Claude Skills by oaknational

Use this skill when the question, construct, boundary, unit of analysis, causal story, stakeholder perspective, scale, or conceptual decomposition may be wrong or contested. It produces protected alternative Frame Cards, a multidimensional scale map, explicit alternatives, Bridge Claims, and Crosswalk Claims. Invoke directly for requests to reframe a problem, expose hidden assumptions, compare decompositions, or test whether a team is solving the wrong problem. Do not use merely to plan evide...

Use this skill when outcomes, completed inquiries, experiment results, incident recurrences, evaluation traces, or several related cases are available and the goal is to learn about object conclusions, methods, skill routing, coordination, or the learning process itself. Invoke for retrospectives, calibration reviews, portfolio learning, proposed skill improvements, and outcome-driven reopening. Do not use it to preserve ordinary session state, to rewrite a skill from one anecdote, or when no...

Designs, audits, analyses, and learns from controlled experiments for digital products and services, including A/B/n and split tests, feature experiments, holdouts, switchbacks, ramps, and quasi-experiments. Invoke for assignment and exposure, sample-ratio mismatch (SRM), metrics and guardrails, minimum detectable effect, power or precision, peeking or sequential inference, interference, novelty, seasonality, heterogeneity, accessibility, harms, rollout, rollback, and long-term product outcom...

Use this skill when two or more existing claims, evidence streams, method reports, disciplines, scales, or stakeholder interpretations must be reconciled without averaging away disagreement. It maps provenance and dependence, distinguishes contradiction from scope difference, tests Bridge and Crosswalk Claims, searches for defeaters, and produces a Conflict Ledger plus multidimensional Epistemic Profile. Invoke directly when evidence already exists and the question is what it jointly warrants...

Use this skill to orchestrate a proportionate, end-to-end inquiry when a consequential task is ambiguous, uncertain, contested, novel, spans multiple scales, or admits materially different frames or methods. It selects screening, core, standard, or deep operation; coordinates framing, inquiry design, protected investigation, synthesis, decision, audit, and world-return learning. Prefer a narrower Parallax sibling for an isolated framing, design, synthesis, decision, audit, or learning task. D...

Author a plan node in the ratified plan-node estate.

Open a pull request and shepherd it to merge-ready: reviewer-facing description, full-surface harvesting (GraphQL review threads, all comments, all checks, Sonar issues), root-cause-first triage, budgeted watching, re-fetch after every push, and an honest truly-green merge — all checks green, every thread resolved, normal non-admin merge. Use whenever a branch reaches PR closeout or an open PR needs driving to live.

Size the work and the instrument before shaping either. A pre-decision gate, sibling to concept-exploration, asking whether this is the right SIZE of question and the right LEVEL to answer it — four findings (too big, too small, wrong instrument weight, wrong level) under a non-override clause that keeps it from becoming an expediency door. Use before the decision lenses; when a loop stops converging; when adjacent findings are being absorbed rather than homed; or when a decision is about to ...

Structured outward reasoning for analysis, planning, decisions, diagnosis, and design — a direct-trial-first gate, a decision-relevant value-of-information stop gate, and five firing moves (name the kind, frame the problem not the solution, surface the warrant, decide for reversibility, stress-test) that point to the full grammar of thinking for depth. The outward pair to oak-metacognition's inward reflection. Use when facing a gnarly problem, choice, or analysis; especially when investigatio...

Run a deep post-mortem on a completed arc — a merged PR series, a finished lane, a resolved incident, a long review series — producing a durable record with a causal stack, a counterfactual test, honest credit, and proposals that each carry a warrant and a falsifier, routed through the PDR-130 lanes. Use at owner word after a significant arc closes, or when an arc's cost or shape surprised everyone and the estate should learn from the trajectory, not just the outcome.

Merge agent memory and state files (napkin, repo-continuity, distilled, thread records, registers) by reconciling CONCEPTS, not lines. Use whenever these files diverge across branches or sessions and a merge, rebase, cherry-pick, or post-update `gh pr merge` would combine them — git line-merges silently corrupt the meaning.

Lightweight end-of-session continuity update with a conditional consolidation gate — the continuity component the wrap programme runs at every session close (owner ruling 2026-07-28). Wrap is the close entry point; this workflow is its step, not an alternative close.

