Drop in one talking-head video and get it finished — transcribe it, search the web for supporting b-roll (article screenshots, memes, images, video clips), present the asset list for approval, then hand off to reel-builder to normalize audio, burn subtitles, cut the b-roll in, and export for Instagram and LinkedIn. Use when the user drops a video and wants illustrations/screenshots/memes added, asks to "edit this like CapCut", wants b-roll found for them, or wants a talking-head take prepped for Instagram/LinkedIn. Sourcing and approval only — all rendering belongs to the reel-builder skill.

Triage US residential real-property assessment review or appeal requests and prepare evidence packets when market value is a material ground. From an address and notice value, separates market, appraised, assessed, equalized, and taxable values; verifies jurisdiction-specific rules and valuation dates; analyzes comparable sales with APNs, stable transaction IDs, and contrary evidence; estimates a supportable requested value; and produces upload-ready English Markdown and PDF. Use for market-value informal reviews, decline-in-value or Proposition 8 requests, protests, grievances, abatements, comparable-sales packages, and defensible reduced-value estimates. Also use to identify the current authority, form, and deadline for unequal or non-uniform appraisal, exemption, classification, ownership, special valuation, or another non-market ground, but stop before comparable-sale valuation or packet generation for those grounds.

Compute perceptually-correct confidence from DJL face_feature cosine distances using piecewise mapping (d ≤ 0.30 → 1.0, d ≥ 0.65 → 0.0, linear between). Includes enrollment averaging, L2 normalization, and the "textbook formula compresses strong matches" anti-pattern. Use when mapping FaceNet/ArcFace cosine distance to a user-facing confidence score, driving a confidence display (semaphore, progress bar, gauge), or diagnosing why a strong-looking recognition still reads as "yellow" or "weak" downstream.

Load whenever the task is to debug, troubleshoot, diagnose, investigate, or root-cause a failure — a bug, error, crash, failing or flaky test, regression, or any "it doesn't work" report — in any language or codebase. Load BEFORE reading the suspect code or proposing a fix, including when the fix looks obvious or the user has already named a suspected cause. Not needed for tests that are expected-red mid-TDD, or for explain-only questions where no fix is requested.

Finds source-visible semantic runtime bugs without running tests or builds. Use for repository bug hunts from an agent, Codex automation, or Claude loop that should report actionable correctness, security, dataflow, state/lifecycle, API-contract, validation, cache/persistence, or error-path bugs, not style, lint, architecture, maintainability, or missing-test feedback.

Use whenever a Flutter app needs tests: unit tests with flutter test, widget tests with testWidgets and WidgetTester, integration_test on simulators, emulators or devices, choosing the layer per change, mockito with build_runner, LCOV coverage and CI.

Use whenever code that reads feature flags through OpenFeature, Unleash, Flagsmith or GrowthBook needs tests: InMemoryProvider for hermetic tests, evaluation details, hooks and targeting context in TypeScript, Java and Python, plus a flag test-matrix workflow. LaunchDarkly has launchdarkly-testing.

Use when a pull request or release changes an OpenAPI spec and breaking changes must be caught: removed endpoints, deleted schemas, narrowed enums, changed responses. Runs oasdiff breaking, grades findings ERR, WARN or INFO and gates CI with --fail-on.

Use when two concurrent database transactions can touch the same rows (balance debits, seat bookings, stock counts), or before changing a default isolation level. Tests which anomalies each level prevents (dirty, non-repeatable and phantom reads, write skew) in PostgreSQL, MySQL and SQL Server.

Use when a project has accessibility debt and a zero-violations gate would block every PR. Builds a CI ratchet that fails only on new WCAG violations a PR introduces, with pre-existing ones baselined per rule and page, from axe-core, pa11y, Lighthouse, WAVE or IBM Equal Access results.

Operate the Graphite Merge Queue — enqueue PRs via label, dequeue, monitor queue status, and reason about stacked-PR queueing. Use when the user mentions Graphite merge queue, "merge when ready", queueing a PR, dequeueing, fast-track, the merge-queue label, "enqueue", "the queue", or when they ask to merge a PR/stack through Graphite (not directly). Triggers on phrases like "queue this PR", "add to merge queue", "enqueue stack", "remove from queue", "why didn't my PR merge", and on edits to `.please/config.yml` under `graphite.merge-queue`.

Configure a GitHub repository for Graphite — branch protection, merge queue, CI triggers, and stack-aware CI optimizations. Use when the user is onboarding a repo to Graphite, mentions Graphite branch protection / required checks / signed commits / merge queue, sees CI failing on `graphite-base/*` branches, asks about merge-queue choice (Graphite vs GitHub vs external), or wants to reduce CI cost for stacked PRs (CI Optimizations, stack-aware CI, `needs.optimize_ci.outputs.skip`). Triggers on `.github/workflows/*.yml` changes for repos using Graphite, the phrases "stack CI", "stacking and CI", "graphite-base", and any branch-protection or merge-queue discussion in a Graphite repo.

Drive the Graphite CLI (`gt`) for stacked pull-request workflows. Use when the user works with stacked PRs, mentions Graphite, gt, "the stack", or wants to create/submit/sync/restack/split/squash/fold branches. Triggers on `gt` commands, "stack of PRs", "stacked diffs", "trunk-based", `.graphite_repo_config`, branches prefixed by a Graphite user (e.g. `lms--`, `pp--`).

Triage an academic paper into a structured reference summary for the agentic-context research repo. Use when given an arxiv ID, arxiv URL, DOI, or paper PDF to assess context management relevance. Creates references/{slug}.md; adds REVIEWED.md entry and REFERENCE_INDEX.md row. Triggers: triage paper, add paper, analyse paper, arxiv, context window, prompt compression, token budget, token reduction, context injection, context summarisation, RAG injection, eviction policy, long context, LLM context, context overflow, tiered context, context pruning, sliding window, KV cache, attention sink, survey paper, benchmark paper, deep dive paper.

Triage a context management tool or library into a structured reference summary for the agentic-context research repo. Use when given a GitHub URL, npm package, PyPI package, CLI tool, MCP server, or library name to assess context management relevance. Creates references/{slug}.md with architecture overview and scope assessment; adds REVIEWED.md entry and REFERENCE_INDEX.md row. Triggers: triage tool, add tool, analyse tool, context window tool, prompt compression library, token budgeting cli, context injection framework, session lifecycle tool, MCP context server, context daemon, context sandbox, context vault, rolling summary tool, context compressor, eviction policy tool, context pruning library, token counter, sliding window manager.

Produce Metis Strategy whitepapers — thought leadership PDFs with a specific structure, visual style, and editorial voice. Use this skill whenever a user asks to create, draft, or structure a whitepaper, thought leadership paper, playbook, or long-form research report for Metis Strategy. Also trigger when the user mentions "whitepaper," "white paper," "playbook," "thought leadership piece," "long-form report," or asks to produce a publishable document that looks like existing Metis research. Output is a polished, print-ready PDF generated from HTML via Playwright — matching the look and feel of published Metis whitepapers. Do NOT use this skill for short reports, proposals, one-pagers, or presentations — use metis-brand-identity for those. Always use this skill for whitepaper requests, even if the user just says "write a whitepaper on X" without further specification.