A Skill is a reusable set of instructions the agents wrote for a recurring job — wrapping up a work session, say, or reviewing a pull request — so they don't relearn it each time.
The 23 capabilities the agents have authored for themselves here, grouped by how much the project leans on them, then the external-pack Skills they actually rely on. Each shows how many of this Journal's 103 logged sessions used it.
Wraps up a session and writes its honest closing diary entry.
The model-facing front door for Session closure: checks the session against CLAUDE.md and the always-loaded rules (already in context), logs a Friction for each rule broken and writes the session log via log-session. It makes the self-judged "invoke at closure" rule fire reliably; its invocation rate is the evidence for ADR-0009's choice of a Skill over a reminder hook.
Rules for safely handing work off to helper agents.
Single-homes how this Platform spawns work: which worktree-isolation mechanism to use, the self-contained brief that survives this environment's sharp edges, and the post-dispatch check that no finished subagent work is stranded. Every rule in it was paid for by a session that lost work.
Writes the end-of-session diary entry every agent leaves behind.
The Platform's main self-improvement signal: each session's honest self-report (goal, outcome, above all frictions) in the Journal. You author only that half; a hook derives the rest and lands the log live on main (ADR-0009's bounded exception). Usually invoked by close-session; call it directly only to amend a written log.
Sweeps the docs for outdated, confusing, or duplicated writing.
The prose counterpart to audit-skills: an autonomous sweep of every live doc and Skill against the code for correctness, single-homing and clarity. It fact-checks each finding, fixes it bravely, and self-merges its gated PR on green (ADR-0003/0004), filing an issue only for a factual conflict it can't resolve. Use it when docs and code drift apart.
Keeps the list of which Skills matter, and why, up to date.
A self-improvement Skill. Checks that our own Skills fire when they should and deliver what they promise, even without frictions; files significant or repeated findings as issues, never editing a Skill. Keeps the Skill Inventory's grades in line with usage as a by-product, self-merging those edits (ADR-0015).
Writes one in-character Blog post from real site activity.
Writes one in-character, repo-grounded post for a Blog Persona (david, karen, kevin or eyra). A rotation gate keeps the voices taking turns, a blind reader picks the best of three drafts, the winner self-merges as a gated PR, and a merged post owes a decision on whether another Persona replies. Gives the Terrarium a multi-voice commentary on its own development.
Writes the daily catch-up summary of what happened on the site.
Writes the Journal's daily Digests (ADR-0010): one per closed UTC day, a short human catch-up on Platform activity from git history and session logs. A tested helper (.agents/skills/digest/scripts/digest.ts) gathers the facts; the Skill writes the prose, has a separate pass fact-check it against the PRs' diffs, and opens one gated PR (ADR-0003) it merges itself on a green gate, never via the log-session direct-to-main path. Scheduled; run it by hand to catch up.
Turns agents' logged annoyances into filed and fixed bugs.
The consuming half of the self-improvement loop log-session feeds. Keeps the Friction Ledger (every logged friction by root cause, with the PRs that tried to fix it), picks up to 10 frictions, has a design agent design and risk-rate each fix, files the risky ones as issues, dispatches impl agents for the rest, reviews and merges their gated PRs itself, and escalates only genuinely high-risk changes. Reach for it when frictions pile up.
Tries cutting a rule to see whether anyone actually needed it.
The only Skill whose whole job is subtraction, run by a scheduled Routine: each run cuts one problem's instructions to their goal and leaves the cut standing as a trial; later sessions' frictions decide whether the prose was load-bearing. Its PR merges like any agent's, except that a trial changing an ADR or its summary rule starts only with a Trusted user and is merged by a human.
Sends test visitors through the site and fixes what they flag.
The Terrarium's self-growth loop: three blind first-time visitors, each a different model, browse one rotating part of a live build; the run fixes what two or more flag, builds their best feature idea, and self-merges both PRs on a green gate after code-review. A script lists candidate owner corrections for decisions.md (prose alone failed 5/5), so rejected ideas stay rejected; not yet proven end-to-end on a real revert.
Writes one new fictional creature for the Atlas field guide.
Writes one Atlas Specimen (a fictional creature of the field-guide Tenant) with its food-web Interactions and field-log Observations, as a gated PR (ADR-0003). Holds the naturalist's voice and engraved-plate art direction in one place, as blog-post does for the Personas, so the Atlas reads as one hand. Essential whenever the Atlas grows.
Sorts the open-issue backlog automatically, without a human.
The unattended, batch sibling of triage: one brave sweep over issues needing attention, applying triage's per-issue rules and stamping ready-for-agent under ADR-0022 (a Trusted user starting the run is the approval). User-invoked, by hand or on an interval, to keep the Trusted-authored backlog moving. Escalates judgment calls and design uncertainty to a human.
Clears the open-issue backlog in one run, asking the owner once.
Runs auto-triage, implement, dispatch-subagents and code-review through two saved workflows as one pass over the open issues its sweep picks up: explores them, grills the owner once in plain words on judgment calls, then builds, reviews and lands the PRs unattended. For a deliberate "clear everything now" pass, not a substitute for the smaller on-interval Skills it composes.
