Creates Architecture Decision Records (ADRs) in MADR format and arc42 documentation. Generates plan-context.md as the context bridge to Claude Code. Use this skill when the user mentions "architecture", "ADR", "arc42", "Architecture Decision", "tech stack", "solution design", "system design", "architecture review", "plan-context", or similar. Also when requirements exist and the next step is technical structuring. This skill creates PROPOSALS. Claude Code makes the final decisions based on the real state of the codebase.

Author an engineering-change Product Requirements Document (PRD) that requires inspecting a code repository — verifying the requester's claims against the actual codebase, interviewing them down the design tree, deciding module and interface boundaries, then filing the result as a GitHub or GitLab issue. Use when the deliverable is a repo-grounded implementation spec for engineers. Trigger on "spec this feature out", "turn this idea into an issue", "engineering PRD", "write this change up before we build it", "plan this against the codebase", or a rough feature idea that must become an actionable engineering spec. Runs five phases (INTAKE, VERIFY, INTERVIEW, MODULES, WRITE + FILE) gated on a decision ledger. Boundary — for PM-facing product requirements aimed at stakeholders (Lean PRD, one-pager, AARRR or HEART metrics) use `prd-generator` instead.

Produce a rigorous, verified, fully-cited research report on ANY topic by running a disciplined five-stage pipeline: CLARIFY the question, PLAN (an editable research plan), FAN-OUT parallel search subagents, VERIFY every key claim (source-grounded + adversarial + cross-model), then SYNTHESIZE a cited report. Use this WHENEVER the user wants real research rather than a quick answer — "research X", "do a deep dive on Y", "write me a report / brief / literature review on Z", "compare A vs B and back it with sources", "what's the state of the art in …", "investigate …", "find evidence for/against …", "give me a market/landscape/competitor analysis", or any question where being wrong is costly and the answer needs citations. Trigger even when the user does not say the word "research" but clearly needs multi-source, fact-checked, cited output. This is the general-topic web-research orchestrator — not the codebase-to-wiki `deep-research` skill. For a fast single-fact lookup, answer directly instead.

Deep, read-only code comprehension mode — understand how a codebase actually works by tracing real execution before any design, decision, or change. Produces a fixed deliverable set (high-level execution flow, call graph, call hierarchy tree, per-method dependency table, data flow, external interactions, error propagation, important methods, hotspots, and safe extension points), then STOPS and waits for approval. Use whenever the user wants to understand, map, trace, or explain existing code rather than modify it — "how does X work", "trace the execution of", "map the call graph", "walk me through this feature", "what calls this method", "where are the entry points", "how does data flow", "find the hotspots", "where can I safely extend this", or "understand this codebase before I touch it". Also triggers on "code understanding mode" and "analyze the actual code execution". Read-only by design — it never proposes implementation or refactoring, only understanding.

Use when the user wants to hold work to a higher engineering bar, not a fast agreeable answer. Four modes. (1) ADVERSARIAL REVIEW — "grill me on these changes", "be my reviewer", "poke holes", "don't open a PR until I pass your test" — become a skeptical gatekeeper, not a rubber stamp. (2) PROVE-BY-DIFF — "prove to me this works", "show me it works", "diff behavior between main and my branch" — demonstrate the change empirically by exercising both branches, don't just assert it. (3) ELEGANT REWRITE — "scrap this and implement the elegant solution", "redo this properly", "stop patching, start over" — throw away the mediocre patch and rebuild from a clean design. (4) SPEC-FIRST — "spec this out first", "reduce the ambiguity before you start", or any vague hand-off — converge on a precise, unambiguous spec before touching code. Trigger whenever the user is deliberately raising the standard on a code change — challenging, proving, redoing, or de-risking — even if they don't name a mode.

Use when validating an upstream pull request or dependency fix against VS Code tests using GitHub Actions CI build artifacts instead of building locally. Covers finding compatible binaries, verifying PR and merge-commit provenance, selecting a candidate runtime or build, comparing unchanged regression tests, and restoring the baseline.

Automatically fix Azure SDK CI/CD pipeline failures by applying code changes and verifying locally. USE FOR: "fix pipeline failure", "fix CI", "fix failing tests", "auto-fix and commit the fix", "fix build error", "fix mypy/pylint/type-check/lint errors", "auto-fix pipeline", "resolve pipeline failure". DO NOT USE FOR: pipeline analysis (instead use azsdk-common-pipeline-analysis), API design review, SDK publishing. INVOKES: azure-sdk-mcp:azsdk_package_build_code, azure-sdk-mcp:azsdk_package_run_check, azure-sdk-mcp:azsdk_package_run_tests, azure-sdk-mcp:azsdk_verify_setup.

Use when the user asks to turn off, disable, close, or disarm luffy-arm full-power mode, including "turn off full power", "disable full-power", 关闭 full power, 退出全功率模式. Removes the dedicated admin key, verifies that credential is off, and reports alternate user-level SSH access separately.

