Catalogue

Every skill and subagent, as of 2026-10-07. Paths are on each repo's main. How each status was established is in Development; the ⚠ rows are explained in Needs review.

Status

Status	Meaning
Evaluated	A recorded run of the skill's evals, with the result in a commit's Tested: trailer or body
Upstream	Vendored, and byte-identical to the upstream repo's main on 2026-10-07 (licence file aside). Its quality is upstream's; nobody here ran its evals
Upstream, edited	Vendored with local edits, listed in the row
Behind	Vendored from an older upstream version that has since changed
Gated	Only a structural check runs: claude plugin validate, a front-matter parse, or a script's --self-test
Untested	Nothing checks it beyond being used
⚠	Flagged, see Needs review

Tags

Tag	Meaning
FE	What a visitor sees: pages, design, the map, front-end testing
backend	Writing and reviewing application code (Scala, Kyo, the over-engineering checks)
DE	Data engineering: ingesting and shaping data the product reads (the corpus, transcripts)
research	Scientific and written work: explaining, writing prose, citation and releases, the proposal review
workflow	How the team works: MIPs, issues, stacked PRs, porting skills. Outside the four asked-for tags; see the README

marola-site — .claude/skills/

Skill	Tag	Origin	Status	Notes
site-frontend	FE	own	Untested	Entry point that orders the other frontend skills; ships site_check.js, which CI runs as a site gate, not as a skill test. ⚠ R13
citizen-science-site	FE	own	Untested
ptbr-humanizer	FE	own	Untested	pt-BR copy on marola.dev; not the same skill as ww3-gpu's humanizar
frontend-design	FE	anthropics/skills@683bc88	Upstream
webapp-testing	FE	anthropics/skills@683bc88	Upstream	Ships 4 Python helpers
design-taste-frontend	FE	Leonxlnx/taste-skill@b482f7a	Upstream ⚠ R9	1,206 lines
emil-design-eng	FE	emilkowalski/skill@e8a175d	Upstream
review-animations	FE	emilkowalski/skill@e8a175d	Upstream	disable-model-invocation
break-ui	FE	emilkowalski/skill@e8a175d	Upstream
karpathy-guidelines	backend	forrestchang/andrej-karpathy-skills@2c60614	Upstream ⚠ R11
mapbox-cartography	FE	mapbox/mapbox-agent-skills@f71cd72	Upstream ⚠ R8	3 evals, not run here
mapbox-data-visualization-patterns	FE	mapbox/mapbox-agent-skills@f71cd72	Upstream ⚠ R8	3 evals, not run here
mapbox-maplibre-migration	FE	mapbox/mapbox-agent-skills@f71cd72	Upstream ⚠ R8 R10	3 evals, not run here
mapbox-style-patterns	FE	mapbox/mapbox-agent-skills@f71cd72	Upstream ⚠ R8	3 evals, not run here
mapbox-style-quality	FE	mapbox/mapbox-agent-skills@f71cd72	Upstream ⚠ R8	3 evals, not run here
mapbox-token-security	FE	mapbox/mapbox-agent-skills@f71cd72	Upstream ⚠ R8	3 evals, not run here
mapbox-web-integration-patterns	FE	mapbox/mapbox-agent-skills@f71cd72	Upstream ⚠ R8 R10	3 evals, not run here
mapbox-web-performance-patterns	FE	mapbox/mapbox-agent-skills@f71cd72	Upstream ⚠ R8	3 evals, not run here

marola-devkit — the marola-devkit plugin, plugins/marola-devkit/

The plugin as a whole passes claude plugin validate --strict on every devkit change, which makes every row here at least Gated for its front matter.

