Tatsuya Shimomoto

A Claude Code mod that turns your global skills, agents, rules and tools on or off per repo with named profiles. In my writing repo, it hid 94 of the 101 skills with one line.

Claude Code skill: forces agents to research existing solutions before building — reduces reinvention

Claude Code skill: forces agents to research existing solutions before building — reduces reinvention

Three Agent Skills for autonomous LLM agents (replayable audit logs, read-only instruments, shadow-mode validation) that make an agent's decisions reconstructible and let a new decision mechanism earn evidence before it steers live behavior; the installable form of AAP ADR-0006/0005.

Writes AI-facing documents (llms.txt / llms-full.txt / FAQ / glossary) optimized for citation by ChatGPT, Perplexity, Gemini. Answer.AI llms.txt standard + GEO-SFE 3-layer static analysis.

An Agent Skill for Claude Code that audits all your installed skills and proposes a verdict for each one, such as Keep, Improve, Retire or Merge, and changes nothing until you confirm. Part of the Agent Knowledge Cycle's Curate phase; install it together with skill-health.

Claude Code skill: audits installed skills for staleness, conflicts, and redundancy

Agent Skill that routes AI work to a script, algorithmic search, an LLM workflow or an autonomous agent with five questions before anything is built, by asking whether accountability can still be distributed after deployment; the installable form of the AAP quadrants navigator (ADR-0009/0010).

Retired Claude Code skill, kept as a public record: a read-only wrapper around the OpenAI Codex CLI that gets a second model family's review of a diff or a design plan. The author's harness now uses OpenAI's official Codex plugin instead.

Claude Code skill: designs and ships a companion JSON-LD knowledge graph (graph.jsonld) next to llms.txt for projects with stable concept-level structure — encodes domain entities and relationships as schema.org triples for LLM citation

Claude Code skill that loads the Authorship Strategy judgment framework into a coding agent, so it can evaluate a concrete plan against your authorship strategy, check its adoption concerns, or explain the framework, answering with value and concerns rather than a checklist.

Agent Skill that keeps the citations of a research repository in sync: it audits whether each external paper cited in the docs also appears in .zenodo.json references and graph.jsonld, then syncs the missing layers bottom-up.

Claude Code Agent Skill bundle for writing & reviewing academic papers (SSRN / arXiv / Zenodo / journals). Bundles an orchestrator skill, a draft skill, and 5 reviewer subagents — first repo to package skills + agents together via install.sh.

A Claude Code skill and six review subagents for writing human-facing articles, essays and newsletters: one central thesis agreed with the author before drafting, a reviewer panel that reads the frozen draft once, and the author's go-ahead before titles and publication.

Agent Skill for developers building LLM pipelines and agents: four patterns for layering deterministic code and LLM calls in one pipeline (guard, filter, judge + enforce, orchestrator + worker), where code owns control flow and the LLM owns meaning.

Agent Skill for Claude Code that audits a repository's context documents (CLAUDE.md, ADRs, README, graph.jsonld, llms.txt) for role overlap and staleness against the code, moves each fact to one home, and flags README findings instead of editing them.

Agent Skill for Claude Code that extracts reusable patterns from a session, checks each one against what actually happened, and saves it only where a future session will find it: absorbed into an existing skill, rule or doc, or promoted to a skill of its own.

Agent Skill with structural security for autonomous LLM agents: dangerous capabilities are never implemented instead of guarded, accumulated content is treated as untrusted, and each agent process gets one external adapter; the installable form of AAP ADR-0001 to 0004.

No change: keep "Claude Code skill that rewrites or reviews a README so first-time visitors can tell what the project is; a fresh judge agent checks the page against your code and returns sentence-level fixes, not a score."

Claude Code skill that runs the release of a DOI-registered research repository on GitHub and Zenodo: verify, sync the citation files, cut the release that mints the new version DOI, then propagate it, following the authorship-strategy identifier-federation decisions (ADRs 0001-0003).

Agent Skill for Claude Code that audits a project repository's non-code assets (tool configs, CI workflows, runbooks and docs) for whether anything still uses them, and gives each a Keep, Update, Retire or Merge verdict on top of structural linters.

Agent Skill for Claude Code that scans your installed skills, finds principles that belong in the always-loaded rules (facts, wiring and traps specific to your environment), and drafts them as rules, editing nothing until you confirm each one.

An Agent Skill for Claude Code that audits your always-loaded rules (~/.claude/rules) for what each line costs in every session and gives each rule a verdict, such as Keep, Improve, Demote to skill or Dissolve. Part of the Agent Knowledge Cycle's Curate phase.

Frozen Agent Skill for designing a research intake filter that admits only information likely to change your next action: search widely, intake narrowly. Kept as a public record; the principle now lives in skills such as search-first.

Agent Skill for Claude Code that measures whether coding agents actually follow a skill, rule or agent definition: it generates scenarios from supportive to competing prompts, runs an agent on each, and reports compliance rates with classified tool call timelines.

An Agent Skill for Claude Code that scans a skill library for references to scripts, agents, files or sibling skills that no longer exist, deterministically and without auto-fixing. The code layer of the Agent Knowledge Cycle's Curate phase.

Frozen public record of an Agent Skill that decides, for one task, whether to write deterministic code or call an LLM: classify the property as structural or semantic, apply a false-positive test, and split enumeration from judgment.

A daily research note per research line: your questions steer the search, TypeSafe Jev makes the judgments, an LLM writes the explanation (ja / en / zh).