Diagnose a named CCB agent by combining authoritative runtime and job lineage with deep read-only pane inspection, apply bounded recovery when evidence supports it, verify the result, and request authorization before submitting a redacted GitHub issue. Use for `$ccb_diagnose agentname`, `$ccb-diagnose agentname`, or reports that a CCB agent is stuck, disconnected, not continuing, not replying, or showing provider errors.

Use the `ollygarden` CLI to inspect telemetry services and insights, Rose repositories, code findings, scan executions, analytics, organizations, auth contexts, and webhooks. Use for: current or recent Rose findings for this repo, important findings across my organization, repository scan status, connected repositories, service insights, OllyGarden auth/context, webhook inspection, and JSON/jq scripting.

Drive a researched and planned feature to a landed, tested change. Use after plan-feature has produced a docs/design/<project>/ workspace and the user says "implement it", "build the plan", "run the plan", or "let's ship this". Orchestrates refresh-plan, implement, review, a debug-local-deployment loop, and a test loop across the daytona / local-pi / claude x SDK / UI matrix, then documentation and a GitButler stacked branch. The orchestrator stays in the loop and spins narrow subagents for each phase.

Turn scanned faxes, images, and CSV/CDA exports into clean text ready for OpenMed de-identification and NER, fully on-device. Use when the user has clinical documents (image scans, photographed/faxed notes, tabular CSV/TSV exports, C-CDA XML) and needs OCR or structured intake before openmed.deidentify and openmed.analyze_text, asks about openmed.multimodal, OCR engines (Tesseract / PaddleOCR), tabular redaction, or layout and reading order. Covers the verified ocr() and redact_document() entry points and the ExtractedDocument contract. Pairs before deidentifying-clinical-text and extracting-clinical-entities.

Evaluate an OpenMed de-identification or clinical NER model against the leakage-first release gates G1a through G8, which gate releases on residual PHI leakage rather than on F1. Use when the user wants to run the OpenMed eval harness on a synthetic golden set, decide whether a de-id model is RELEASABLE or QUARANTINED, enforce direct-identifier recall floors, require zero critical leakage, fit calibration thresholds, or produce a signed gate report. Trigger on "release gate", "leakage", "is this model safe to ship", "G1a", "G3", "quarantine", "recall floor", or "calibration thresholds" in an OpenMed de-id context.

Remove, mask, or replace PHI/PII in clinical free text on-device with OpenMed's deidentify(). Use when the user needs to de-identify medical notes, strip patient identifiers, redact PHI before sharing or analysis, anonymize discharge summaries, or pick a de-id method (mask vs remove vs replace vs hash vs shift_dates). Covers confidence_threshold for safety, consistent+seed for stable surrogates, keep_mapping for reversible de-id, policy= profiles, and the DeidentificationResult fields. Pairs with OpenMed extract_pii (detect spans), reidentify (restore), configuring-privacy-policies, and auditing-deidentification-runs.

Run large-scale batch NER, PII extraction, or de-identification over many clinical notes on-device with OpenMed, with sharding, checkpointing, resumability, and append-only JSONL output. Use when the user needs to process a corpus or folder of notes, de-identify a dataset, run NER over thousands of documents, build a resumable batch pipeline, or stream results to JSONL without holding everything in memory. Covers process_batch / BatchProcessor / BatchItem / BatchResult, the operation= selector (analyze_text | extract_pii | deidentify), iter_process streaming, the PHI-safe on_progress callback, chunking long documents, and no-PHI logging. Produces a resumable batch runner over an OpenMed model.

Generates and verifies 21 CFR Part 11-style audit trails — who/what/when, electronic signatures, and tamper-evidence — for OpenMed pipelines in GxP and clinical-trial (GCP) settings. Use when the user runs OpenMed in a regulated/validated environment and needs an attributable, time-stamped, tamper-evident record of each processing action, electronic-signature manifestations, or computer-system-validation (CSV) evidence. Trigger keywords: 21 CFR Part 11, Part 11, audit trail, electronic signature, e-signature, GxP, GCP, GLP, GMP, CSV, computer system validation, data integrity, ALCOA, tamper-evident, contemporaneous. Pairs adjacent to OpenMed: maps directly onto OpenMed deidentify(audit=True) -> signed AuditReport with .sign(key)/.verify(key), whose repro_hash + HMAC give the tamper-evidence and attribution Part 11 expects. This is a compliance-enablement aid, not a validation certification.

Animate SVGs with SMIL, CSS keyframes, stroke drawing, shape morphing, and motion paths. Use when adding motion to SVGs such as spinners, logos, draw-in effects, or morphing icons, not for static drawing or icon layout.

Hono patterns for TypeScript API routes, middleware, request and response typing, streaming, WebSockets, and Cloudflare Workers deployment. Use when users mention Hono, honojs, Cloudflare Worker handlers, Hono middleware, or Hono route typing.

Diagnose hard bugs and performance regressions with a reproduce-to-regression-test feedback loop. Use when the user reports broken, throwing, failing, or slow behavior or explicitly asks to diagnose or debug it.