ninja nerdingPrompt like Theo on Opus 5.5

Prompt like Theo on Opus 5.5

Theo read through Anthropic's Opus 5.5 guide on camera, and showed his own work while doing it: two real prompts, the questions he asks every day, and his CLAUDE.md, edited live.

Jump to his CLAUDE.md, or start with the formula behind every prompt.

Original sources

Theo · Getting the most out of Opus 5.5 youtu.be Anthropic · Getting the most out of Opus 5.5 claude.dev Theo · Opus 5.5 livestream youtu.be Theo · Wow. youtu.be

Paper-friendly: no video, no buttons, sources spelled out.

fakenerd.ai/theo-opus-5-5 · Video: youtu.be/ejjBbaq9RmY · Timestamps are minutes into the video.

The video

28 minutes, posted Sept 24, 2026. Theo reads Anthropic's guide (by Addy Osmani) and reacts. Every timestamp on this page opens the video at that moment.

2:28 The guide starts: three things to try first
4:14 A 6.5-hour Max run, and the handoff prompt
9:36 It stopped at phase 0: the follow-up prompt
10:40 Say what done looks like
11:43 Stop telling it to think hard. Max vs XHigh
14:54 Steering a running task
16:19 Design: name what you don't want
18:19 Tell it which stops you want
18:59 Editing his CLAUDE.md on screen
20:19 Subagents
21:09 Checking results: his risk questions
22:30 Cross-lab code review
24:39 Mark what it can't confirm
25:53 Flagged messages and “show your reasoning”
28:38 “Do not touch max mode”

Nine signals

The whole video in nine boxes. Each links to the part of this page that explains it.

0110:40
Every prompt has three parts
What you want, what done means, and an out. Theo calls handing over the whole task at once “prompting wider”.
How
025:42
Say what you don't want
His handoff prompt rules out the thing that would bother him most. “Being explicit about what you don't want is even better.”
How
035:42
Give it an out
Models are trained not to give up, so they answer their own questions and push on. One line that says when to stop and ask fixes that.
How
049:50
When it stops early, name the stop
His run did phase 0 of 7 and stopped. His reply named that stop, ruled it out, and set a finish line he could click.
How
0511:43
Stop telling it to think hard
It already thinks before every reply. Change the effort level instead of adding “think hard” lines.
How
0612:57
High or XHigh, never Max
Max takes away its ability to think less. On his Skatebench runs it scored 79% to XHigh's 78%, for 13 times the cost.
How
0718:59
Put standing permission in CLAUDE.md
His Approvals section says yes to whatever the task needs. He pasted Anthropic's stop rule onto the end of it, on camera.
How
0821:57
Ask about risk before you read the diff
“What are the risks of merging this code today?” is his most common prompt. It tells him whether a big change needs reading line by line.
How
0920:52
Say subagents are allowed
Opus 5.5 is shy about starting helper agents. He tacks a line onto prompts that says it may use them however it chooses.
How

Words on this page

Thirteen terms, in plain words. Skip this if you already prompt coding agents every day.

Opus 5.5: Anthropic's AI model, released Sept 22, 2026. The guide Theo reads is about how to prompt it.
Claude Code, T3 Code: Claude Code is Anthropic's coding agent. T3 Code is the app Theo watches his agent threads in.
Thread: One conversation with an agent, working on one task. One of Theo's ran for 6.5 hours.
Subagent: A helper agent the main agent starts to take on part of the work. Opus 5.5 is shy about using them unless told it may.
CLAUDE.md: A file of standing instructions for Claude, kept in a repo. Theo's covers pull requests, model choice, subagents and approvals.
Fleet: Theo's set of computers (Leftbook is one), all run from one private repo of the same name.
Environment variables: Settings and secrets a project needs to run, kept outside the code. Theo names them in his prompt so the agent copies them without asking.
Effort level: How much the model may think before it answers: Low, High, XHigh or Max. Each is a cap, except Max, which also stops it thinking less.
Token: The unit AI models count and charge in, roughly a piece of a word. More thinking means more tokens.
Skatebench: Theo's own test for AI models. This page uses his Opus 5.5 runs to compare XHigh with Max.
PR: Pull request. A proposed code change, waiting to be reviewed and merged.
Review bots: Automated checkers that comment on a PR. Draft PRs don't get them, so his CLAUDE.md says to open real ones.
Babysit a PR: Keep watching a PR: check what the review bots found, fix the real problems, stay quiet when nothing is new, stop when they pass.

