Skillquarium

A collection of AI agent skills, organized as an Obsidian vault for easier human navigation.

• managed by vercel’s skills
• place at ~/.agents (the repo root is the skills CLI’s home; skill folders live in skills/)

Screenshot

Setup

mkdir -p ~/.agents && cd ~/.agents
git init
git remote add origin [email protected]:stanfish06/skillquarium.git
git fetch origin
git checkout -f master
./skillquarium install

install symlinks every vault skill into each agent’s skills folder (npx skills add . -s '*' -g). It needs Bun; the skillquarium launcher installs the CLI’s own dependencies on first run.

gstack, career-ops, and UI/UX Pro Max are optional and are skipped by default. Opt in with --extras:

./skillquarium install --extras gstack          # Garry Tan's gstack workflow
./skillquarium install --extras career          # career-ops workspace
./skillquarium install --extras ui-ux           # UI/UX Pro Max seven-skill bundle
./skillquarium install --extras gstack career ui-ux
./skillquarium install --extras all             # all optional extras
./skillquarium install --extras=gstack,ui-ux    # comma form also works
./skillquarium install --help                   # full flag list

Navigation

• vault/index.md — start here: all skills grouped by domain, plus an A–Z list.
• vault/skills.base — filterable / sortable table (by domain, status, rating).
• vault/recipes/ — goal-oriented workflows that chain skills together.
• vault/maps/ — one map note per domain, with cross-links between domains.
• Scientific Expert Profiles — browse the discipline index and its per-discipline maps; each lists primary experts first, then cross-disciplinary experts, with bridges to broader capability maps.

Skillquarium TUI

Run the OpenTUI interface from the vault root:

./skillquarium

The launcher requires Bun and installs the pinned OpenTUI dependency on first use. It has two tabs, switched with Tab, 1/2, or a click:

• Skills toggles model invocation.
• Benchmarks lists the skills under evaluation in eval/skills/

Keyboard controls (Skills tab):

• / focuses search; arrows or J/K move through results; M marks or unmarks the current row.
• C toggles Claude Code, X toggles Codex, and Space toggles both for all marked rows. With no marks, they affect only the current row. A mixed batch is normalized on; a fully enabled batch is normalized off.
• F cycles status filters; G cycles categories. Shift reverses either cycle.
• Ctrl-S saves all states; Ctrl-R reloads them; Ctrl-P performs the guarded pre-commit reset.
• Q or Esc quits when search is not focused.

The table shows four combined states:

• enabled — Claude Code and Codex may both invoke the skill automatically.
• disabled — only explicit invocation is allowed in both products.
• mixed — the Claude Code and Codex fields disagree; toggling makes both enabled.
• error — malformed metadata must be repaired before the tool will change it.

Save writes .skill-vault/data/skill-toggle-state.json. Reload reapplies both products from that snapshot.

The tool writes disable-model-invocation in SKILL.md for Claude Code and policy.allow_implicit_invocation in agents/openai.yaml for Codex. This prevents skills from loading into agent’s context window automatically, but you can still invoke them manually if needed.

Scriptable commands use the same safe backend:

./skillquarium list
./skillquarium enable academic-paper atac-seq
./skillquarium enable --product claude academic-paper
./skillquarium disable --product codex atac-seq
./skillquarium disable academic-paper atac-seq
./skillquarium toggle academic-paper
./skillquarium save
./skillquarium load
./skillquarium pre-commit-reset
./skillquarium tui --query "single cell"

./skillquarium -h lists every subcommand. Beyond the toggles above: doctor, catalog, preview, build, validate, embed, tokenizer, install, overrides, drift, soften, update, import, eval, query, grep. Each takes --help.

As of 2026-08-09, claude code and pi can reliably toggle skills and reduce context usage, codex is not working well. As a workaround, one can use pi as the harness for gpt models.

Searching the vault

./skillquarium query "raw fastq to enriched pathways"   # ranked skills, --k N, --json, --explain
./skillquarium grep "AnnData"                           # literal text in skills/, grouped by skill
./skillquarium embed                                    # refresh vault/embeddings/ (committed)
./skillquarium tokenizer                                # retrain vault/tokenizer/tokenizer.json (committed)

embed needs a llama.cpp /v1/embeddings URL in the gitignored .skill-vault/config.local.json. No other command contacts it.

query ranks with BM25 over ASCII words (runs of [a-z0-9]), then appends up to query.bpeExtra (default 5) skills that BM25 over the committed BPE model’s word pieces finds and that list lacks, tagged [bpe #N]. The ASCII list itself never changes; --no-bpe or bpeExtra: 0 turns the addon off. The retrain-tokenizer workflow retrains and commits the model whenever skill markdown changes on master; locally, tokenizer retrains it with skill-tokenizer on PATH or at tokenizer.bin; queries encode in-process and never run the binary.

Regenerating the navigation layer

After adding or removing skills, rebuild the wrappers, maps, index, and knowledge graph:

./skillquarium build