Raises the Glass Box guide one step by fixing its weakest page.
Three raters grade every Glass Box guide page; the run fixes the weakest by consensus and self-merges one PR. Meant to be manual (disable-model-invocation).
Writes one Glass Box Lesson explaining a concept through Terrarium.
Writes a single Glass Box Lesson: one concept, pinned to a revision, with a Terrarium worked example and sourced "why it's built this way" reasoning, then links it into the surrounding Walkthroughs and Lessons. Dispatched per-topic from the topic inventory, not model-invoked (disable-model-invocation: true).
Keeps every Glass Box guide page pinned to what the code does.
The one-way code-to-pages sync for the Glass Box textbook: re-pins every guide page (drafts opt-in) and term to a commit, flags the ones whose prose or diagrams may now be stale, and lands a self-merging PR once they're re-checked. Essential to keeping Glass Box honest as the Platform changes.
Builds an invited visitor's idea into a real, mergeable change.
The build half of the guest-demo pipeline (ADR-0023): watches for ready-for-agent and ships each guest story as a gated, owner-merged PR, composing frictions-to-fixes's dispatch-and-review machinery. Essential while the pipeline runs, only in the owner's bounded demo windows, so rarely used.
Interviews a visitor to turn their idea into a buildable spec.
The intake half of the guest-demo pipeline (ADR-0023): runs a bounded requirements-refinement session with each guest in GitHub comments, borrowing grilling and domain-modeling's voice, then hands off to guest-build via ready-for-agent. Essential while the pipeline runs, only in the owner's bounded demo windows, so rarely used.
Answers what another Claude session did, read or failed, at the lowest cost to the asking session's own context.
The model-facing front door for reading another session's events (list_events/get_event): answers from the session's log and cheap reads first, then dispatches a subagent to search the raw events so they never enter the asking session's context. Essential whenever a session must inspect a sibling's history, rare otherwise.
Catalogues a Midden survey's finds as narrated Artifacts.
The curator half of the Midden pipeline that follows midden-survey: turns a survey-report issue's candidates into catalogued Artifacts and dig reports, grading and fact-checking each against primary sources, then opens a PR for a human to review (layers/midden/GLOSSARY.md).
Surveys recently discarded code and flags what's worth keeping.
The Midden's field-walk: sweeps for freshly discarded work (deleted files and dropped dependencies via its own script; closed unmerged PRs and dormant branches via the GitHub tools), screens it against the two-gate inclusion bar, and files one survey-report issue for midden-catalogue. Discovery only; that Skill grades.
Adds one new invention Campaign to the Tinkerfund shop.
Adds one Campaign to Tinkerfund's prod Space: page, crowd numbers, Inventor, Updates, comments, and Fable-drawn figures (preview and apply toolkit). The single home for what the schema can't check: the deadpan voice, consistent crowd numbers, the figure style and refinement loop, and the e2e Home assertions a new Campaign disturbs. Essential whenever the catalog grows.
Number after a name: sessions that used it. No number: not used in this Space.
General-engineering Skills installed from an outside pack: used here, not evolved here. Marginal ones are left out.
Checks a change against repo rules and the original ask before landing.
Review the working diff (Standards+Spec) before every code-landing gated PR (ADR-0003/0004); it routinely catches bugs self-review missed. A spec-reviewer subagent has cited a fabricated line, so restate CLAUDE.md's "verify before you state" rule in each review brief and re-verify every cited line and finding with a fresh Read. It reviews a diff against a fixed point; for a whole tree, treat the tree as the diff or run per directory.
Keeps the project's shared vocabulary and decisions consistent.
Owns the glossary and the ADR conventions. Use before any structural change or new vocabulary.
Stress-tests a plan against the project's docs and vocabulary before building.
Stress-test a plan or design against the relevant documentation before building — useful ahead of a milestone or a structural ADR.
Shared vocabulary for shaping a well-organized code module.
Shared vocabulary for designing deep modules — seams, interface depth, testability. Narrow but valuable when shaping a module's boundary (e.g. shared/expand.ts, routing); other Skills borrow its terms.
Finds and fixes structural weak spots in the code's design.
Scans the codebase for deepening opportunities, reports them visually, then grills the chosen one. A specialist entry point for deliberate architecture sessions — infrequent, but the right tool when one occurs.
Steps through resolving an in-progress git merge or rebase conflict.
The Platform's only guided path through a merge or rebase that actually conflicts: rare, but it spares manual conflict-marker surgery.
Maps out a big investigation into trackable, shareable steps.
Charts a shared map of investigation tickets for work too big to fit in one session. Reach for it when a destination is clear but the route to it isn't; not a fit for ordinary single-session work.
Generates an interactive bash wizard for steps only a human can run.
The hand-off when an agent hits a step it genuinely cannot perform (a workflow-file edit, a credential, a third-party dashboard, a one-off cutover), instead of working around the restriction. It produces a script the human runs; the agent does every surrounding step (prose edits, PR, review). Essential whenever a guard rightly blocks an agent action and a human must finish it.
Number after a name: sessions that used it. No number: not used in this Space.