One prompt, and what goes around it

Top row: the parts of one prompt. Bottom row: what Theo does before, during and after. Click any box to see what it is, his wording, and a fill-in version to copy. Or press Start and follow one real run.

One prompt · darker boxes are the formula

Around the prompt · before, always, during, after

Follow one real run

Six beats, using the run Theo shows: moving a 6.5-hour thread to another computer, then getting it to finish. Watch which boxes light up.

One prompt · part 1 of 4 · In the prompt

Watch at 5:42

What you want

The whole job in one message, said plainly, so the agent can plan the full run instead of waiting for the next step. Then add what you don't want. Theo thinks that line does even more.

His wording · Typed · handoff prompt

“My goal here is to do the whole implementation on that machine in one pass […] I don't want you to be spamming my machine with computer use requests and browser control while I'm using it.”

Fill it in · FakeNerd's version

I want you to [the whole task]. My goal is to [the outcome] in one pass.
I don't want you to [the thing that would get in your way or break something].

What you want, what done means, and an out

Theo's read of the guide's first example: “the perfect simple one, two, three.” He calls handing over the whole task “prompting wider”.

What you want

The whole task in one message. Also what you don't want: “being explicit about what you don't want is even better.”

What done means

A finish line it can check. Built, screenshot taken, PR up. Top models “keep going until they get there.”

The out

When it may stop and ask. Without one, it answers its own questions and plows on, because it was trained never to give up.

The guide's example, which Theo holds up

Migrate the payment endpoints from the old client to the new one.
Done means: every endpoint uses the new client, the old client is
deleted, and the test suite passes.
Stop and ask me only if a test fails for a reason you can't explain.

His prompts

Both copied off the screen. Left: what Theo typed. Right: why each line is there, mostly in his own words.

1 · Move a run to another machine

5:42

A thread had run 6.5 hours on his laptop on Max. He wanted it on Leftbook, another computer in his fleet, without doing the SSH, cloning and env-var copying by hand. It took about 17 minutes. “This is the prompt that I think does a good job of giving the agent an out.”

01
I want you to kick off an agent run on my computer Leftbook. You should be able to see how to connect through the fleet repo on this machine.
Points at the known way in, so it doesn't try anything sketchy to connect or trip a security guard.
02
I want you to make sure: 1. This repo is cloned. 2. All the needed environment variables are there. 3. You can run Opus 5.5 through Claude code in T3 Code, such that I can still see the thread how I normally do in T3 Code.
“Make sure” means do it and check it worked. Naming the env vars up front is the permission: it won't stop to ask before copying them.
03
My goal here is to do the whole implementation on that machine in one pass, with however many subagents and whatever else are needed, with full computer use capability on the other machine.
The goal, said plainly. Also a green light for subagents, which Opus 5.5 is shy about using.
04
I don't want you to be spamming my machine with computer use requests and browser control while I'm using it.
What he doesn't want. “Being explicit about what you want is great, but being explicit about what you don't want is even better.”
05
If anything does not work as expected when you move this workload over to Leftbook, don't hesitate to ask questions so we can get it working right.
The out. Models are trained hard not to give up, so they answer their own questions and push on. This line makes stopping to ask allowed.

2 · When it stops too early

9:50

The new run did phase 0 of 7 and stopped. His reply names that exact stop and rules it out. “Now I do not have to look at this thread again for at least an hour or two.”

01
I want you to hold off on all questions until the very end, when you get through to phase 7 in completion. I don't want you to stop after every phase to ask for permission to continue. I want you to do the whole thing.
Written after the run did phase 0 and stopped. Names the stop it made and forbids it.
02
Continue working on this port until the entirety of it is completed, and don't inform me until you have a Tailscale link I can click on that is the complete rewrite so I can test it and make sure everything works.
A finish line you can click. Done isn't “I think it works”, it's a link Theo can open.
03
Don't worry about how it's hosted. Don't worry about how it's built. Just worry about completing the work as assigned.
Takes side quests off the table.

Phrases he reuses

Questions Theo asks all the time, and when.

“What are the risks of merging this code today?”
“Probably my most common prompt.” Asked after a long run leaves a PR too big to want to read.
21:57
“How much risk do you perceive with this change? What's the worst thing that would happen if we merge this right now?”
Same job: decide if the diff is worth reading line by line.
22:16
“What's the status of this work? What do you need from me?”
Coming back to an old thread he no longer remembers.
21:48
“What's the state of the work?”
Checking a run that went on too long. Just a question, not “continue”.
5:02
“continue”
When a run stops with “Want me to continue?”. The CLAUDE.md rule below makes this rarer.
19:46
“Use subagents and workflows however you choose. I trust your judgment there.”
Tacked onto the end of prompts. Opus 5.5 is less willing to spin up subagents unless told it can.
20:52
“This sucks. Make it better.”
With a screenshot, an arrow drawn on it, pasted into the chat. “Works way better than you would think.”
18:13
“I just updated the CLAUDE.md in this repo. Commit it and push it out to the rest of the fleet.”
How the CLAUDE.md reaches all his machines. No sync script, the agent does it.
19:28
“Mark the things you can't confirm.”
On reviews. Anything it can't check (other hardware, a secret, a matter of taste) gets flagged for him.
24:39

His CLAUDE.md

On screen at 19:04. Lines 46–74, copied from the video frames. Not published anywhere else: there's no copy on his GitHub. “[…]” marks words hidden behind his webcam.

fleet/corefiles/CLAUDE.mdLines 46–74

46## Pull Requests
47–48(scrolled past, not readable)
49- Make sure titles follow conventions from the repo. They should be simple and easy to understan[…] Conventional commit styles in projects that use them, i.e. "fix(web): new threads no longer spik[…]
50- PR descriptions should aim for simplicity. Open with a minimal, clear description of the probl[…] Follow up with how you solved it.
51- Add a blurb to the end of the PR description about what model and harness is making the change[…]
52- **Open a real PR, not a draft.** Drafts do not get review-bot coverage.
53- **Rebase onto latest `main` before opening.** Stale branches conflict and waste a review round[…]
54- When asked to monitor or babysit a PR: poll checks and comments newer than the last push; veri[…] bot finding against the source before acting on it; fix real ones and dismiss false positives with a written reason; fix CI failures, distinguishing real breaks from known infra flakes. If nothing is new, stay quiet — do not post filler comments. Stop when the repo's review bots are green on the latest commit.
55- Merge only per the disposition given in the request (merge when green, or stop and report). If none was given, report and ask.
56
57## Model selection
58
59The user's request always wins. If the user names a model, use the latest version of that model family for the work they describe. "Use Opus" means the latest Opus. "Use Fable" means the latest Fable. "Use Sonnet" means the latest Sonnet. This applies to the main session, subagents, and any tool that takes a model parameter. The defaults below only apply when the user has not said which model to use.
60
61## Claude Code subagent guidance
62
63Never use Haiku or Sonnet for subagents. Generally speaking, you should use the currently selected model, But if you have a specific reason to use a different model, ask and I will approve.
64
65## Approvals
66
67Generally speaking, I approve of the actions, commands, and tools required to complete the task I requested.
68
69For example, if I ask you to show me an HTML write-up, I expect you to publish that HTML if necessary. If I ask you to create a UI pull request with screenshots, I expect you to upload those screenshots and include them in the description. These are examples, not an exhaustive list. Use your judgment to apply the same principle to similar situations.
70
71Approve the commands and tools needed to complete the task. Ask me first only when there is a real concern about exposing sensitive information or an action goes far beyond what I requested in an irreversible way.
72
73When a step doesn't need my input, keep going. Put status notes in the same message as your next action.
74Stop and ask only when you can't continue without me, or before anything destructive: deleting data, force-pushing, or changing anything outside this repository.

