Skip to content

Repository files navigation

skill-issue

the agent setup i actually use: twenty-one skills, guard hooks, a save command, and a memory layer that survives every session. works with claude code, cursor, and codex - anything that reads markdown skills, on hosted or open-source models.

each skill is small on purpose: a router, a few opinions, reference files it loads only when needed. no rule dumps - the model already knows how to code. these encode judgment: what to weigh, where the failure hides, when to stop.

setup

paste this repo link into your agent and say "set this up". it follows SETUP.md - skills, plugins, hooks, memory, all of it.

manual: clone it, symlink the skill folders into ~/.claude/skills (or ~/.cursor/skills), install the superpowers plugin so skills fire on intent. exact commands in SETUP.md.

the skills

skill what it does
design-jeev the whole design pipeline: shape, taste rules, build, polish. lab mode renders directions to react to
tune move one dial: bolder, quieter, colorize, distill, animate, clarify
check production-readiness: audit, harden, normalize, adapt, optimize
done-check the done check. greps the diff for shortcuts, runs your tests, maps requirements to hunks
test build the measuring stick first: evals for AI behavior, characterization before refactors
research cheap gatherers, one strong judge, cited memo mapped to your repo
paulgraham stress-test a startup idea. five frameworks, one verdict
ios swift and swiftui: concurrency, navigation, HIG review
python uv-first conventions and the fastapi service shape
i-have-adhd long replies shaped for ADHD: next action first, state restated
apple-user-doc-prose user-doc voice: imperative, calm, plain
html-default-style self-contained HTML house style, no CDN slop
universal-memory the memory layer - see below
unslop remove AI writing tells without inventing anything. compiled from a ten-skill bake-off; rule zero is never add a fact the source lacks
variate design variations of one file on your own localhost, flipped with the arrow keys (vendored, MIT)
defuddle clean markdown out of cluttered web pages (vendored from kepano, MIT)
obsidian-markdown obsidian-flavored markdown: wikilinks, callouts, embeds, frontmatter (vendored from kepano, MIT)
obsidian-bases obsidian .base files: views, filters, formulas (vendored from kepano, MIT)
obsidian-cli drive an obsidian vault from the command line (vendored from kepano, MIT)
json-canvas .canvas files: nodes, edges, groups (vendored from kepano, MIT)
last30days what people said about a topic in the last 30 days, across reddit, x, youtube, hn (fork of mvanhorn's skill, MIT)

skills fire on intent once superpowers is installed. or invoke by name: /design-jeev, /check audit, /done-check.

memory

three files, one law each. SOUL = how it behaves, always. MEMORY = what it knows about you. project CLAUDE.md = repo rules. all loaded every session, so nothing gets re-taught, and a correction becomes a file edit on the spot. templates in universal-memory/templates, global wiring in templates/CLAUDE.global.md.

and no - you don't hand-make memory.md files when context fills. sessions auto-compact.

hooks + commands

hooks/bash_guard.py blocks the irreversible stuff before it runs: recursive rm outside temp dirs, force push, DROP TABLE, hook-bypass flags, commit trailers. hooks/file_guard.py keeps writes inside the project. commands/save.md banks a session's learnings with one word. wiring is SETUP step 7.

your own skills

drop a directory with a SKILL.md in ~/.claude/skills/:

---
name: your-skill
description: Use when [what should trigger it]
---

the description is the router - if a skill isn't firing, fix the description, don't write routing tables.

why the skills are shaped like this

progressive disclosure. design-jeev used to be 1091 lines that loaded whether you wanted them or not. detail moved behind reference files read on demand.

opinions over rules. the font section used to ban 23 typefaces. now it says the reflex pick is the problem - shorter, and actually true.

one owner per fact. motion durations live in one file. severity levels in one file. everything else points at them.

About

agent skills that encode judgment: design, review, testing, writing, research. works with claude code, codex, cursor, and any agent that reads markdown skills, hosted or open-source models

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages