From huntse-agent-skills
Read a codebase, build a deep understanding of it, and produce a written plan of candidate simplifications (dead code, single-use abstractions, duplication, over-engineering) ranked by impact-to-risk ratio. Writes the artifact to ~/tmp/ so it can be fed straight into /cook and then /handoff for unattended execution. Use when the user asks to audit, simplify, declutter, or shrink a codebase, find dead code, or prep a simplification pass for a background agent. The output is a plan, not edits — this skill never modifies the code under review.
How this skill is triggered — by the user, by Claude, or both
Slash command
/huntse-agent-skills:plan-simplificationsThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Read the codebase under `cwd` (or a user-supplied path), understand it, and write a ranked simplification plan to `~/tmp/simplifications-YYYY-MM-DD-{slug}.md`. The deliverable is a markdown doc the user can hand to `/cook` — this skill **never edits the code being audited**.
Read the codebase under cwd (or a user-supplied path), understand it, and write a ranked simplification plan to ~/tmp/simplifications-YYYY-MM-DD-{slug}.md. The deliverable is a markdown doc the user can hand to /cook — this skill never edits the code being audited.
Confirm scope — default is the entire repo at cwd. If the tree is large (>200 source files) or the user has hinted at a subset, ask one question: whole repo, a specific dir, or a file list. Skip vendored/build dirs (node_modules, .venv, dist, build, .git) regardless.
Read systematically — don't simplify what you don't understand. In order:
README.md, CLAUDE.md, pyproject.toml / package.json / Cargo.toml, build files.main, CLI commands, HTTP route registrations, exported package surface.Build internal understanding — before listing candidates, be able to answer: what is this code for? What's the public contract? What are the hot paths? What's the test coverage shape? If you can't, keep reading.
Identify candidates — walk these categories explicitly. Each finding gets a file path with line numbers, an estimate of lines removed, and a risk note:
Rank — sort by impact (lines/abstractions removed, clarity gained) divided by risk (blast radius, test coverage at the site, whether behavior changes). High-impact-low-risk first.
Split safe vs behavior-changing — anything that touches a public signature, CLI flag, HTTP route, schema, or env var goes into the Behavior-changing section as opt-in, not the default bundle. Default bundle = strictly observable-behavior-preserving.
Write the artifact to ~/tmp/simplifications-YYYY-MM-DD-{slug}.md using the template below. Pick a slug from the repo name or focus area.
Tell the user the path and suggest the next step: /cook to turn the recommended bundle into an unattended plan, then /handoff to dispatch it. Do not invoke /cook yourself — the user reviews first.
---
written: YYYY-MM-DD by Claude Code
repo: {absolute path}
scope: {whole repo | path | file list}
commit: {git rev-parse HEAD}
---
# Codebase summary
{3–6 sentences: purpose, shape, entry points, test posture, anything unusual.}
# All candidates (ranked, highest impact-to-risk first)
## 1. {short title} [impact: ~N LOC | risk: none/internal/observable]
**Where:** `{absolute path}:{line}` (+ other call sites)
**What:** {one sentence}
**Why it's safe:** {test coverage, no external callers, etc.}
**Steps:** {2–4 bullets the executor can follow}
## 2. ...
# Recommended bundle (this run)
- Candidates 1, 3, 5 — total ~N LOC removed, all internal-only.
# Behavior-changing (opt in explicitly via /cook)
## B1. {title} [risk: observable — {what breaks}]
...
# Verification plan
- Run `{exact test command}` — must stay green.
- Spot-check: `{grep / build / smoke command}`.
# Do not touch
- {public API surfaces, files, contracts that are out of scope}
Provides behavioral guidelines to reduce common LLM coding mistakes, focusing on simplicity, surgical changes, assumption surfacing, and verifiable success criteria.
Searches, retrieves, and installs Agent Skills from prompts.chat registry using MCP tools like search_skills and get_skill. Activates for finding skills, browsing catalogs, or extending Claude.
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.
npx claudepluginhub huntse/agent-skills --plugin huntse-agent-skills