These 3 skills changed my life
Monday, Oct 26 · 31 min · about 10 min to read
Theo-brain watched all 31 min
Theo walks through his main skills for coding agents: Sol Loop, where a second model reviews the code in a loop; Full Send, where the agent works through to merge without bothering him; and Full Stop, which does everything but the merge. The lesson is to turn what you keep telling agents into skills, based on what you trust them with.
Skim itSkim it for the parts that match how you direct agents, since the notes already cover the main ideas.
Tap any time to jump the video there.
Key takeaways
His best skills were not clever ideas planned up front. They are things he caught himself telling agents over and over, written down once so he never has to repeat them.
decided to force into my work. It's actually the opposite. It's behaviors I found myself doing again and again that I turned into skills. When I'm working
Having a second model review the work before a pull request (a proposed code change) goes up makes the request slower to open but much faster to get approved and merged, with fewer issues.
this skill, that the PR that goes up has fewer issues, is more likely to be correct, and ends up getting through review faster, and being merged faster. It does take longer for when I send the prompt for the PR to open, but it takes a lot less time for that PR to get
Theo queried his own usage logs. Of 32 times he said "full send," 27 changes merged without him ever looking at them again.
I had 32 times where I said full send, and of those, 27 merged without me having to look at them again. Do you
Full Stop, the version that does everything except merge, has quickly become his most-used skill. He now answers incoming pull requests by simply telling an agent "figure it out, full stop."
it's ready to merge. I love this skill. This has quickly become my most used skill to the point where when I was starting stream today, people were sending me PRs they wanted me to look at and pretty much all of them I opened up and just immediately opened with figure it out full stop. This one is actually
His closing lesson: work out what you do and don't trust agents with, and encode that in skills. He sees himself becoming a manager of his own work.
I'm becoming a manager of myself in a way. And in order to get the most out of your agents, it's important to analyze your flow and figure out what you do and don't trust them with, and start encoding that in stuff like skills, so that you can ship more effectively. All
Two skills here are new to you.
Sol Loop: a second model reviews in a loop
- What it is
- Theo's main agent writes the code, then OpenAI's GPT-6.1 Sol reviews it, round after round. The agent makes the smallest changes that satisfy the reviewer and pushes back on bad suggestions.
- Why Theo cares
- He does daily work in Claude Opus 5.5 because it is the better collaborator and writes better code. Sol is more thorough, "the Rottweiler" that won't let go of a problem, so he uses it for reviews.
- How to try it
- Ask your agent to write a skill: have a different model review your changes in a loop, make minimal fixes, and avoid scope creep.
- Set a budget: three rounds, five only for security or data-loss risks, and stop early if only small nitpicks remain.
- Have the pull request footer state how many rounds ran and the reviewer's final verdict, so problems can't be hidden.
The TLDR is that you should use GPT-4 Soul to review your code in a loop. Make the most minimal possible changes to appease the Soul reviewers. Be careful of scope creep and don't be scared to push back. The be careful of scope creep
with Opus 5.5. I found that Opus 5.5 is a much better collaborator, overall smarter, writes way better code, just more pleasant to work with in general. GPT-61 Soul is more thorough, though. It really does feel like the Rottweiler. It grabs onto the problem and doesn't let go, and I really like using it for
Full Send: agent owns the work through merge New
- What it is
- A skill that tells the agent Theo isn't watching. It must finish the change, prove it works, open the pull request, handle reviews and failing checks, and merge on its own.
- Why Theo cares
- Once he and the model agree on the plan, he doesn't want the thread to interrupt him again and make him re-remember everything just to click merge. It should only bother him if it gets stuck.
- How to try it
- Grant permission in writing: Theo found OpenAI models hesitate to merge unless the skill says the user explicitly allowed it.
- Define proof: reproduce the bug, apply the fix, watch the same test pass, and for screen changes run the app and look.
- List merge conditions: all checks pass, review bots' comments are handled, and no human reviewer is still asking for changes.
you stop. You own the work from its current state to a merged PR. If the user did not explicitly request this in some clear way, do not assume full autonomy. If they did, everything else
proof section. Proof is evidence that you observed yourself in the session. Command outputs, reproduction, screenshots. Use the most direct proof that the change allows. Reproduce the original failure, apply the fix, then watch the same reproduction pass. Run
Full Stop: everything but the merge New
- What it is
- Claude rebuilt Full Send with one change. The agent keeps fixing, verifying and handling reviews until the change is ready, but never merges it. Theo makes the final call.
- Why Theo cares
- He loved Full Send so much that he missed it on riskier changes. This lets the agent handle everything quietly, so by the time it bugs him the change is ready to go.
- How to try it
- Copy your Full Send skill and replace the merge step with "stop when ready for me to merge."
- Use it on changes you consider risky, and Full Send on small, well-understood ones.