Lines 73–74 were added on camera: Anthropic's stop rule, pasted onto the end of Approvals.

What each section does

Pull Requests 46–55
Short titles in the repo's commit style. PR body: the problem first, then the fix, then which model and harness made it. Real PRs only, never drafts, because drafts skip the review bots. Rebase first. When babysitting a PR, check each bot finding against the code, stay quiet when nothing is new, and merge only if told to.
Model selection 57–59
What he says wins. “Use Opus” always means the newest Opus, so the file never goes stale when a model ships. Applies to subagents and tools too.
Claude Code subagent guidance 61–63
No Haiku or Sonnet subagents. Subagents run on whatever model the main session uses, unless it asks first.
Approvals 65–74
Blanket yes to whatever the task needs, with examples (publish the HTML, upload the screenshots) and “use your judgment” for the rest. Ask only for secrets or big irreversible moves. Lines 73–74 are the Opus 5.5 stop rule, added on stream.

What changed since Opus 5.5 shipped

Opus 5.5 came out Sept 22, 2026. Everything Theo has said about his CLAUDE.md since then:

Sept 23
Deleted his “unslop” rules (the lines that police how the model writes) to hear how Opus 5.5 talks on its own.
“I don't think I'm going to bother re-adding it cuz it's doing a great job.”
“Wow.” · 10:01
Sept 23
Pasted Anthropic's stop rule onto the end of Approvals (lines 73–74), then had Opus 5.5 commit and push it to every machine.
“I'm putting my money where my mouth is. I'm pasting this directly into my CLAUDE.md.”
This video · 18:59
Sept 23
Said no to the guide's other CLAUDE.md idea: ending every run with “Blocked on me / Changed / Found”. Asks for status when needed instead.
“I'm not going to go that far. When I do want this, I usually just ask.”
This video · 21:36
Sept 23
On stream, suggested a one-liner that has the same effect as the stop rule.
“Add to the CLAUDE.md something like: assume you have my permission to keep going.”
Livestream · 2:31:38

Steal this · his Approvals section

## Approvals

Generally speaking, I approve of the actions, commands, and tools required to complete the task I requested.

For example, if I ask you to show me an HTML write-up, I expect you to publish that HTML if necessary. If I ask you to create a UI pull request with screenshots, I expect you to upload those screenshots and include them in the description. These are examples, not an exhaustive list. Use your judgment to apply the same principle to similar situations.

Approve the commands and tools needed to complete the task. Ask me first only when there is a real concern about exposing sensitive information or an action goes far beyond what I requested in an irreversible way.

When a step doesn't need my input, keep going. Put status notes in the same message as your next action.
Stop and ask only when you can't continue without me, or before anything destructive: deleting data, force-pushing, or changing anything outside this repository.

Where it lives

fleet/
├── .claude/
├── .local-backups/
├── bin/
├── bootable-codex/
├── common-tools/
├── corefiles/
│   ├── AGENTS.md
│   ├── CLAUDE-subagent-guidance-202…
│   └── CLAUDE.md        ← the file on screen
├── docs/
├── reports/
├── scratch/
├── scratch-log-audit/
├── skills/
│   ├── claude-only/
│   ├── command-center/
│   ├── universal/
│   └── PROPOSED-GLOBAL-RULES.md
├── snapshots/
├── tests/
├── theo-boxes/
├── vibeproxy/
├── .env
└── .gitignore

One private “fleet” repo runs all his machines. The file sits in corefiles/ next to an AGENTS.md. No sync script: he tells the agent to commit and push it to the fleet.

Leave it on High or XHigh. Never Max.

The levels are caps: “think up to this much.” Max is different: “Max isn't just making it so the model can think more. It is removing its ability to think less.” His Skatebench runs on Opus 5.5:

SkatebenchXHighMax

Avg tokens per answer338~5,000

Avg time per answer6 s50 s

Slowest answer31 s600 s

Score78%79%

One more right answer, for 13× the cost and 15× the tokens. Low “kind of sucks”. Between High and XHigh he can “barely notice the difference.” 12:57

Anthropic's guide vs Theo

Mostly agreement. Where Theo goes further or disagrees, it's the useful bit. Read the full guide, which also covers the task list in a file, fast mode and flagged messages.

Say what done looks like
10:40
Guide
Whole task in one message, name the finish line, then let it cook.
Theo
Agrees, it's what he calls “prompting wider”. His formula: what you want, what done means, and an out.
“Think hard” lines
11:43
Guide
Delete them. It already thinks before every reply.
Theo
Agrees, loudly. “It knows that it should think. It is smarter than you probably think.”
Effort level
12:57
Guide
Change effort instead of asking it to think.
Theo
High or XHigh, barely any difference. Never Max: “Max is a will”, it removes the model's ability to think less.
Add to a running task
14:54
Guide
Type a follow-up while it works.
Theo
Stopped restarting threads. Steering now adds to the list instead of wiping it.
Design
16:19
Guide
List the styles you don't want.
Theo
Agrees, and adds a screenshot with an arrow: “This sucks. Make it better.”
When to stop
18:59
Guide
Put a stop rule in CLAUDE.md.
Theo
Did it, on camera. Lines 73–74 of his file.
Subagents
20:19
Guide
Ask it to split big audits across subagents and check each result.
Theo
Says Opus 5.5 is less willing to use them unless told. Adds “use subagents however you choose” to prompts.
Summary format
21:09
Guide
CLAUDE.md: end every run with Blocked on me / Changed / Found.
Theo
Skips it. Asks “what's the status, what do you need from me?” when he needs to.
Code review
22:30
Guide
Have Opus 5.5 review the diff before a person does.
Theo
Switch labs for review. OpenAI models dig harder. Still: Opus 5.5 found almost twice what Opus 5 did in his test.
Show your reasoning
25:53
Guide
Don't ask for its internal reasoning. Ask “explain why you chose this in three sentences.”
Theo
Annoyed: “what was your reasoning for these changes?” is how he'd naturally ask, and it can trip a flag.

Try it, smallest version first

Six steps on one small task, in any coding agent. The prompt and the last part of the CLAUDE.md are FakeNerd's starters, built on Theo's formula and his file. You can stop after step 3.

01
Pick one small task with a finish line you can see
Something you can check yourself in a minute: a form that should show an error, a page that should load, a test that should pass. Our example below uses a contact form.
02
Set effort to High
In your agent's settings, not in the prompt. Try XHigh only if the result falls short. Leave Max off, and delete any “think hard” lines.
Why not Max
03
Write the prompt in three parts
What you want (and what you don't), what done means, and the out. Copy the starter below and swap in your task.
The parts, one by one
04
Add the CLAUDE.md starter
Paste it into a CLAUDE.md file at the top of your project. Your agent reads it on every run, so you stop repeating yourself.
His full file
05
If it stops early, name the stop
Don't just type “continue”. Say which stop it made, that you don't want it again, and repeat the finish line.
His follow-up prompt
06
Before you keep the change, ask about risk
“What are the risks of merging this code today? Mark the things you can't confirm.” Then check the things it marked.
His other check-in questions

Starter prompt · ours, his formula

I want you to make the contact form show an error when the email box is empty.
I don't want you to change how the form looks or touch any other page.

