AjeyDS/skills

記録された最新のソース活動: 2026年10月19日 10:24

SkillsMP カタログ更新: 2026年11月12日 12:52

収集済み skills: 11

GitHub スター: 0
GitHub フォーク: 0

AI アシスタントでインストール

このプロンプトをコピーして、使用中の AI アシスタントに送ってください。

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/AjeyDS/skills.

リポジトリの README

README の元ファイル：README.md

README の元ファイルを読む

README 全文を表示

# skills [![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE) <img src="assets/hero.jpg" alt="Five small robots at a workbench, each picking up a different tool from a shelf: a magnifying glass, a compass, a telescope, a root diagram and a book" width="100%"> Your coding agent is clever. It still doesn't know what shipped last week, which of its own lines are dead weight, or how far to trust the branch another agent just pushed. A skill is a small folder of knowledge that an agent loads only when the task needs it. These are the ones I wrote for my own work. They carry the facts, checks and pitfalls a model is missing, and leave the planning to the model. Think of it as handing a sharp colleague the right notebook, not a script to read from. Every skill follows the open [Agent Skills](https://agentskills.io) format, so it works in Claude Code, Claude.ai, Codex, Cursor, Gemini CLI and other agents that read skills. ## A closer look ### ai-slop-detector <img src="assets/ai-slop-detector.jpg" alt="A robot inspector with a magnifying glass and a clipboard of gauges flags duplicated and dead blocks on an overgrown building, next to a clean, trimmed wing" width="640"> Code written by agents tends to grow sideways: near-duplicate helpers, files nothing imports, packages that never existed. This skill runs deterministic analyzers first and reads second, so the audit starts from evidence instead of a vibe. [Read the skill →](skills/engineering/ai-slop-detector) ### external-agent-delegation <img src="assets/external-agent-delegation.jpg" alt="A lead robot at a desk hands task cards to three robots in parallel lanes; each lane ends at a review gate of different thickness, from a light single gate to a heavy double gate" width="640"> Hand work to Codex, Cursor or Gemini running in parallel, then review each result as hard as the model behind it deserves. A strong model in its strong zone gets a light check. A model released this morning gets the double gate. An optional refresh pass looks up the newest models and their benchmarks, even ones that shipped a few hours ago. [Read the skill →](skills/engineering/external-agent-delegation) ### freshness-check <img src="assets/freshness-check.jpg" alt="Small scout drones fly out from a planning board to floating web pages and return with fresh leaves and clippings; one brings back a coral warning card pinned onto the plan" width="640"> A plan is only as current as the model that wrote it. Before you build, cheap scout models search the web to test the plan's assumptions, and the main model decides what actually changes. [Read the skill →](skills/engineering/freshness-check) ### frankenstein-code <img src="assets/frankenstein-code.jpg" alt="A robot in a lab coat stitches together a friendly patchwork robot from parts taken out of labeled crates; each stitched part carries a small provenance tag, and rejected parts sit in a bin marked with an X" width="640"> Half of what you want to build probably exists on GitHub already. This skill splits the goal into parts, finds open-source code for each one, checks its license and activity, takes only the pieces you need, and tags every piece with where it came from. [Read the skill →](skills/engineering/frankenstein-code) ### causal-root-cause-analyst <img src="assets/causal-root-cause-analyst.jpg" alt="Cross-section of soil under a small plant: most roots are grey and crossed out, one glowing coral root is traced by a robot holding a lantern and a magnifying glass" width="640"> A headline is the plant you can see. The cause is somewhere in the roots. This skill collects the competing explanations, tests each one against evidence, and tells you which root it really was, in plain language. [Read the skill →](skills/research/causal-root-cause-analyst) ### study-notes-pdf <img src="assets/study-notes-pdf.jpg" alt="A colorful concept map on the left, a magnifying glass zooming into one node, and a bound study guide with diagrams and a