Skip to content

Letting coding agents finish the job, without letting them ship alone

What we took from Theo Browne's tstack skills: proof before a PR, a capped review by a second model, and a merge gate a person still owns.

Sections
  1. The gap
  2. What a skill is
  3. Three ideas worth copying
  4. What we changed
  5. What we do not know yet

The gap

Coding agents are good at writing a change and poor at finishing one. They call a task done without showing it works. They ask a question at the end of a long message, and the thread looks finished. Each review round adds one more cheap-looking fix, and the PR grows. Someone still has to watch CI and the review bots.

On 27 October 2026 Theo Browne, who builds T3 Code, published the agent skills he uses every day as tstack. We read them the same day and installed most of them.

What a skill is

A skill is a Markdown file that an agent such as Claude Code or Codex loads when its description matches the request. Theo writes the descriptions as trigger phrases, not documentation. The workflow loads only when someone says "full send", "full stop" or "sol loop", so nothing runs unasked.

SkillSayWhat it does
full-send"full send"Proves the change, reviews it, files the PR, watches it and merges it
full-stop"full stop"The same, but stops with the PR ready for a person to merge
sol-loop"sol loop"A second model reviews the change in rounds until both agree
council"council"Three models from three providers review and score the work
babysit-pr"babysit"Watches CI and review bots, and fixes the real findings
takeover"take over"Finishes someone else's PR and keeps their credit

In the video he says 27 of his last 32 full-send runs merged without him opening the thread again. That figure comes from his own setup, which ships a release every six hours and can roll one back.

Three ideas worth copying

Proof is something the agent watched

The skill defines proof as evidence the agent saw in the same session: command output, a reproduction, a screenshot. Reproduce the failure, apply the fix, then watch the same reproduction pass. Run the checks CI runs. Add one focused test that fails without the fix. Its plainest line: "Green mocked tests do not prove that the user experience works."

Review by a different model, on a budget

Claude writes the code and OpenAI's GPT Sol reviews it. Theo's reason is that models trained differently miss different things. The loop has hard limits:

sol-loop/SKILL.mdthe budget, as published
- Default: 3 rounds total. Round 1 is full. Later rounds check only
  changed code and open findings.
- Extend to 5 only for security, permissions, auth, data loss, or
  state/concurrency, while a blocker or should stays open with evidence.
- Stop early if a round returns only nits.
- One reviewer only. Sol must not start subagents or reviewers.
- Keep a ledger: id, severity, outcome, usage if returned. Never guess it.

The writer may push back with evidence, and it declines nits in one line. The aim is the smallest change that meets the intent. Review should make the change correct, not bigger.

The merge gate is written down

A PR is ready only when all of these hold on the latest commit: the proof covers every change, required checks are green, review bots have finished and every finding is fixed or dismissed with a reason, no human review still asks for changes, and the branch is up to date and mergeable. No admin bypasses.

What we changed

The tstack README suggests keeping personal settings in the repo's agent instructions, so the skill files can be updated without losing them. We installed the skills unmodified and wrote our rules beside them.

AGENTS.mdour settings, trimmed
- Full send is allowed in every app repo, but only when asked for by
  name. Without it, the agent stops at a PR that is ready to merge.
- The testing permission does not override how we test: no production
  resources, builds on the shared CI runners, scoped checks while
  working and the full suite before merging to main.
- Product behaviour and merge-conflict choices go to a person, through
  the question tool, with the recommended option first.

Three reasons sit behind those lines.

  1. Merging is asked for, never assumed. On our apps a merge to main deploys, so the agent merges only when we say "full send" and the gate passes. Any other request stops at a PR that is ready for us to merge.
  2. Permission to test is not permission to test anywhere. The skill lets the agent run the app, a browser or a simulator without asking. We keep that, but tests never touch live data, and heavy builds queue on our shared runners instead of one laptop.
  3. Decisions stay with people. The agent decides how to build what was asked. What the product should do, and which side wins a conflict, is still a question for us.

What we do not know yet

We installed these on 27 October 2026, so we have no numbers of our own. Over the next month we will count how many PRs reach "ready to merge" without a nudge, how many of the reviewer's findings were real bugs, how many rounds each PR takes, and whether anything had to be rolled back. We will add those numbers here.

Want an app built with proof at every step?

We are bringing these review and merge gates to the apps we build and run.