Liteagents keeps getting thinner and wider.
It started as scaffolding for early models that needed to be told everything. Models got smarter, so the toolkit got smaller. Wide lanes instead of tight rails — constraints on what matters, not scripts for how to do it.
That leaves 13 skills. Five things carry most of the load.
THE RULES — AGENT_RULES.md
One generic, model-agnostic rules doc, loaded every session. It sets what matters and leaves wide room on how: simple over clever, every line has a purpose, surgical changes, exhaust the stdlib before adding a dependency, POC before you design. /remember keeps your copy current against the template that shipped with your install.
MEMORY — /stash → /remember
Your project memory and how the model actually behaves with you, in one markdown file that's hot in context every session. /stash snapshots a session before it compacts. Every 5 stashes, /remember folds them in, and reads across all your projects for where you and the model collided: corrections, dead ends, abandoned flows. Those become rules. No database, no RAG.
SHIPPING — /branch-review → /release → /refactor
Two things always hurt: where do I start this review, and how do I ship without drowning in nits.
/branch-review marks HEAD and reviews the whole branch from that mark — full security pass every run, medium depth by default, and it only reports what it can actually confirm. Nits get pushed to a fix ledger for later.
/release does the docs sweep, version bump and local commit, then hands you a ready-to-merge PR. It never pushes.
/refactor clears that ledger whenever you feel like it.
All of it saved to disk, so you don't have to remember where you left off.
DOCS — /docs-builder
The recent one I'm gladdest about. My PRD often grew past 3,000 lines carrying decisions, findings and learnings across five subdirectories, and I'd lost track of what was in there. /docs-builder indexes everything under three, splits the oversized docs, and keeps them clean. It runs on its own as your markdown grows, or point it at any long file to split it.
UI — /live-canvas
A terminal agent can't see your screen, and "that padding, on that card, but only on mobile" costs ten minutes to describe. /live-canvas puts your UI on a local page. You click the thing that needs changing, type what you want, and send the batch. It also generates variations of a screen so you can pick instead of guess.
npx liteagents
https://github.com/hamr0/li...