bookmark ribbon on the right" width="640"> Big picture first, then the zoom. You get a researched study guide with diagrams and examples, delivered as a PDF you can actually sit down and read. [Read the skill →](skills/learning/study-notes-pdf) ## All skills ### Engineering | Skill | What it does | Needs | |---|---|---| | [ai-slop-detector](skills/engineering/ai-slop-detector) | Audits a codebase for AI-generated quality decay (duplication, dead code, over-engineering, hallucinated packages) using deterministic analyzers first and reading second. | Python 3.9+; optional lint tools | | [bootstrap-system-md](skills/engineering/bootstrap-system-md) | Explores a repo and writes a `SYSTEM.md`: one tool-agnostic reference every coding agent reads before working, plus shims and a hygiene gate. | Any agent | | [external-agent-delegation](skills/engineering/external-agent-delegation) | Hands repo work to parallel coding agents (Codex, Cursor, Gemini) and calibrates review depth to the model that did the work. | Other agent CLIs | | [frankenstein-code](skills/engineering/frankenstein-code) | Finds open-source code that already solves part of a goal, vets it, takes only the needed pieces and records provenance. | Web + GitHub search | | [freshness-check](skills/engineering/freshness-check) | Sends cheap scout models to the web to test a plan's assumptions against what is current before you build on it. | Claude Code subagents | | [goal-anchored-delegation](skills/engineering/goal-anchored-delegation) | Carries the user's *goal*, not just the plan, into every handoff and grades how much freedom the executing agent gets. | Any agent | | [ui-ux-audit](skills/engineering/ui-ux-audit) | Audits a frontend against UI/UX basics (states, accessibility, tokens, spacing, type, color, hierarchy) and reports fixes with file:line. | Any agent | ### Research | Skill | What it does | Needs | |---|---|---| | [causal-root-cause-analyst](skills/research/causal-root-cause-analyst) | Digs into what is really behind a claim, headline, metric move or incident, tests competing causes against evidence, and tells the verdict as a plain story. Includes DiD, RDD, IV, PSM, Benford and GRIM scripts. | Web search; Python for data mode | | [saving-raw-sources](skills/research/saving-raw-sources) | Collects research sources as RAW files, one per source, with the original wording kept intact. | Web fetch | | [synthesize-research-insights](skills/research/synthesize-research-insights) | Reads a RAW source folder and writes plain-language insights with pointers back to each source. | Local files | ### Learning | Skill | What it does | Needs | |---|---|---| | [study-notes-pdf](skills/learning/study-notes-pdf) | Builds a researched, big-picture-first study guide with diagrams and examples, delivered as a polished PDF. | Python + Playwright | ## Install <details open> <summary>Any agent (skills CLI)</summary> ```bash npx skills add AjeyDS/skills ``` Pick one skill with `--skill <name>`, a target agent with `-a <agent>`, or see the list with `--list`. </details> <details> <summary>Claude Code (plugin marketplace)</summary> ``` /plugin marketplace add AjeyDS/skills /plugin install engineering@ajeyds-skills /plugin install research@ajeyds-skills /plugin install learning@ajeyds-skills ``` </details> <details> <summary>Claude.ai</summary> Download a skill's `.zip` from the [latest release](https://github.com/AjeyDS/skills/releases/latest), then go to **Customize → Skills → + → Upload a skill**. Code execution must be on. </details> <details> <summary>Manual</summary> Copy a skill folder into your agent's skills directory, for example `~/.claude/skills/<name>/` (Claude Code), `~/.agents/skills/<name>/` (Codex and others), `~/.cursor/skills/` or `~/.gemini/skills/`. </details> ## Notes - Skills that run scripts list their requirements in the `compatibility` field of their `SKILL.md`. - `external-agent-delegation` ships the author's dated model notes as an example. Copy `references/your-setup.md` to `your-setup.local.md` (gitignored) and fill in your own agents; let your own record replace the notes over time. - Illustrations were generated with Codex's built-in image generation. - Issues and suggestions are welcome. ## License [MIT](LICENSE)

リポジトリの README を読み込んでいます…

このリポジトリの skills

分類待ち