skills

This repository is the source of truth for a small set of reusable agent skills.

The default distribution model is conservative:

• keep each skill under version control in this repository
• avoid public registries and auto-install workflows by default
• link only the required skills into a project workspace when needed

Repository layout

skills/
  <skill-name>/
    SKILL.md
scripts/
  install-skill.sh

Each skill must live at skills/<skill-name>/SKILL.md.

Current skills

Planning flow

• triage: choose the lightest adequate planning route for a development task
• lite-spec: write a compact execution brief for bounded implementation work
• metaplan: review and tighten specs, plans, and task breakdowns before autonomous implementation

Source-of-truth review

• sot-integrity: audit a source-of-truth artifact for authority, evidentiary grounding, trust scope, and conflict with repository reality before implementation or orchestration relies on it

Artifact cleanup

• deslop-history: clean final user-visible artifacts from accepted decisions or issue/PR discussion by removing historical residue and process framing
• deslop-prose: clean academic or professional prose by removing hype, generic metadiscourse, decorative structure, and claim-evidence mismatch while preserving technical meaning

Writing structure

• exposition-flow: turn repository-grounded source material into a self-contained, dependency-ordered document or presentation for an intended audience — audience prerequisites, define-before-use, discourse-role separation, and compression that preserves prerequisite closure

Orchestration

• light-orchestration: choose between single-agent execution and a minimal multi-agent split, and produce strictly bounded subtask contracts when a split is justified

Handoff

• handoff-prompt: generate a compact prompt for the next agent or stage

Notebook review

• marimo-reactive-review: review an existing marimo notebook for reactive design quality and refactor opportunities

Data inspection

• headson: preview large JSON, YAML, JSONL/NDJSON, text, source code, multi-file data, and Jupyter notebooks with structure-aware hson previews under explicit budgets

Computational reproducibility

• computational-reproducibility: keep computational experiments reproducible through explicit preflight checks, manifest expectations, source/artifact separation, fallback policy, and failure semantics

Lean / formalization

• lean-formalization-discipline: keep Lean / Mathlib proofs sound and axiom-clean while refactoring — no sorry on false statements, honest sorry over hidden axioms, self-contained in the repo, and lake build + #print axioms re-verified every commit

Mathematical writing

• math-claim-integrity: audit the structural and logical integrity of a mathematics paper — quantifier scope, proof/computation honesty, theorem hierarchy, contribution-list discipline, stale claims, and standing assumptions
• math-notation-consistency: audit a LaTeX mathematics document for notation drift — definition sites, aliases, symbol reuse across scopes, and back-references after gaps
• math-semantic-preservation: review or rewrite mathematical prose so edits, paraphrases, and terminology migrations preserve the exact source meaning — referent identity, operation vs result, quantifier and equality mode, provenance, and notation role
• wabun-math-style: detect and correct Japanese-language anti-patterns in mathematical writing — epistemic hedges, tense and voice discipline, connective and particle logic, borrowed vocabulary, terminology SoT drift, and concept-preserving translation

Research review

• research-significance: assess whether a mathematical or applied-mathematical research direction is locally significant — separating correctness, novelty, and scholarly value — before prose polishing or source-of-truth treatment

Repository guidance

• growing-agents-md: seed, lint, or update a compact canonical AGENTS.md with deterministic guardrails

Git maintenance

• git-workspace-cleanup: workspace-level Git cleanup — merged branches, merged worktrees, stale remote-tracking refs, and stale worktree metadata across all repos under the workspace root

External skills

• external-skill-review: review a candidate external skill against the repo policy, record approved entries in a project-local catalog, and recommend install commands

Dev containers

• devcontainer-cli: review, validate, and troubleshoot Dev Container setup with explicit CLI-first guidance

Shell scripting

• shell-script-engineering: create, edit, or review non-trivial Bash/sh where correctness depends on execution context, permissions, filesystem mutation, argument parsing, idempotency, or CI/container behaviour