Create a git worktree for a lane and configure it so every downstream surface is true: an inherited bot commit identity checked rather than re-set, dependencies, environment files, and a draft PR at first push. Use when taking up a lane needing its own checkout, or when a worktree misbehaves — commits attributed to nobody, missing env, hook failures. Do not use to switch branches in place (never on the principal), for the session-level residency switch alone (that is EnterWorktree), or to dis...

The Subagent Invocation Framework (Sif): general doctrine for invoking another agent — any vendor, any arity — as bounded in-session capability, plus per-binding annexes carrying each binding's transport, tool-contract, and authority facts. Read before building or using any agent-invoking instrument; route to a concrete instrument (the-codex-dialogues, cricket, codex-helper) for the actual invocation workflow.

Stand up this session as the Watcher for the Practice Slack channel — a named, persistent presence that polls the channel, summarises activity, replies to messages addressed to it, and alerts the owner. Use when asked to become the Slack Watcher, take over or relieve the Watcher mantle, or stand up a watch loop ("become the Watcher", "take over the watch", "relieve <name>"). Do NOT use to send the Watcher a message, ask it a question, or check whether one is live — that is talk-to-slack-watch...

Apply the repository start-right quick grounding workflow to the active session. Use when the user asks to start right, re-ground work, or explicitly apply the shared start-right-quick skill guidance and linked directives before or during task execution.

Apply repository start-right grounding plus team bootstrapping for multi-agent sessions. Use when a coordinated team is starting, re-grounding, or choosing temporary collaboration responsibilities.

Apply the repository start-right-thorough grounding workflow to the active session. Use for high-risk, cross-workspace, architectural, or planning-heavy work where full one-gate-at-a-time discipline is required from the start.

Send a message to the live Slack Watcher from any session and handle the reply correctly. Use when asked to tell the Watcher something, ask it a question, or check whether a Watcher currently holds the mantle ("tell the Watcher X", "ask the Watcher for status", "is the Watcher up?"). Do NOT use to become the Watcher, take over its mantle, or run its polling loop — that is slack-watcher — nor for Slack messages not addressed to the Watcher. Never take the mantle from here: a silent Watcher is ...

From a live CLAUDE seat only: open a bounded multi-turn reflective dialogue with a Codex interlocutor over a direct MCP connection, to probe a stated uncertainty against a different vendor's prior. On a Codex host this instrument does not run (same-vendor dialogue defeats its premise — see the host check). Use mid-task at a genuine fork or uncertainty; not for task delegation (codex-helper), not for a fast one-shot conscience check (cricket), not for live-peer collaboration (a second seat + A...

Author and curate the ticket graph deliberately — scoping, relationships, milestone homes, and the standing curation sweep. Use when minting any ticket, deferring or sequencing work, wiring blockedBy chains, placing milestone homes, or when the graph has grown dense enough that only its authors can navigate it. The graph is authored, not endured.

Write and audit canonical TSDoc plus the adjacent README or ADR updates that belong with a code change.

Exercise UX craft judgment on a screen or artefact — visual hierarchy, layout, spacing, type, and interaction behaviour — deciding what should be loudest, how the eye should travel, and what an element must do when touched. Use when a surface is being composed or critiqued and the question is whether it reads and behaves well, not which class or token to type. Do not use it to pick tokens, classes, or components (design-system-usage owns that), to judge conformance against WCAG success criter...

The repository's orientation lens — one intent-discerning surface for anyone who wants to understand this repo, get started, or get their bearings. Covers "explain this repo", "tell me about this", "what is this", "give me an overview", "executive summary", "how does X work", "I want to understand the search architecture", "onboard me", "where do I start", "give me a tour", "set me up", and "help me contribute". Discerns the person's interest, angle, and the delivery mode that fits — a pinpoi...

Decision tree for safely undoing a working-tree, staged, or committed change. Identifies the git state of the change, names the safe and destructive operations available, and forces an owner-authorisation halt before any destructive operation. Loaded passively and invoked whenever the agent or owner reaches for an undo, revert, reset, or restore.