Done means: submitting with an empty email shows a short message under the box, nothing gets sent, and a filled-in form still sends as before. Show me a screenshot of the error.

If you can't find the form, or fixing it means changing how it sends, stop and ask me. Otherwise keep going.

Line 1 is what you want, line 2 what you don't, then what done means, then the out.

Starter CLAUDE.md · drawn from his

## Approvals
<!-- Theo's lines 67 and 71, word for word -->
Generally speaking, I approve of the actions, commands, and tools required to complete the task I requested.

Approve the commands and tools needed to complete the task. Ask me first only when there is a real concern about exposing sensitive information or an action goes far beyond what I requested in an irreversible way.

<!-- Anthropic's stop rule, which Theo pasted in as lines 73-74 -->
When a step doesn't need my input, keep going. Put status notes in the same message as your next action.
Stop and ask only when you can't continue without me, or before anything destructive: deleting data, force-pushing, or changing anything outside this repository.

## When you finish
<!-- FakeNerd's addition, from Theo's habit of asking it to mark what it can't confirm -->
Tell me how you checked that it works. Mark the things you can't confirm.

Approvals is Theo's wording, minus his examples. The stop rule is Anthropic's, as he pasted it. “When you finish” is ours. The comments say which is which.

FakeNerd's take

Copy the Approvals section first

It's the one part of his file that's whole on screen, and the stop rule at the end is Anthropic's, not his. Paste it, then cut the examples that don't match your work.

The formula beats the prompts

What you want, what done means, and an out. That works in any agent and costs nothing. His two prompts are long because his tasks are, not because long is better.

Never Max rests on one test

Skatebench is Theo's own benchmark, and the gap is 78% to 79%. A good reason to try XHigh first, not proof that Max is wrong for every job.

His file is a snapshot, not a template

Some lines were scrolled past or hidden by his webcam, and there's no published copy. A lot of it assumes his setup: a fleet of machines and review bots on every PR. Beginners can start with Approvals alone.

Read this before token-maxing

This page is how he words one run. His token-maxing page is how he runs many at once. Learn to set a finish line and an out before you multiply threads.

Related on this site: Theo on token-maxing, the same author running ten threads at once. Rauch's workflow, another rule file and another habit of questioning the agent after a run. Agentic benchmarks, for judging models on more than one test.

Attribution

Theo · Getting the most out of Opus 5.5 (Sept 24, 2026). Prompts and CLAUDE.md read from the video; quotes from its captions.
Getting the most out of Opus 5.5, Anthropic's guide by Addy Osmani.
Theo · Opus 5.5, GPT-6 Sol, GPT-6 Luna… livestream (Sept 23), where the CLAUDE.md edit first happened.
Theo · Wow. (Sept 23), on deleting his unslop rules.
t3.gg

The video

Nine signals

Every prompt has three parts

Say what you don't want

Give it an out

When it stops early, name the stop

Stop telling it to think hard

High or XHigh, never Max

Put standing permission in CLAUDE.md

Ask about risk before you read the diff

Say subagents are allowed

Words on this page

One prompt, and what goes around it

What you want

What you want, what done means, and an out

What you want

What done means

The out

His prompts

1 · Move a run to another machine

2 · When it stops too early

Phrases he reuses

His CLAUDE.md

What each section does

What changed since Opus 5.5 shipped

Leave it on High or XHigh. Never Max.

Anthropic's guide vs Theo

Try it, smallest version first

Pick one small task with a finish line you can see

Set effort to High

Write the prompt in three parts

Add the CLAUDE.md starter

If it stops early, name the stop

Before you keep the change, ask about risk

FakeNerd's take

Copy the Approvals section first

The formula beats the prompts

Never Max rests on one test

His file is a snapshot, not a template

Read this before token-maxing

Attribution