Skills
The Claude Code plugins and skills I keep enabled, the ones I turned off, and my global CLAUDE.md.
Plugins and skills change how the agent works. They cost context just by being enabled, so this list is kept short on purpose.
Behaviour plugins
These load at session start and shape every response.
ponytail: forces the laziest solution
that works. It runs a ladder before any code is written: does this need to exist, is it
already in the codebase, does the standard library do it, does the platform do it, can
it be one line. It stays because agents over-build by default, and a smaller diff is
cheaper to review. It also ships ponytail-review and ponytail-audit for hunting
over-engineering in existing code.
caveman: compresses the agent's prose. It removes articles, filler, and hedging. Technical terms, code, and error strings stay exact. It drops back to normal prose for security warnings and irreversible actions, and for anything written to disk (commits, docs, PRs). It stays because output tokens are the slow, expensive ones.
superpowers: a set of process skills for TDD, systematic debugging, brainstorming, writing and executing plans, git worktrees, and a "verify before claiming done" check. It stays for the debugging and verification skills, which push the agent to find a root cause instead of patching the symptom.
Ponytail and superpowers pull in opposite directions. Ponytail says do less. Superpowers says follow the process. The tension is useful: the process skills decide how to approach a problem, and ponytail keeps the answer small.
Review skills
- swe-guidelines: architecture review through seven lenses (object model, contracts, context, storage, async, network, delivery), plus scaffolding commands.
arch-*skills:arch-compareto choose a pattern,arch-buildto implement it,arch-diagnosefor one misbehaving symptom,arch-reviewfor a full audit,arch-adrto record the decision.- postgres-ops-review: audits a Postgres server itself: config, bloat, autovacuum, backups, replication.
- grilling: makes the agent interrogate my plan instead of agreeing with it.
Design skills
- Figma plugin: reads and writes Figma files: design to code, and code back to Figma.
- ui-ux-pro-max: a reference skill of styles, palettes, font pairings and stack conventions, so UI work starts from a deliberate direction.
/mock-options: my own command. It builds N design variants on a throwaway/mock/<slug>route, I pick one, it applies the winner and deletes the route.
pstack: the skill library
pstack comes from Cursor's plugins
repo. I run it adapted for Claude Code and Codex, installed under ~/.agents and
symlinked into each tool, so all three agents share the same skills.
| Kind | Skills |
|---|---|
| Explaining | how (how does this work), why (why is it this way), teach, bro (plain-language restatement) |
| Writing | unslop (cut AI tells from prose), technical-writing |
| Parallel work | arena (N candidates, graft the best parts), swarm (fan out workers), interrogate (multi-model adversarial review) |
| Thinking | architect, figure-it-out, blast-radius, reflect, recall, show-me-your-work |
| Principles | 23 principle-* skills, one engineering rule each: fix root causes, subtract before you add, make operations idempotent, test behaviour not implementation, guard the context window |
It also brings two agents: comment-sicko, a reviewer that deletes needless comments,
and poteto-agent.
The principle skills are the part I would copy first. Each is a few lines, and naming one in a prompt is shorter than explaining the rule again.
Bigger workflows
- greenfield: reverse-engineers a codebase into behavioural specs with provenance.
- iterative-development (same marketplace): extract requirements, scope the simplest core, run iterations, audit progress.
What I turned off
Installed but disabled: the Vercel plugin, clangd-lsp, rust-analyzer-lsp, the lean4
skills and a cybersecurity skill pack. A further dozen research skills (literature
review, paper lookup, data analysis) sit disabled in a skills-disabled/ folder.
None of them are bad. Every enabled skill adds its description to every session, and the agent has to consider it on every prompt. I enable a pack for the project that needs it and disable it afterward.
Global CLAUDE.md (excerpt)
The rules every project inherits.
## Implementation
Write the code yourself in the main session. Subagents are for read-only fan-out
(searching many files, parallel investigation), not for building.
Use bun, not npm. Default new web/Vite projects to TypeScript.
## Done means verified
Before claiming done, merge-ready, or fixed: run build/tests/lint yourself, or actually
load the affected screen, and report the result. If a check fails, say so with the output.
## Commit granularity
One behavior per commit. Stage explicit paths. If the message needs "and", split it.
Target <=5 files per commit. Merge or rebase branches, never squash-merge.
## Recommending options
When presenting choices, give 5 options and mark one recommended. One line of why.
## rtk
Prefix Bash with `rtk` when output is large and detail is not needed. Use the native
command when exact bytes matter.
## Database
Schemas in 3NF: no derived columns. Table names singular. Avoid reserved words.