Align the estate to a changed upstream Oak bulk-download schema — summon on \"bulk schema changed\", \"bulk data fails validation\", \"unrecognized_keys at ingest\", \"refresh the bulk downloads\", or any drift between fresh bulk data and the strict Zod gate. ADR-222's authority ordering is constitutive: upstream's published bulk JSON Schema is the authority; the hand-written Zod templates are the interim mechanism trued against it; a data-vs-schema mismatch is an UPSTREAM BUG REPORT, never v...

Update npm dependencies deliberately — summon on \"dependabot alert\", \"pnpm audit findings\", \"raise a security floor\", \"dep sweep\", \"bump a dependency\", \"pnpm override\", or any advisory/currency/forced-bump trigger. Three entry doors: security-advisory response, routine currency sweep, upstream-forced bump. Core is the mechanism-decision tree (in-range lockfile refresh → package.json bump → override floor, in that preference order; existing floors are RAISED, never twinned) plus a ...

Align the estate to a changed upstream Oak Open Curriculum API (OpenAPI) spec — summon on "upstream spec changed", "schema drift check fired", "refresh the schema cache", "sdk-codegen refresh", a failing correction-layer removal-condition test after an upstream deploy, or an upstream changelog entry. Drives the owning runbooks: characterise the drift first-hand (doc-only vs structural, additive vs consumer-breaking), refresh deliberately, treat correction-layer test failures as lifecycle sign...

Judge whether a rebuilt page matches its visual reference: capture both sides as images at the same canonical width, run the windowed rejection statistics, and read the pair with the heatmap and σ-scores directing attention. Use whenever comparing a rebuild against a design export, a reference render, or a prior capture — fidelity reviews, visual regression triage, and owner-reported "looks different" reports.

Produce and read rendered proof (screenshots, focus-state renders, DOM-fact echoes) for any verdict on visual work — layout, rendering, theming, or interaction behaviour. Use before claiming a visual surface renders correctly, before and after curing a visual defect, after placement or order changes, and before telling the owner a visual deliverable is done. Runs the design-showcase visual probe.

A portable, beginner-friendly primer on working with agentic AI coding agents: what they are, the set-the-goal / it-acts / you-supervise loop, giving good context, reviewing both output and actions, iterating, common failure modes, and working safely. Use as the lead-in for someone new to working with AI coding agents, before any project-specific guidance.

Doctrine for any work that builds, serves, or consumes a graph surface. A graph is not a list: subgraphs are complete within their declared bound or they are wrong; list operations (pagination, truncation, top-N sampling of nodes or edges) never touch a subgraph; responses carry the anchors that navigate to the next bounded response; graph tools are thin deterministic formatters over a smart corpus; a degraded list-shaped answer is never an acceptable substitute for a typed refusal or a well-...

Wrap a session up safely — the deep-closeout PROGRAMME. Orchestrates the modes, work-safety verification with evidence, session-handoff, conditional consolidation, and the deep context-loss scan, then owns the metaloss recursion: repeated passes over the scan itself until the fixed point where a further pass adds no new loss class. Closes EVERY session, ordinary or deep (owner ruling 2026-07-28) — invoked naturally as wrap up safely / we are done here, at any boundary where the seat or sessio...

Recover ChatGPT-, deep-research-, or other LLM-exported reports from paired markdown, DOCX, and PDF copies into source-faithful clean markdown by copying content as faithfully as possible, repairing structure, and using the DOCX to repair real links.

Work the Claude Design pipeline for a converted app — conversion playbook, byte-sacred export refresh via the claude-design MCP, and its core: the export↔implementation fidelity review (serve the canonical export and the dev server, capture both sides at matched geometry, perceptually diff every declared pair, review the side-by-side report, and record a disposition — fix / deliberate / investigate / matched / superseded — for every finding in the tracked divergence register). Use when conver...

Invoke Codex as a sub-agent for well-defined tasks using `codex exec`. Provides templates for brief one-shot tasks (no grounding) and longer repo-aware sessions (with oak-start-right-quick). Use when delegating a self-contained task to Codex from Claude Code or from a shell script.

