agentic-skills

Research and academic writing: deep research briefs, literature reviews, sourced comparisons, external lookups, plain-English explanations, paper drafts, peer reviews and paper-against-code audits. Ships researcher, writer, verifier and reviewer agents that gather evidence, draft from it, cite it and critique the result.

9 skills · 4 agents

Splits work across Claude models. Ships three read-only Haiku subagents (repo-map and call-sites query a local Graphicle code graph and check it against the source, api-docs fetches library documentation), a delegate skill that tells the main session when to use them, a setup skill that configures the main model, the Opus advisor, Flightdeck, permissions and the CLAUDE.md rules, and hooks: a Sonnet gate on six risky Bash command families and a Stop check for missing test reports and breaches of the global CLAUDE.md writing rules. The gates are second opinions that fail open on timeout.

2 skills · 3 agents · 1 MCP server · hooks

Design and build in a loop, for Python and for TypeScript or JavaScript (Node.js, MERN). An architecture design skill picks module boundaries and design patterns from named forces, has a reviewer check the draft and ends at a human gate. An implement skill then writes acceptance tests first, has them reviewed, locks them with a hook, implements in the main session and runs three read-only reviewers in a clean context.

2 skills · 3 agents · hooks

Authoring workflows for Claude Code primitives: skills, hooks, subagents and instruction files (CLAUDE.md and rules). Each follows a draft, test, review and iterate loop against the current reference docs, and delegates to its siblings when a job needs more than one primitive.

4 skills

Consolidates Claude Code native auto memory. A SessionStart hook decides when a pass is due and has the main session start a visible dreaming:dreamer subagent, which writes a proposal and a report to a staging directory. The dreaming:dream skill shows the diff, snapshots memory in git and applies it only on approval, and a PreToolUse hook confines the subagent to staging. Requires jq.

1 skill · 1 agent · hooks

Git workflows: reviewed commits that split unrelated changes and carry the conversation's reasoning into the message, and a pull request flow with tests, scoped lint and type checks, a review gate, CI-gated merge and resync. Never bypasses hooks or force-pushes.

2 skills

Hugging Face hub assistant — model and dataset profiles, prompt and chat templates, inference providers, licences, linked papers, and search for models suited to agentic or tool-use workloads. Bundles the official Hugging Face MCP server, so no separate connector setup is needed.

1 skill · 1 MCP server

Guards chezmoi-managed dotfiles as a hook. Refuses an Edit, Write or NotebookEdit whose target is a chezmoi destination file, and names the source file to edit instead. Edits inside the chezmoi source directory always pass. The managed list is cached for 30 minutes, and the hook fails open whenever chezmoi is missing, uninitialised or errors. Requires jq and chezmoi.

hooks

Project dotfile management through the dotset CLI — .gitignore, .uvgroups, .envrc, .cleanup, .skyignore, and .rsync-exclude. Use when adding or removing ignore patterns, managing UV groups, or initialising a project's dotfiles.

1 skill

Diagnostic-first debugging on Opus at high effort. Gathers environment and run-condition evidence, reads the code, the full terminal output and the installed library source, reads upstream issues and PRs in full, then delivers a ranked-hypothesis report with the provenance of every claim, and stops there before any fix.

1 skill

Consolidates the Claude desktop app's memory store through the mcp__memory__ file tools. The dream skill writes a verbatim backup file first and has you save it, then merges duplicates, drops stale lines and adds durable facts from the current conversation, and writes only on approval. A revert mode restores the backup. Skill only, no agents or hooks.

1 skill

Multi-agent architecture selection for LangChain and LangGraph. Argues for a single agent first, then maps constraints onto subagents, handoffs, skills, router, or a custom workflow — delivering an architecture decision record with an explicit call and token cost model plus a minimal scaffold.

1 skill

Deterministic JavaScript and TypeScript quality gates as hooks. They refuse writes containing private keys, auto-fix formatting and lint after every edit, and block the turn until the files changed in it pass Biome (or ESLint and Prettier when the project configures those) and the tsc type check. Requires jq. Every tool is taken from the project's node_modules/.bin and never from npx.

hooks

Deterministic Python quality gates as hooks — refuses writes containing private keys, auto-fixes formatting and lint after every edit, and blocks the turn until the Python files changed in it pass ruff and ty. Requires jq; ruff, ty and pre-commit-hooks are each resolved from PATH first and fall back to uvx.

hooks

Redirects pip to uv as a hook. Refuses a Bash command that runs pip install or pip uninstall, whether spelled pip, pip3.12, .venv/bin/pip or python -m pip, behind VAR=value, sudo, env or xargs, anywhere in a compound command or backtick substitution. Read-only pip subcommands, uv pip, pipx, pipenv and pip mentioned inside quotes or a heredoc all pass, so a command hidden inside bash -c or a quoted $(...) is not seen. The refusal names uv add or uv remove inside a uv project and uv pip outside one. Requires jq.

hooks

Writing for client deliverables: a copyedit pass for English and German reports, assessments, proposals, statements of work, emails and slide text that removes AI tells, fixes locale formats and consistency, keeps every number and commitment intact and reports what changed. Ships a deterministic prose lint with a compare mode.

1 skill