| name | rtk-skill |
| description | Use when running shell commands that produce verbose output (git, test, build, lint, package managers, docker). Prefix with rtk to save 60-90% tokens. MindMesh projects use explicit rtk prefix (no global hook). |
| version | 1.1.0 |
RTK Skill (Rust Token Killer)
RTK is a CLI proxy that filters and compresses command output before it reaches the LLM (typically 60–90% token savings).
MindMesh setup (this repo)
MindMesh integrates RTK via AGENTS.md + this skill — we do not run rtk init / global hooks.
You must prefix commands explicitly:
rtk <original-command-and-args>
Resolve binary (try in order):
rtk on PATH (common in dev environments)
bunx rtk (project-local via @mind-mesh/rtk)
./node_modules/.bin/rtk
Verify:
rtk gain
rtk --version
Golden rule
For any shell command that prints more than a few lines, run it as rtk <cmd> instead of bare <cmd>.
Applies to: git, tests, builds, linters, package managers, docker/kubectl, ls/grep/find, gh, etc.
Critical: built-in Read / Grep / Glob
Claude Code, Cursor, and similar agents often have native Read/Grep tools that bypass Bash hooks.
Those tools do not auto-rewrite to RTK. For token-efficient file/search workflows, use:
| Instead of native tool | Use |
|---|
| Read large source file | rtk read path/to/file.rs |
| Read signatures only | rtk read path/to/file.rs -l aggressive |
| Grep / search repo | rtk grep "pattern" . or rtk rg "pattern" |
| Find files | rtk find "*.ts" . |
| List directory | rtk ls . |
Exception: .mind-mesh/agent/context.md and short knowledge/*.md — read directly (already dense).
Command reference (by category)
Git (high savings)
rtk git status
rtk git log -n 20 --oneline
rtk git diff
rtk git diff --staged
rtk add -A && rtk git commit -m "msg"
rtk git push
rtk git pull
Tests (failures-focused, ~90% savings)
rtk cargo test
rtk test cargo test
rtk bun test
rtk vitest
rtk jest
rtk pytest
rtk go test
Build & lint
rtk cargo build
rtk cargo clippy
rtk tsc
rtk eslint .
rtk ruff check
rtk next build
Package managers
rtk pnpm list
rtk bun install
rtk pip list
Files & search
rtk ls src/
rtk read src/lib/foo.rs
rtk read src/lib/foo.rs -l aggressive
rtk grep "fn handle_" .
rtk find "*.svelte" .
rtk diff file1 file2
Containers / cloud (when used)
rtk docker ps
rtk docker logs <container>
rtk kubectl get pods
rtk gh pr list
Errors only
rtk err npm run build
Global flags
rtk -u git status
rtk -v cargo test
When command fails
RTK may collapse output but preserves exit codes. On failure, look for a tee path in output:
FAILED: 2/15 tests
[full output: ~/.local/share/rtk/tee/....log]
Read that log if you need the full unfiltered output — do not blindly re-run the same verbose command.
Bypass RTK (rare)
RTK_DISABLED=1 git status
When NOT to use RTK
| Use RTK | Use other MindMesh skills instead |
|---|
| Shell command output | Architecture → mind-mesh-knowledge-skill / context.md |
| git test build lint | Symbol relations → codegraph-skill |
rtk read for code slices | Structured repomix index → repomix-context-skill |
Workflow order: MindMesh knowledge → codegraph → repomix slices → rtk for remaining shell work.
Analytics (optional)
rtk gain
rtk gain --history
rtk discover
Do not
- Run
rtk init or rtk init -g — MindMesh owns agent guidance via AGENTS.md
- Assume hooks rewrite commands — you must type
rtk prefix
- Re-run identical verbose commands after RTK already summarized them
- Use wrong
rtk package (verify with rtk gain)
Complements
- mind-mesh-knowledge-skill — what project knowledge to read first
- codegraph-skill — AST/call-graph queries
- repomix-context-skill — grep repomix pack for source sections