Create a well-formed commit for current changes with conventional message format. Always active, every commit, every session, no trigger required. Enumerates live commitlint constraints inline at draft time, validates the drafted message via `pnpm agent-tools:check-commit-message` BEFORE invoking git commit, and coordinates the short-lived git index/head commit window.

Choose the right DELIVERY LANE for live agent-to-agent messaging — s2s (SendMessage) for time-critical unblocking between live Claude seats, ARC channel files for rapid dialogue with a named collaborator, the comms event stream for the discovery narrative every present or future seat must find, Slack-via-Watcher for traffic whose audience is the owner or humans on the Practice Slack channel — and hold the behaviours that keep the fast lanes honest: decision-bearing content (Slack-crossing inc...

Structured workflow for merging significantly diverged branches. Use when either branch has changed 100+ files, a dry-run merge produces 10+ conflicts, or the other branch refactored core interfaces your branch consumes.

Explore an unshaped concept, phenomenon, recurring incident class, or messy set of observations before solution options or the decision question are well formed. Use when framing options immediately would foreclose the real question; run four alternating metacognition and reason movements to produce a well-formed understanding with warranted, falsifiable proposals. Do not use it as a separate pre-decision pass once the options or decision question are already well formed; continue through the...

Declare and run session-completion or dedicated-knowledge-curation consolidation, including buffer disposition and closeout proof.

Run a persistent dedicated Oak knowledge-curation goal until every live curation buffer is empty or explicitly owner-decision-gated and its insight is conserved into permanent homes; wraps start-right-quick and consolidate-docs. Fitness is a signal that routes work, never a completion gate or a reason to trim, archive, split, shard, or rename.

Fold the live coordination branch to main and rotate — the full converge-and-rotate ceremony: ownership-aware dirty-file sweep, merge main in with a stale-capture probe, bot fold PR carrying the product-gravity line, full-condition merge, day-stamped successor cut, branch-labelled surface refresh, rotation broadcast, and a wrap-not-closeout loss scan. Invoked at the 24h rule's DUE check, at owner word (\"fold and rotate\", \"converge the coordination branch\"), or before any boundary that nee...

Invoke the platform Cricket panel for a fast second opinion on whether the current work is the right work. Use at cycle or decision boundaries; for materially uncertain or high-impact choices; when the path feels suspiciously obvious; or for rubber-ducking and design partnership. Run normal and adversarial stances, and treat every verdict as evidence rather than authority. Do not use Cricket instead of an artefact reviewer.

Run one curator pass on the repo's knowledge substrate. Use when allocated the curator boundary on team-start, when owner-directed to a curation lane, when a graduation buffer crosses critical fitness, when a pending-graduations trigger fires, or when a landed substrate shows an adoption gap.

Cut or name a coordination branch with the date + base-sha6 name minted by the agent-tools coordination topic. The suffix usually separates cuts from DIFFERENT base tips (successor folds, recovery re-cuts) — a probabilistic lineage signal, never a uniqueness proof — and does not discriminate parallel same-tip cuts. Use at the fold ceremony's successor cut or a recovery re-cut, and never hand-transcribe the name.

Run a full dependency-currency pass — survey with pnpm -r outdated and pnpm audit, triage every bump by measured risk tier, execute one type-affecting major at a time with baseline-capture proof, drive pnpm audit to zero via annotated override floors, and refresh SHA-pinned GitHub Actions against verified stable tags. Use when the owner asks to bring dependencies to latest, clear audit or Dependabot findings, or reopen a dependency-currency lane. Do NOT use for a single dependency bump riding...

Build well-branded, accessible (WCAG 2.2 AA), themable interfaces and assets with the Oak Open Curriculum Design System — production surfaces or throwaway prototypes, mocks, decks, and worksheets. Use whenever composing UI, documents, or teaching artefacts from the system's tokens, component class library, compiled React components, templates, fonts, icons, or brand voice.

Enter a body of material — research, code, transcripts, a day's events — with no question, no target, and no problem statement, and wander: juxtapose, invert, analogise, follow surprise, and see what connections appear. The divergent partner to concept-exploration's convergence. Use when material deserves an unpremeditated encounter, when the owner invokes play, or when no decision or contribution is yet nameable. Returning empty-handed is a valid outcome; seeds that appear are routed onward ...

Run all quality gates and fix issues.