Six procedures your agent loads on demand: explain a thing to the right audience, stress-test a plan, get an independent verdict before "done", split heavy work across models, measure how agent-ready a repo is, and learn a codebase with spaced repetition. Works in Claude Code, Codex, Cursor, OpenCode, pi and any tool that reads SKILL.md.
npx skills@latest add mneves75/skills
git clone https://github.com/mneves75/skills.git ~/.claude/skills/mneves-skills
Each skill is a folder with a SKILL.md: a description the agent matches against your request, then the procedure it follows. No servers, nothing phones home.
Explains one idea with the Feynman method, calibrated to the listener: child, layperson, executive, junior, expert. Answers in your language.
Say: "explain OAuth to my dad", "ELI5", "explique de forma simples"
How to use, with an example →A twelve-step pass on a draft: first principles, hostile panel, six alternatives compared on weighted criteria, pre-mortem, five-year test. Returns the improved deliverable.
Say: "challenge this", "stress-test", "best possible answer"
How to use, with an example →A fresh context checks the work against criteria frozen up front, with real evidence: tests, renders, the actual binary. Verdict is PASS, FAIL or BLOCKED. Only PASS allows "done".
Say: "verify it", "prove it", "fix all"
How to use, with an example →One model plans and reviews; others type. Heavy backend work goes to Codex through codex-lane, which keeps a thread alive across rounds so fix-ups reuse the executor's reasoning.
Say: "delegate", "use codex", "heavy task"
How to use, with an example →Scores a repository on how well it supports an agent's loop: context, action, verification, iteration. Nine pillars, a percentage, a level from L1 to L5. Runs locally.
Say: "why does the agent struggle here?", "agent readiness"
How to use, with an example →You explain part of the codebase; the agent reads the real code, asks what you skipped, and turns each gap into a flashcard scheduled with SM-2 in a per-project SQLite file.
Say: "let me explain the auth flow", "quiz me", "what's due?"
How to use, with an example →A skill is text the agent reads at the right moment. The install just puts the folders where your tools look.
npx skills add mneves75/skills links each skill into the agents on your machine (-g for user-wide, --all -y for no prompts). Or clone the repo into a skills directory.
Every SKILL.md starts with a description that names when to use it. When your message fits, the agent loads the procedure. You can also call one by name: /mneves-verify.
Steps, rules and output formats are followed with your tools and your permissions. Two skills ship a small script (codex-lane, srs_db.py); the rest are Markdown only.
mneves-agent-readiness ships a Bun + TypeScript assessor with no runtime dependencies. Point it at a repo and it tells you where the agent's feedback loop is broken.
cd /path/to/your-project
bun --bun /path/to/skills/tools/readiness-check.ts \
--format=html --output=report.html
# static checks only, fast
bun --bun /path/to/skills/tools/readiness-check.ts \
--skip-tests --skip-build
# gate a CI job
bun --bun /path/to/skills/tools/readiness-check.ts \
--min-level=39a8a13f, static checks, 2026-08-26
Style 88% · documentation strong · task discovery and product signals weakest
SKILL.md files are short. Reading one is the fastest way to know what your agent will do.
Write your ownFolder name equals frontmatter name, a description that says when to use it, no home paths. CI checks the rest.
mneves-fable-orchestrator routes the heavy executor to GPT-6-astra at reasoning high (was GPT-5.6-sol); the GPT-5.6-terra bulk tier is dropped, exceptions are pinned per lane.HOWTO.md: mneves-eli5 and mneves-expert-review examples are verbatim claude -p runs (claude-fable-5, 2026-08-26); mneves-teach-back-srs shows real srs_db.py output; the readiness example was already a real run.mneves-fable-orchestrator adopts five rules from steipete's codex-first (2026-08-20 revision): every hard prohibition in a spec carries an escape hatch (the minification incident), check for a live worker with pgrep -fl "codex exec" before editing, verify beyond the diff (merged surface, uncommissioned commit messages, guard files, test-helper edits), a diagnosis list for executors that die instantly or go quiet (with lane resume), and never passing credentials via -c (use a CODEX_HOME overlay).site/index.html): install commands with copy buttons, one card per skill linking into the how-to, how it works, the readiness tool with the sample score, start-here links, recent changelog.