Recently added

Market sizing skill that estimates TAM/SAM/SOM using top-down and bottom-up approaches with evidence-backed reports.

Turn "fix it until it's OK" into a stopping condition a machine can decide — for DOCUMENT verification loops. Put the exit criterion into a deterministic detector's exit code, gate on a SET of detectors (never one), keep a monotonicity floor so deletion can't satisfy the gate, and enforce the generator/grader boundary mechanically rather than by instruction. Ships 11 dependency-free detectors (63-120 lines, meant to be edited), a template that wires the conventions into code, a test runner, and a fixed document format that collapses the per-document setup to a single yes/no. Deliberately narrow, and measured: the gate certifies TRACEABILITY, not correctness — in trials, a version that added five verbatim-checked quotes and a version that added none both passed the same gate. Says so in every output. Scope is prose; code has off-the-shelf checkers and different economics. Trials also surfaced that the detectors' own edit points become escape hatches — a skip-mark, a section name on the skip list, a zero declaration — and that under a tight turn limit the model stops iterating and reads the verifier source to solve it analytically. Both are documented rather than papered over: what cannot be closed is stated in the output. Pre-1.0.

Guardrails and a minimal working loop for co-writing a LONG document with an AI — tens of thousands of characters to book length, including academic long-form. Long documents fail by drifting rather than by being written badly, so this skill targets the three measured decay axes (long input, long output, long session) with rules that hold across whatever other skill is running: never single-pass the whole document, carry a spine file (purpose, outline, glossary, style contract, claims) outside the context window, write sections sequentially with the previous section's tail loaded, repair the seams that decomposition provably creates, keep the outline revisable, and gate on a deterministic scan plus a specific human review list rather than AI self-scoring. Every rule traces to a cited finding in references/failure-modes.md, including the disagreements. Ships a stdlib-only drift_scan.py that reads the spine and enforces it, now including an INTENT ledger — the exit condition written before the work, so a silent change of aim becomes visible instead of being noticed three sections later — and CONTENT checks that ask whether the draft does what the spine said it would (claim coverage, unverified claims, terms used before their definition, unsourced assertions), plus references/content-review.md — the per-section procedure that produces the human review list for what needs meaning. Surface checks: declared style-contract and glossary violations, cross-section contradiction of typed facts (numbers, dates, versions, 可否) plus a ledger-driven candidate list that narrows what a human must read side by side, 文体 mixing, over-long sentences, notation drift, cross-section duplication, and section-length collapse — with exit codes suitable as a loop's stop condition, plus an optional SudachiPy backend that lifts terminology-drift recall from 0/8 to 6/8 on unlisted variants. Bilingual JA/EN, with Japanese long-form rules sourced from 文化審議会建議「公用文作成の考え方」and the JTF style guide.

Literature search, citation chasing, snowballing, author analysis, and BibTeX export using free official APIs (Semantic Scholar + OpenAlex + CrossRef). Honest Google Scholar alternative — no scraping, no paid services.

A minimal, portable, cross-agent persistent knowledge base. Stores knowledge as plain Markdown — one KB is one directory that works by copying the folder: no server, no database, syncable via Git or OneDrive. Supports multiple KBs, each declaring its own format (prose / entities, with room for more), and typed relations between records that are written once and traversed in both directions. Cross-agent adapters verified on Claude Code, Codex, and Antigravity (agy); bundles a stdlib-only validator and graph tool.

Rewrite AI-generated or AI-assisted text to read as human-authored. Reduces common AI tells (em-dash overuse, 'delve'/'leverage', tricolons, reflex bullet lists, 'いかがでしたでしょうか', generic closers) while preserving meaning, register, and factual content. Bilingual: English and Japanese. Reader-experience focused, not detector evasion.

Runs multi-source research with hallucination control built into the architecture — parallel retrieval subagents that return only quoted spans, an append-only Source Ledger where no claim ships without a source row, blind per-claim verification, and disagreement surfaced rather than averaged.

Lists GitHub Issues, aggregates by label, analyzes priorities, and organizes tasks.

Faithful translation skill that produces source-faithful translations across any language pair, with a parallel sentence ledger for verification, terminology preservation, and translator's notes.

Automated fact-checking skill with structured checklist mode. Extracts verifiable claims and citations from documents (Markdown, PDF, DOCX), verifies them via web search and Puppeteer headless browser across 7 verification dimensions (link validity, content matching, citation granularity, numeric accuracy, fact/speculation distinction, source reliability, cross-section consistency), and generates a comprehensive Markdown fact-check report.

Verifies the validity of references and citations in reports and papers, conducting evidence-based fact-checking.

Design event-driven, asynchronous, and real-time services beyond request/response REST: messaging (queues/streams), async workers, scheduled/round-based jobs, and WebSocket/SSE push, with correctness under concurrency — idempotency, ordering, delivery guarantees, and backpressure.

