skill
radin
🐀 the agentic stack for those who have to save tokens
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/shortcuts/radin itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 25, 2026A Claude Code plugin for a solo dev on a small subscription. One curl | bash
gives you:
- A backlog-driven workflow. Tasks live as files in your repo, so they survive the conversation. Plan them, execute them one commit at a time, review the result. Any run resumes where it stopped.
- A curated token-reduction stack, installed whole: rtk, caveman, ponytail, codebase-memory-mcp, thermo-nuclear, mattpocock-skills, unslop, technical-writing. radin delegates to them instead of reimplementing them. It never forks or vendors them.
Install
curl -fsSL https://raw.githubusercontent.com/shortcuts/radin/main/install.sh | bash
Needs curl, tar, bash, and the claude CLI. Two files are C, the TUI and the
JSON surgery behind radin repair. Install builds both with cc (Xcode
Command Line Tools on macOS, gcc on Linux) and skips each with a warning when
no compiler is there. Companion tools pull their
own stacks; one that fails is reported and skipped, and radin still installs.
rtk installs through Homebrew, mise, or its own
installer. The install asks which.
Install asks three questions: one about radin-batch's sub-agent models,
one naming the package manager for companion tools, one about sub-agent effort. Add -s -- --yes to take the defaults. Add -s -- --verbose to
trace every command, for an install that hangs or fails.
Update radin itself with radin update. It asks the sub-agent model and effort questions
again, reapplies every companion's configuration, and installs only the
companions that are missing. Re-run install.sh to update the companions too.
Use it
| Command | What it does |
|---|---|
/radin-record | Turn a bug, idea, or aside raised mid-session into a backlog task |
/radin-plan | Write a step-by-step plan for one task, touching no code |
/radin-batch | Work the backlog: prioritize, plan the tasks /radin-record flagged, implement, commit, one task per commit |
/radin-do | Record one ad-hoc prompt as a backlog entry, plan it when flagged, then execute just that task |
/radin-review | Two-axis pass over a scope, repo rubrics plus the originating entry's spec; kept findings go to the backlog or straight to /radin-do |
/radin-stats | Every companion tool's own savings report, side by side |
/radin-uninstall | Remove what install.sh put in ~/.claude |
radin | Full-screen backlog browser for humans. Bare radin on a terminal is the human entry point. No agent involved; ? lists keys. Off a terminal it prints the usage text and exits non-zero, so a skill or pipe never hangs in a full-screen app. radin help prints that text on purpose |
Typical loop: record, plan (optional), execute, review. Review findings feed the next
execute. Each skill runs in your own conversation, so it can ask you
questions; only leaf work goes to sub-agents. /radin-batch confirms the
order before it starts, and every task's state is on disk, so interrupting it
costs nothing. Want your thread free? Dispatch it from claude agents or a
second claude session. The worktree-per-task answer keeps concurrent runs
off each other's checkout.
Where things live
One directory per repo, <repo-root>/.claude/.radin/, holds the backlog index, one
markdown file per task, plans, reviews, and execution state. Commit it to
share the backlog with your team, or gitignore it to keep it private. radin
never touches your .gitignore.
Every tool goes through the radin CLI (radin backlog, radin state), so
nothing hand-edits those files. A task carries a category
(feat/fix/chore/refactor), an optional priority and depends_on, and
can sit under an epic whose DESCRIPTION.md every child inherits.
What it shops for
| Tool | What it does |
|---|---|
| rtk | Wraps dev commands, cuts their output 60-90% |
| caveman | Terse agent prose, ~65% fewer tokens |
| ponytail | Makes the agent think like the laziest senior dev in the room |
| codebase-memory-mcp | Indexes the repo into a knowledge graph, so the agent asks instead of grepping |
| thermo-nuclear | Maintainability review skill |
| mattpocock-skills | Engineering skills; radin delegates interviewing and research here |
| unslop | Rewrites prose to strip AI-slop patterns |
| technical-writing | Technical-writing rules for docs and agent-facing prose |
unslop and technical-writing are fetched from cursor/plugins main on every
install and update, unpinned. An upstream edit changes agent behavior without
review.
Two caveats. A graph answer is a pointer, never proof of absence, so read the
file before editing. After codebase-memory-mcp update, run
radin repair, because that update reruns a config write that
drops other tools' hooks. radin snapshots ~/.claude/settings.json and
~/.claude.json into ~/.claude/.radin/backups/ before that write and never
deletes a snapshot; codebase-memory-mcp uninstall undoes upstream's own side.
Full list in
docs/DECISIONS.md.
Hacking on radin itself? AGENTS.md, CONTRIBUTING.md.