Eight agent skills I use every day — for deciding, remembering, and improving, not just generating code. Small, composable, model-agnostic. Built for Claude Code, Codex, Cursor, and any agent that reads a SKILL.md.
Most of an agent's failures aren't coding failures. The agent agrees when it should push back. The big decision gets made in a chat window. The setup quietly rots. The session ends and the context is gone. These skills are built for exactly those moments.
They aren't a framework. Each skill is a single self-contained directory you can read in a minute and fork into your own. No process owns your work; the skills give you sharper tools and get out of the way.
- Run the skills.sh installer. It auto-detects which coding agents you have (Claude Code, Codex, Cursor, …), prompts you to pick which skills to install, and copies them into the right place. Works with any public GitHub repo — no registration needed.
npx skills@latest add ngmeyer/skills-
Pick the skills and agents you want.
-
That's it. Each skill is self-contained — no per-repo setup step. Invoke them directly:
/six-pager /council-review /adversarial-review
/claude-md /session-close /skillforge
Other install methods (Claude Code plugin marketplace, manual)
For the official Claude Code plugin flow. Two steps:
# 1. Register this repo as a plugin marketplace
/plugin marketplace add ngmeyer/skills
# 2. Install individual skills
/plugin install six-pager@ngmeyer-skills
/plugin install council-review@ngmeyer-skills
# ...one per skill you want
Then /reload-plugins to activate. Marketplace-installed skills are namespaced — invoke them as /ngmeyer-skills:six-pager, /ngmeyer-skills:council-review, etc.
git clone https://github.com/ngmeyer/skills.git ~/ngmeyer-skills
# Symlink the ones you want into your agent's skills dir
ln -s ~/ngmeyer-skills/skills/productivity/six-pager ~/.claude/skills/six-pagerI built these to fix the failure modes I kept hitting with Claude Code and other agents — the ones that have nothing to do with whether the code compiles.
"Where all think alike, no one thinks very much."
Walter Lippmann, The Stakes of Diplomacy
The Problem: An agent's default is to agree. Ask "is this a good plan?" and you get a confident yes. That's the most dangerous failure mode in a thinking partner — it feels like validation and it's worth nothing.
The Fix is manufactured dissent. Make the agent argue against you before you commit:
/council-review— runs your question through five advisors with distinct reasoning methods, who peer-review each other anonymously, and a chairman synthesizes the verdict. Use it for open questions with real stakes./adversarial-review— a single critic that actively tries to break a finished thing: a PR, a spec, a plan, an argument. Use it before you ship.
These pair: decide with /council-review, then stress-test the artifact with /adversarial-review.
"There is no way to write a six-page, narratively structured memo and not have clear thinking."
Jeff Bezos, 2017 Amazon shareholder letter
The Problem: Decisions worth a week of consequences get made in a few sloppy chat turns. The bullet points hide the gaps. Nobody writes down what would make this fail.
The Fix is to force the prose. Writing is the thinking instrument, not the record of it.
/six-pager— generates an Amazon-style narrative decision memo (or a PR/FAQ for a launch), under Strunk's prose rules, with a required premortem: assume it's twelve months later and this failed — name the three most likely causes now, while you can still act on them.
"Don't live with broken windows."
David Thomas & Andrew Hunt, The Pragmatic Programmer
The Problem: A CLAUDE.md starts sharp and drifts. It claims facts that no longer match the code, accumulates bloat, sometimes leaks a secret. And the wider setup — skills, workflows, context files — decays one small unaddressed thing at a time.
The Fix is regular, opinionated maintenance:
/claude-md— two modes:auditfinds drift, leaked secrets, duplicates, and budget bloat across all yourCLAUDE.mdfiles;improvemeasures one against Anthropic's best-practice rubric and proposes concrete rewrite diffs./weekly-setup-improvements— audits a week of work and writes a forward-looking improvement report: a closure check on last week's actions, zombie-action kills (dropped twice → killed, never re-surfaced), and auto-drafted scaffolds for the skills you keep wishing you had. Safe to schedule weekly.
"The palest ink is better than the best memory."
Chinese proverb
The Problem: Everything the agent learned this session — which decisions you made, which paths turned out to be dead ends, and why — evaporates when the window closes. Worse, a folder move or a second working directory can split your project memory into two namespaces, and the work you saved appears to vanish.
The Fix is to write memory down, and to repair it when it splits:
/session-close— reconciles a session's outcomes into persistent project-memory files with section-aware merges. Not a session dump — an update./session-recover— finds duplicate project directories that point at the same logical project from different paths, unifies the memory into the canonical one, and captures the lesson so the split doesn't repeat.
"Without data, you're just another person with an opinion."
W. Edwards Deming
The Problem: You write a skill, it works well enough, and it never improves. "Better" stays a vibe. There's no measurement, so there's no progress.
The Fix is to treat a skill like any other artifact you can measure and tune:
/skillforge— two modes:forgescaffolds a new skill the right way (frontmatter, progressive disclosure, helper scripts, mandatory Gotchas);optimizemakes an existing skill measurably better at its outcome, not just its packaging — a quality audit, domain outcome-research, and a held-out train/validation A/B that catches overfitting. Every skill in this repo was built and tuned with it.
The fundamentals of good thinking don't change because an agent is doing the typing. You still have to disagree before you commit, write to find out what you actually think, and maintain what you build. These skills are my best effort at making an agent do that by default. Enjoy.
Code-facing skills.
- adversarial-review — Single-critic red-team of a finished artifact: edge cases, hidden assumptions, failure modes. Actively tries to break it before it ships.
- claude-md — Two modes for
CLAUDE.mdfiles:auditfinds drift, leaked secrets, and bloat across your projects;improverewrites one against Anthropic's best-practice rubric. - session-recover — Recover lost context by merging duplicate project-memory directories that a folder move split into two namespaces.
Thinking and process skills, not code-specific.
- council-review — Run a decision through a 5-advisor multi-agent debate (DMAD) with anonymous peer review and a chairman synthesis. Built to beat sycophancy.
- six-pager — Amazon-style narrative decision memo or PR/FAQ, under Strunk prose rules, with a required premortem.
- session-close — Reconcile a work session into persistent project memory with section-aware merges — not a session dump.
- weekly-setup-improvements — Audit a week of work and produce a forward-looking improvement report: closure check, zombie-action kills, auto-drafted skill scaffolds.
- skillforge — Forge a new skill, or
optimizean existing one to a measurably better V2 (quality audit + outcome research + train/validation A/B).
Every skill is one directory with a SKILL.md (YAML frontmatter: name, description) plus optional references/ (progressive-disclosure depth) and scripts/ (deterministic helpers). Skills live under skills/<category>/<name>/.
See CONTEXT.md for the design philosophy and CLAUDE.md for how to add or evolve one.
MIT © 2026 Neal Meyer