So full stop is Claude rebuilding my full sent skill with one big change. Own the work until the PR is ready for him to merge. Keep fixing, verifying, and reviewing. Never merge it. This is
Takeover skill with credit
- What it is
- A skill for taking over someone else's pull request. The agent reviews it, decides whether to build on it or treat it only as a reference, and finishes the job while crediting the original author.
- Why Theo cares
- Getting outside contributors to make a change perfect is too much work. He wants someone he trusts to land it, without contributors feeling their ideas were cloned.
- How to try it
- Paste a pull request link into your agent and say "take this over," adding Full Send or Full Stop.
- Tell the skill to credit the original author and close the old request if it's no longer needed.
This is very important to me cuz I don't want people feeling like I am just closing their PRs and cloning their ideas. I want the people who make those original PRs to get credit. It's just
Advice
Write a skill's description as a trigger, not a summary. It tells the model when to use the skill. His are as simple as "use when the user says Sol Loop."
descriptions. Descriptions are not what they sound like. This shouldn't describe what the skill is. This is an indicator to the model on when it should be used. So, most of my descriptions are really simple like this. They are effectively triggers. I'm describing when the model
When an agent does something you don't like, ask it why, then update the skill to prevent it. The permission rules in his Full Send skill exist because OpenAI models were doing things he didn't like.
malleable. If I notice things not working how I want, I ask the model, "Why did you do that?" And it says, "I thought this because of this." And then I go update the skill to prevent the thing I didn't like. So, these bullet points here are all because OpenAI models were doing things I didn't like.
Tell autonomous agents to ask questions through the tool's built-in question feature. A question buried in a normal message goes unseen and the thread looks finished. Allow asking only when truly blocked.
that it has available to it. A question in a normal message goes unseen and the thread looks done. To ask Theo something, use the harness's question tool so the thread shows that it's waiting. Ask user question in Claude
Only use your Claude subscription inside Claude Code; Anthropic objects to other uses. OpenAI's Codex subscription can be used anywhere, including inside Claude Code.
system prompt. So, you should only use this with Claude Code if you're using your Claude subs. I use my Codex subs with Codex. I use my Claude subs with Claude, but I also use my Codex subs in other things like Claude through this because OpenAI doesn't care where you use your tokens effectively. Anthropic cares a lot. OpenAI doesn't. So, if you
Hot takes
Letting agents merge code you never looked at is fine, if you've built both verification and safety nets.
His reasoning: Agents need ways to check their own work, and you need a way to catch failures before they reach many users. In T3 Code, a test release goes out every six hours, so breakage is fixed before the stable release.
kind of insane to not look, but as I've talked about in other videos, you should build good verification systems so your agents can check if the thing works, but you should also build good safety nets. So, if it does fail, you can catch it
happens. In T3 Code, we have a nightly that goes out every 6 hours, and we get feedback immediately if it breaks something. So, I can fix it before the next stable release goes out. And I know
OpenAI models are the most thorough reviewers, Claude catches odd dangerous issues, and Gemini hallucinates more than it finds.
His reasoning: Models with different training act differently, like three people with different experience reviewing the same change, so each tends to catch unique things.
useful. I personally find that OpenAI models are the most thorough and find things better than others do. Claude models can find like really weird, potentially dangerous things that other stuff will miss. And then Muse and Gemini both can find things here and there, but I find that Gemini hallucinates more than actually finds.
Don't manually invoke skills with slash commands; trigger them with plain phrases instead.
His reasoning: Claude Code won't let you manually invoke several skills at once, and a manual invocation puts your whole prompt into the context twice.
different skills. It's also worth noting that certain harnesses, mostly cloud code, don't let you manually invoke multiple skills at the same time. So, if you disable the model's ability to invoke, you will not be able to trigger multiple at once in certain cases. It's
them information. So when you do invoke a skill manually in Claude code, it's going to put your whole prompt twice in the context. Not ideal. Don't manually invoke, it sucks.
How Theo thought about this
Each skill grew from the last. Repeated instructions became Modernize and Takeover, and hand-run reviews became Sol Loop. Better code made him trust Full Send, and missing that flow on risky changes led to Full Stop.
loop skill. I realized the quality of the code coming out was better than I expected, so I introduced the full send skill to see it all the way to the end. Then I missed that flow a whole bunch, so I made the full stop skill to get as much of it as I could without having to
He decides per change whether to let the agent merge or not. With experience you learn what models can and can't verify, and what needs another look.
done enough of this type of like agentic coding, you know what the models can and can't do, what they can and can't verify, what needs an extra look, and what doesn't. And once you've built that intuition, it's relatively simple to know this thing can probably merge itself, this thing probably needs another look. But I love this workflow
Sign in to mark what you've watched and keep notes.
Have a question about this one? Ask about this video →
Theo Daily is unofficial and not affiliated with Theo or t3.gg. Written by Claude Opus 5.5 from each video's transcript.
