status: active, description, tags, language, created, updated, author: riccardo. Add the optional GitHub fields only when subtype: repo.
- Propose destination path:
references/<subtype>s/<id>.md.
- For
repo:<id>is<owner>-<repo-name>(kebab-case). - For
articleandtemplate:<id>is a slug derived from the
page title (lowercase, hyphenated, stripped of punctuation, capped at ~50 characters).
- If the destination already exists, suffix with
-v2(or higher)
and surface the collision in the plan.
- Compose body with the canonical template (frontmatter, the
# <Name> heading, a Markdown link to the URL, a ## Why this is interesting section left empty, and an optional ## Notes section). The "Why this is interesting" block is left blank by default — the user fills it later.
Output
Present a consolidated plan as a Markdown table, one row per URL:
This is the start of the file. Read the full SKILL.md on GitHub.
Security
- Security scanPassed, 95 of 100Pattern checks on the skill's files.
- Model reviewPassed
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingskills/librarian/SKILL.md
Popularity
- 7-day trend score
- 0
Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/aithos-rr/aithos-selection/librarian)