So frequently, we find ourselves just downloading and toying with some new skills to see how the behavior feels for our workflows. This repository is the new staging ground for my own approved tools.
converge finds unnecessary complexity through consumer tracing and evidence.
It can audit a repository, review a finding, and apply an authorized cleanup.
It started in Bar Savvy and now discovers each project's architecture and verification commands.
Run equip and select converge to install it. Start with $converge audit in the target project.
A number of these skills come from engineers that have worked hard and promoted their work publicly. It'd be extremely disingenuous to not support them and the impact that their skills have had on my engineering.
The skills below started with their work.
Matt Pocock — mattpocock/skills
The deepest single influence here. Nine started as theirs:
codebase-design, grill-me,
grilling, handoff,
improve-codebase-architecture, tdd,
to-spec, writing-great-skills,
zoom-out.
shadcn — shadcn/improve
improve — survey a codebase as a senior advisor and hand off prioritized,
self-contained plans for other agents to execute. Read-only on source.
Vercel Labs — vercel-labs/agent-browser
agent-browser — the browser automation layer I reach for on every
dogfood and QA pass.
Matt Van Horn — mvanhorn/last30days-skill
last30days — research what people are actually saying about a topic right
now, across Reddit, X, YouTube, HN and the rest. Nothing else here does that job.
Peter Yang — petergyang/no-ai-slop
no-ai-slop — edits drafts into sharper, more human writing without
flattening the writer's voice. MIT; the upstream LICENSE is retained in the skill directory.
Ayoub Ghriss - ayghri/i-have-adhd
i-have-adhd - Shape output for a reader with ADHD: lead with the next action, number multi-step work, restate state across turns, suppress tangents, give specific time estimates, make wins visible.
Vendored copies drift from upstream once adapted. Check the source repo for the canonical
version and its license before redistributing any of these. last30days
ships without its upstream assets/ directory — 14 MB of demo media that SKILL.md never
references.
rules/ holds universal engineering rules: one markdown file per rule, with a
regex trigger and a scope in the frontmatter and the corrective instruction in the body.
They are the minimum standard of care for any project, defined once and enforced in every
harness. hooks/rule-gate.mjs evaluates them as PreToolUse/PostToolUse
hooks in Claude Code and Codex CLI; omp reads the same files natively through the
~/.agents/rules symlink and enforces them as TTSR rules. The format and the tier model are
documented in rules/README.md.
home/AGENTS.md is the prose tier: global guidance that loads into every
session but is not mechanically enforced.
Run install.sh to deploy everything global: the ~/.agents/rules symlink,
the ~/.claude/CLAUDE.md symlink, and the rule-gate hook wiring for Claude Code and Codex.
For skills, run the interactive installer from any directory:
equipBoth install.sh and a plain npm install in this repo put equip on your
PATH: the prepare script symlinks it into npm's global bin directory. Pick the
skills you want, then pick a scope. Both scopes use one layout, the same one
npx skills uses: the skill lives in .agents/skills, and .claude/skills
points at it with a relative symlink. Globally the skill is itself a symlink
to this repo, so git pull updates it in place. In a project it is a real
copy, so collaborators get everything from that repo.
Releases are cut as git tags by .github/workflows/release.yaml
via changesets — the package is private, so nothing publishes to npm; pin to a tag if you want
a stable snapshot.
Packaging this as a Claude Code plugin marketplace — so others can
/plugin marketplace add eggdev/skillsand install pieces à la carte — is the intended distribution path. Not done yet.