Find the essence of an artifact and strip away what is inessential — the via-negativa counterpart to doc-refactor, across prose/docs, specs & requirements, source code, and diagrams. Separates what is load-bearing for the artifact's ONE purpose and audience from redundancy, gold-plating, decoration, filler, speculative generality, premature detail, and scope creep, then PROPOSES cuts with rationale and a 'what would be lost' note. Deliberately meaning-changing (unlike doc-refactor, which only rearranges): diagnose-first, non-destructive, propose-then-confirm. Anchors every cut to a stated purpose + audience, respects Chesterton's Fence (flag, don't cut, what you don't understand), and preserves correctness/behavior. Hands off the code edit to /simplify or /code-review, the redraw to document-figures, a separate short derivative to document-summary, pure rearrange to doc-refactor. Bilingual JA/EN.

Thin conductor for end-to-end EA-driven delivery: sequences requirements → design → middleware selection → implementation → service across the specialist skills, carries ea-model.yaml as the through-line, enforces a gate before each handoff, and tracks progress plus back-edges in a delivery ledger. Delegates all phase work; holds no phase logic itself.

Document and literature summarization skill with Executive and Professional modes, mandatory source-grounded Claim Ledger, and explicit Source vs Inference separation to prevent hallucinated content.

Extract figures from existing documents (PDF, Word, PowerPoint, web) with provenance and create new structural diagrams (Mermaid-first), producing a Figure Ledger that chains into document-summary.

Critical, read-only review of a prose document — the prose counterpart of /code-review. Hunts substantive weaknesses (unsound argument, buried conclusion, thin evidence, unstated assumptions, missing counterargument, audience mismatch, internal contradiction) across a genre-weighted rubric, adversarially verifies its own findings to kill sycophancy and nitpicks, and returns a prioritized findings report the author decides on — WITHOUT editing the text. Hands off rewriting to doc-refactor, fact-checking to verify-content, restyling to ai-tell-reducer. Bilingual JA/EN.

Refactor prose documents — restructure and de-duplicate without changing meaning, the way code refactoring preserves behavior. Runs a diagnose-first workflow (reverse outline → issue inventory → confirm structural moves → refactor → change log) that preserves every claim, fact, figure, and the author's voice, and flags substantive problems instead of silently fixing them.

Guided, knowledge-grounded co-authoring workflow for substantial documents (proposals, specs, decision docs, technical articles, whitepapers). Three optional, composable stages — context gathering, section-by-section drafting, and reader-testing — grounding facts in a domain knowledge base with cited sources and consistent terminology. Drafts into a working file, preserves human [人] edits, and hands off critical review to doc-review, restructuring to doc-refactor, fact-checking to verify-content/evidence-check, de-AI restyle to ai-tell-reducer/humanize-prose, and translation to faithful-translation. Bilingual JA/EN.

Design knowledge restructured so an AI can look it up without guessing. Use-case playbooks (decision deck / technical explanation / pitch / refactoring an existing deck / tables / charts / structural diagrams / corporate-template compliance) that route to atomic rules, each carrying a tier (MUST / SHOULD / HOUSE), a cited source, and a completion condition written in a detector-compatible form. Decided values ship as W3C DTCG tokens; named anti-patterns say what not to do. Router only — it holds no judgement itself and hands file generation to pptx-build.

Turns raw records of an international academic conference, a commercial/vendor conference or trade show, a standards meeting, a customer visit, or an internal discussion into a reader-targeted report: a Markdown report as the source of truth plus a derived executive .pptx. Fixes reader / ask / disclosure scope / original objective / event profile first, inventories every input (AI transcripts, photos, captured slides, proceedings, handouts, colleagues' notes) with reliability and confidentiality, restructures by issue rather than by session, and separates fact / interpretation / hearsay with a lightweight evidence ledger. Requires a coverage statement for multi-track events, counted evidence for trend claims, and a GA/Preview/roadmap classification for vendor announcements. Ships per-profile report and deck templates.

Comprehensive data analysis expert covering statistical insights, visualization, and machine learning.

Competitive analysis skill that researches competitor products via web search, compares features/pricing/differentiation, and generates comparison matrices.

Guides a safe, selective commit and — when you say yes to the one question it asks up front — carries the change through to release: version bump, CHANGELOG, annotated tag, push, GitHub Release, and package publish, with an explicit approval gate at every outward-facing, irreversible step. Reviews changes before committing, suggests file-specific commits for large changesets, and prevents blanket 'git add -A'. Bundles a self-contained commit helper that blocks AI-signed commit messages and a read-only release preflight that checks tree state, upstream sync, tag collisions, version files, CHANGELOG, publish targets, and gh auth.

Commits changes, pushes to remote, and reports progress to GitHub Issues in one step.

Expert code reviewer providing constructive feedback on quality, security, performance, and maintainability.

CLI development expert for designing and implementing user-friendly command-line tools.

Checkpoint management for AI instruction systems. Tracks task start, progress, and completion, and manages instruction usage state.

Business model design skill that generates structured documents following Business Model Canvas and Lean Canvas frameworks.

Comprehensive business consulting from strategy to execution using McKinsey/BCG methodologies.

Detects and runs the appropriate build command for your project. Supports Node.js, Rust, Python, Go, Makefile, and more.

A lightweight, persistent do-it-later inbox. Park a task in one line and pick it up in a later session — hybrid storage (repo BACKLOG.md when inside a git repo, else ~/.claude/backlog.md), a merged list that shows repo + personal at once, and archive/grooming to clear out done items. Invocable as /backlog and on natural phrases like 「これ積んでおいて」 / 「add to backlog」. stdlib-only, cross-agent.

Bridge an ArchiMate EA model's Application layer to implementation: derive a requirement→component→service→api→task traceability matrix, OpenAPI skeletons per ApplicationService, DataObject entity stubs, and a per-component task list, then check for orphans (unimplemented requirements, untraced components).

Read, review, interpret, and surgically edit existing Archi-native .archimate model files in place — round-trip-safe (preserves ids, geometry, folders, styles). Validates against the ArchiMate 3.2 metamodel via a version-aware native↔canonical vocabulary bridge, and generates Markdown reports and PlantUML views for quick review without opening Archi. Brownfield counterpart to archimate-ea.

Conversational ArchiMate 3.2 enterprise-architecture facilitator. Elicits requirements and designs the full layered model with the user into one YAML running model (single source of truth), validates it against the metamodel, then generate-forwards PlantUML views and Open Group Exchange XML for the Archi tool.

Reduce the "AI-ness" of writing — uniform sentence rhythm, reflexive hedging, vague abstraction, formulaic scaffolding, inflated vocabulary, and surface tics (em-dash / bold / rule-of-three overuse) — while preserving meaning, facts, register, and the author's voice, and without fabricating anything. Works on Japanese and English text.

Delegate a task to another CLI coding agent (Codex, Antigravity/agy, or Claude Code) from within your current agent. Explicit-invocation only, read-first routing, a preview→confirm→apply gate for any write, and an OS sandbox (bubblewrap) confining every write step.

Comprehensive academic research support including paper writing, literature reviews, citation management, and research methodology.

Deterministic orchestrator for CLI coding agents. Spawn parallel agents in per-task git worktrees, verify their work, and merge results; no model in the coordination loop, so runs replay byte-identically.

A senior backend and data engineering colleague. Designs systems in the order the work runs (purpose, process, architecture, data, expression, verification), produces a design dossier sized by a scored intake, and holds the result to mechanical gates where absent evidence is NO-DATA and never a pass.

Capture session knowledge into structured entries and consolidate them into durable guidance, author paste-ready hand-off briefs, convene fresh-eyes review panels, audit a large file corpus by blind fan-out, maintain a compaction-surviving control anchor for long autonomous runs, behaviorally evaluate skills, keep live awareness of the installed toolkit, sign agent-assisted work with machine-generated provenance trailers (exact model + versioned tool stack, never a co-author line), and run the tool-dogfooding feedback loop: per-session feedback reports captured into each tool's repo and triaged into leverage-ordered improvement backlogs.

Superpowers-derived process disciplines in a calibrated register: fit-ranked tool dispatch, capacity dispatch (task-to-model/effort routing on a calibrated rubric), calibration-first skill authoring, red-green TDD, root-cause-first debugging, design-before-code brainstorming, evidence-before-claims verification, technically rigorous review reception, and midweight planned execution (executable-cold plans + per-task subagent review loop) — deduplicated against the craft-collection toolkit and eval-gated from birth.

Discipline for evaluation acts: structure an experiment and its write-up so the Methods reconstruct without the conversation and the uncertainty is declared, not hidden — a typed record.yaml across a probe / measurement / decision tier ladder, with every load-bearing rule a script that exits non-zero rather than a line of prose.

Modern Python engineering standards and stack-agnostic data-engineering discipline, with ruff/uv enforcement hooks and a self-refreshing toolchain manifest.

Create portable persona folders for Claude Code and Codex

Reusable skills, commands, agent profiles, hooks, and platform adapters for AI-assisted creative work.

Claude Code efficiency and hygiene toolkit. Surfaces token spend across the 5h / 7d / Sonnet-7d billing windows with per-model / per-agent / per-skill attribution, and audits your effective Claude Code configuration (custom + plugin agents and skills) for overlap and conflicts. Ships an interactive dashboard, a conversational usage-analysis skill, and a structured config-audit skill.

Extract subtitles/transcripts from YouTube videos via CLI or browser automation

Generate or edit images using Google Gemini API via nanobanana

Leverage OpenAI Codex for non-interactive automation mode and hands-off task execution