Delivery workflow

• github-driven-workflow: enforce issue-first, PR-gated delivery with no direct main pushes, independent review requirement, and deterministic merge gates
• github-project-board-maintenance: rate-limit-aware GitHub Projects v2 board maintenance — REST-first candidate discovery, single GraphQL snapshot, mechanical status-update planning, and guarded apply

Randomness

• randomness: prefer a bundled PRNG script for random selection; fall back to a String Seed prompt pattern only for low-stakes or creative-diversity use

Coding-agent orchestration

• claude-code-advanced-orchestration: operational guidance for delegating implementation work to Claude Code as a worker behind a coordinator agent (session identity, permission strategy, worktree/tmux patterns, dynamic agents, effort selection, bare-mode auth, MCP trust, remote-control caveats)

Text linting

• textlint-cli: invoke textlint correctly from the CLI in repositories with existing configuration — detection, repo-local invocation, fix mode, diagnostics, CI-oriented formats, and MCP mode

Naming guidance

Prefer short, command-friendly names.

Use names that:

• are easy to invoke directly as slash-style commands
• distinguish categories without excessive prefixes
• describe the skill's role in a few syllables
• avoid avoidable collisions with common built-in commands such as /plan

Tool compatibility

This repository targets multiple coding agents, but each tool discovers reusable guidance differently:

• Claude Code: project-local skills in .claude/skills/
• Codex: project-local skills in .agents/skills/
• GitHub Copilot CLI: project-local skills in .github/skills/ or .claude/skills/
• Gemini CLI: project-local skills in .agents/skills/

Installation model

The installer script links skills from this repository into a target workspace by default. It does not depend on a remote marketplace or registry. It runs with python3 when available and falls back to uv. Use --copy when a workspace needs its own editable copy. Use --force to replace an existing real directory or file. For external skills installed directly from upstream repositories, see docs/external-skills.md.

Supported targets:

• claude: install to .claude/skills/
• codex: install to .agents/skills/
• copilot: install to .github/skills/
• gemini: install to .agents/skills/
• all: install to all supported agent locations

Examples:

./scripts/install-skill.sh all all
./scripts/install-skill.sh triage codex .
./scripts/install-skill.sh lite-spec copilot /path/to/workspace
./scripts/install-skill.sh metaplan gemini /path/to/workspace
./scripts/install-skill.sh handoff-prompt all /path/to/workspace
./scripts/install-skill.sh triage codex /path/to/workspace --copy
./scripts/install-skill.sh triage codex /path/to/workspace --force

APM installation

This repository is compatible with APM (Agent Package Manager).

Install individual skills via subpath:

apm install t-uda/skills/skills/triage
apm install t-uda/skills/skills/metaplan
apm install t-uda/skills/skills/lite-spec

Install with pinned commit:

apm install t-uda/skills/skills/triage#abc1234

APM automatically deploys skills to detected target directories (.github/skills/, .claude/skills/, .agents/skills/, etc.).

The manual installer script remains available for environments where APM is not installed.

Install-first, compile-minimal

This repository favours native install over compiled instruction files:

• apm install is the preferred path. Skills land directly in each tool's native directory.
• apm compile output (AGENTS.md, CLAUDE.md) is treated as a compatibility layer only.
• Compiled artifacts in this repo stay compact: durable policy, minimal guardrails, and pointers. Procedures live in skill files, not in compiled documents.

Current APM scope

Phase 1 covers skills only. Prompts, agents, instructions, hooks, and MCP are not distributed from this repository yet. See docs/apm-consumption.md for supported primitives, target-directory mapping per tool, and subpath install patterns.

Project instruction files

These are separate from skills, but they affect project-local behaviour:

• GitHub Copilot supports AGENTS.md, GEMINI.md, and .github/copilot-instructions.md
• Gemini CLI loads hierarchical GEMINI.md files and can be configured to accept other context filenames

The installer in this repository manages skills only. It does not generate repository-wide instruction files such as AGENTS.md or .github/copilot-instructions.md.

Notes

• Keep repository files and skill artifacts in English.
• Keep skills narrow, explicit, and easy for agents to apply.
• Prefer project-local installation unless there is a clear reason to install globally.