How Theo uses skills to review and ship T3 Code
A look at Theo’s everyday workflow: review a change with Sol Loop, let Full Send carry it through merge, or use Full Stop to keep the final decision.
Getting an agent to write code is one part of shipping a change. Someone still needs to check the result, respond to review, deal with failing checks, and decide whether it can merge.
In These 3 skills changed my life, Theo explains how he turned instructions he kept repeating while working on T3 Code into three reusable skills. Each handles a different part of getting a pull request finished.
This article describes his published workflow. We have checked the skill files, but have not independently run or measured the setup.
Review the change with Sol Loop
Sol Loop gives the agent implementing a change a separate reviewer running GPT Sol. The implementing agent checks the feedback, makes focused fixes, and can disagree when it has evidence that a finding is wrong.
The review has a budget. By default, it runs for up to three rounds. The first covers the whole change; later rounds concentrate on changes and unresolved findings. There is only one reviewer at a time, and cosmetic suggestions alone should not keep the loop running.
The practical detail is the stopping condition: exhausting the review budget does not count as approval. Unresolved disagreements are brought back to the user.
Watch the explanation at 4:40 · Read the Sol Loop skill.
Let Full Send finish the job
Full Send gives the agent responsibility for taking a change through a merged pull request when the user explicitly requests it. Its workflow includes implementation, direct verification, Sol Loop review, opening the PR, and handling CI failures and review feedback.
Verification needs observable evidence. For a bug, that can mean reproducing the failure and seeing the same reproduction pass after the fix. For an interface change, it means running the app and looking at the result.
Before merging, the skill requires passing checks, completed review, evidence covering the change, and a branch that is ready to merge. If the agent cannot meet those conditions, it needs to raise the blocker. The permission to merge does not automatically extend to production changes or deleting data.
Watch Full Send at 15:23 · Read the Full Send skill.
Keep the final decision with Full Stop
Full Stop carries out the same preparation but leaves the PR open for a person to merge. The agent continues fixing, verifying, and handling review until it can give a useful handoff.
Its instructions also cover automatic merging: the agent must prevent the PR from merging through auto-merge or a merge queue. At handoff, it distinguishes a change that is ready from one waiting for required approval or blocked on something else.
This gives the work a clear endpoint: a reviewed change and the evidence needed to make the final decision.
Watch Full Stop at 21:06 · Read the Full Stop skill.
What it looks like in practice
Near the end of the video, Theo describes taking a contributor PR for a Windows Explorer fix, asking an agent to take it over, and combining that with Full Send. He reports that it produced and merged a replacement while preserving the contributor’s credit. Watch the example at 26:57.
The useful pattern is that the instructions cover the work around the code: review, proof, follow-up, and the point where a person needs to step in. The skills make those expectations reusable across tasks.
Theo reports less review back-and-forth and faster merges. Those are observations from his own workflow, rather than independently measured results for other teams.
Try one part of the workflow
Our suggested starting point is a small change you already understand. Read the instructions, adapt them to your repository, and compare the resulting PR with the way you usually work. Full Stop gives you a chance to inspect the evidence before deciding to merge.
The tstack repository contains the skills and installation instructions. Its README lists support for Claude Code and Codex, an authenticated GitHub CLI, and Python for the PR-watching script. Sol Loop also needs access to the reviewer model; T3 Code provides a delegation route for that. Install the supporting skills listed in the README as well.
Source notes checked October 28, 2026. The skill links above point to the version reviewed for this article; the repository may have changed since publication.