Plan Mode measures twice. Attack Mode cuts once.
TGL is a developer harness for Muse agents: an AI-native, spec-driven way to turn an AI coding agent from improviser into a disciplined build team. Agents are powerful but improvisational: brilliant one turn, sloppy the next. TGL makes disciplined development the default.
Point your Muse agent at this repo and say "run TGL" (or load
SKILL.md into its context). Then:
- "TGL, plan X" enters Plan Mode: your agent grills you one question at a time, writes a spec and a sliced plan, and waits for your sign-off. No code yet.
- "go attack mode" launches Attack Mode: your agent executes the signed plan one slice at a time, reviews each slice against the spec, red-teams the build, and ships.
Plan Mode: the war room. Recon (one question at a time, always as tappable choices, and your agent reads the project memory first so it never re-asks a settled question), SPEC.md (what and why, no code), PLAN.md (numbered slices with exact files, interfaces, tests, commands, expected output, and a Review Focus per slice). Nothing is built until you sign the plan.
Attack Mode: the assault. One subagent per slice, each with fresh context. The commanding agent reviews each slice against the spec before the next begins, and sends you a SITREP at every boundary so the build never goes quiet. Then a fresh agent plays red team and attacks the build the way a real user would. Red team creed: break it before they do.
The chain of command between them: Attack Mode does not march without signed orders. No approved plan, no assault. It is the only enforcement mechanism, and it is absolute.
| Without TGL | With TGL |
|---|---|
| One-shot builds, silent for an hour | Sliced execution with a SITREP at every boundary |
| "It works on my machine" | Verified in the real surface at real size |
| Bugs found by users | Bugs found by your own red team first |
| Context lost between sessions | LEDGER.md remembers every decision |
Two ways in, pick whichever feels right.
Conversational (the Muse-native way). Point your Muse agent at this repo and say "install TGL". It reads the skill, walks you through setup step by step, explains what each mode does, and runs the first-run tutorial with you. No terminal required.
TUI installer. Run python3 install/tgl-install.py in a terminal. It is
stdlib-only Python, so it works over the plainest SSH connection:
- Shows you what TGL is and how the two modes work
- Checks the environment, then copies the skill into your Muse skills folder
- Verifies the install by reading it back (never trust a build report)
- Offers the guided first run
Every step has a help tooltip: press h at any prompt.
Flags: --yes to skip pauses, --no-color for plain output,
--target DIR to choose the install location.
After either path, two phrases run everything: "TGL, plan X" and "go attack mode."
- docs/tutorial.md: guided first run on a tiny example project. Shows exactly what you type and see, including what a SITREP looks like when a build checks in.
- docs/modes.md: Plan Mode vs Attack Mode deep dive, the chain-of-command gate, and what red-team verification catches.
SKILL.md: the harness itself. This is the product.install/tgl-install.py: the TUI installer.templates/: SPEC, PLAN, and LEDGER templates plus the release checklist.docs/: the tutorial and the modes deep dive.assets/branding/: logo, repo banner, and mode diagrams.
TGL is not one person's workflow written down. The stranger test is a standing rule: if something only works on one person's machine, it does not ship. No hardcoded identity, no assumed tooling, no assumed memory system. Built connector-ready: clean packaging, separable concerns.
Muse agents needed the disciplines that already work elsewhere: GSD's phase isolation, Superpowers' spec-before-code, sliced review, and persistent ledger. Rebuilt for how agents actually work: subagents instead of git worktrees, the spec file instead of hooks, real-surface verification instead of build reports, SITREPs instead of silence.
In private dogfood now, getting bloodied on real builds.
MIT. See LICENSE.