Sendable Studios TGL · The full workflow
You bring the idea. TGL gives your Muse agent a way to ask better questions, agree on the work, and follow it through.
Install TGLPlan Mode
Before code, there are choices. TGL makes room for them.
Measure twice.Your agent reads the available project context, then asks one question at a time with a recommended option. You decide who the build serves, what it should do, and what can wait.
SPEC.md describes what you’re building and why. It records
the goal, scope, and expectations without jumping into implementation
code.
PLAN.md breaks the work into slices. Each names the files,
interfaces, tests, commands, expected results, and what the review should
focus on.
Read both documents. Ask for changes. Your agent records your approval against the agreed version and date before moving into the build.
A workflow example · A personal bookmark app
These short excerpts show the role of each document. Your agent writes the full versions around your project.
Goal
Save links, add tags, and find them
later.
First version
One user. A saved-link
list, tags, and search.
Out of scope
Shared accounts, team
permissions, and a browser extension.
Success looks like
You can save a link,
close the app, return, and find it by its title or tag.
Work
Build the link form, storage, and
saved-link list.
Check
Add a valid URL. Confirm it
appears. Reload and confirm it remains.
Edge case
An invalid link gets a useful
message.
Review focus
The happy path works, saved
data persists, and keyboard use is clear.
Attack Mode
Your approved documents guide the work. Each piece gets a fresh set of hands and a review.
Then make it real.A fresh subagent takes on a slice of the approved plan. The lead agent coordinates the work and reviews it against the spec before moving on.
At each slice boundary, you get a check-in: what changed, what was checked, and what needs attention. A failed check calls for a fix, not a quiet move to the next step.
A fresh red-team agent tries to break the build the way a real user might. It looks for gaps across the whole experience, including the edges between slices.
Review the result and any remaining issues. Muse uses its available tools and permissions; the TGL skill does not grant access or publish on its own.
TGL tells your agent how to work. Your approval and your project’s permissions still matter.
A few useful answers
Want to read the instructions your agent follows? They’re all in the open.
Read the skill on GitHubTGL stands for Touch Grass Later. It is a Muse-native skill for planning and building software with AI agents. The skill gives your agent a repeatable workflow: clarify the goal, write a spec and plan, get your approval, build in reviewed slices, and test the result.
The TGL skill is free, open source, and MIT licensed. You can read, use, and adapt its source. The hosted connector currently has no TGL fee or account requirement. Muse has its own account and usage requirements.
TGL is written for Muse. The skill is a set of instructions that your Muse agent reads and follows, using the tools and permissions available in its environment. Support for other agents has not been verified.
You review SPEC.md and PLAN.md, request changes, and explicitly approve the agreed versions before asking Muse to build. TGL instructs your agent to record that approval and wait for it. It is a workflow rule, not a technical lock on every action an AI agent could take.
No. The hosted connector supports Plan Mode and provides skill files. To use Attack Mode, save your spec and plan, install the full skill from GitHub into an environment Muse can read, and give those documents to your agent. Muse then performs the build.
Ask your Muse agent to confirm it can read the installed TGL SKILL.md. Then say “TGL, plan” followed by your idea. It should begin with one focused question and clear choices. A planning-style reply alone does not prove the skill was installed.
Your idea is a good place to start