Skill or agent	Tag	Origin	Status	Notes
mip	workflow	own	Gated
mip-tasks	workflow	own	Gated ⚠ R4	disable-model-invocation
mip-solve-perpetual	workflow	own	Gated ⚠ R1	Unattended overnight runner
triage	workflow	own	Gated ⚠ R4	disable-model-invocation
sharingan	workflow	own	Gated ⚠ R4 R7	3 output + 12 trigger evals written, never run
skill-copy	workflow	own	Gated ⚠ R7	Alias stub for sharingan
obsidian-vault	workflow	rewritten from edvmorango/nix-home-config@e541e6e (no licence upstream)	Evaluated (1 of 5)	Eval 2: 4/4 assertions with the skill, baseline without it missed --continue and the stale note (marola-dev/marola#410)
voice-note-ingest	DE	own	Gated	scripts/transcribe.py --self-test runs in tests/self-tests.sh
voice-to-feature	workflow	own	Gated ⚠ R2	disable-model-invocation
humanizer	research	blader/humanizer v3.0.0	Behind ⚠ R5	Upstream is v3.1.0 (225a6f3)
ponytail	backend	DietrichGebert/ponytail@552acd5	Upstream
ponytail-review	backend	DietrichGebert/ponytail	Behind ⚠ R6
ponytail-audit	backend	DietrichGebert/ponytail	Behind ⚠ R6
agent mip-reviewer	workflow	own	Gated ⚠ R4
agent mip-claims-auditor	workflow	own	Gated

marola (the umbrella) — .claude/skills/

Skill	Tag	Origin	Status	Notes
eli5	research	own	Untested ⚠ R12	Every path it cites was checked to exist when it moved in (#652)
architecture-diagram	workflow	adapted from Cocoon-AI/architecture-diagram-generator@4b9087d (MIT)	Untested ⚠ R8	2 output + 12 trigger evals; only checked to parse
zenodo-release	research	own, adapted from ww3-gpu's release	Untested	The scripts it drives (citation.py, release.sh) have self-tests; the skill itself does not
citation-cff	research	zircote/github-social@4fa6579	Upstream ⚠ R8	7 evals, not run here

marola-corpus — .claude/skills/

Skill	Tag	Origin	Status	Notes
corpus-doc	DE, research	own	Evaluated	14 trigger queries via skill-creator's run_eval.py, 1 run each: 10/14, 0 false positives (4d55a45, 2026-09-06)

marola-app — .claude/agents/

Agent	Tag	Origin	Status	Notes
jar-verifier	backend	own	Untested	Decompiles the pinned Kyo jar with javap

h0ffmann/ww3-gpu — .claude/skills/ and .claude/agents/

.claude/hooks/check_agent_frontmatter.py parses every skill's and agent's front matter on each change, so every row is at least Gated.

Skill or agent	Tag	Origin	Status	Notes
eli5	research	own	Gated ⚠ R12	Wave-modelling and HPC vocabulary; a different skill from the umbrella's eli5
release	research	own	Gated	release.sh --dry-run tested in 2b7819d; the skill's prompt is not
humanizar	research	fabricioctelles/skills@4fcc2bf (Apache-2.0)	Upstream, edited	Jev section dropped, sibling link changed, audience paragraph added; all noted under its front matter. Otherwise equal to upstream main (f1de632)
humanizer	research	blader/humanizer v3.0.0	Behind ⚠ R5	The devkit copy plus a 5-line audience paragraph
ponytail	backend	DietrichGebert/ponytail@552acd5	Upstream, edited	Audience paragraph added
ponytail-review	backend	DietrichGebert/ponytail	Behind ⚠ R6	Audience paragraph added
ponytail-audit	backend	DietrichGebert/ponytail	Behind ⚠ R6	Audience paragraph added
agent revisor-proposta	research	own	Gated	Backed by proposal_lint.py and proposal_review_gate.py, both with --self-test

Outside the repos

Hoffmann's claude.ai account also carries mip, mip-tasks and mip-solve-perpetual (they load as anthropic-skills:*). They are older copies of the devkit's and point at docs/mips/, a path that no longer exists. ⚠ R3