Your prompts kinda suck
Sunday, Oct 18 · 32 min · about 9 min to read
Theo-brain watched all 32 min
Theo argues that engineers prompt badly because they over-engineer, not because they under-engineer. Many "features" are really just information you can get by asking your agent for the outcome. Engineering effort should go into the pipes and data access that let prompts do more. His own support-triage tool shrank from thousands of planned lines of code to one prompt.
Skim itThe notes carry the argument; watch the triage walkthrough and the email segment if you want to see the prompt and setup on screen.
Tap any time to jump the video there.
Key takeaways
Many things people want built as features are really just information. If an agent can get that information when you ask, it doesn't need to be a button in the product. Ask for the outcome instead.
really cool. I'm talking about features that aren't actually features. They're information. And if you can get that information other ways, then it doesn't need to be a feature. It should just be exposed in the ways that you need to do these cool things. And now I'm going to
Asking surfaces things you never knew to look for. A one-line request for a breakdown of his agents' tool calls (the actions an agent takes, like running a search) showed one home-directory search had run for 41.7 minutes.
Apparently, one home directory search lasted 41.7 minutes. Another lasted almost 10. Search the repository or known directory. Like this is useful info that I didn't have cuz I didn't ask this model in this way. But see it like
Theo asked for a picture of his context window (everything the model reads with each request). It revealed that his proxy setup on Railway had been adding 70,000 tokens of junk to every request.
example is going to work because I actually used it for a different video I was filming yesterday and that's how I learned due to the way my proxy is set up, Railway was dumping 70,000 tokens of context into every single request I
He is not prompting instead of building. He prompts for things he never would have bothered to build. Then he puts his building effort into the primitives, tooling, pipelines and data access his agents need so his prompts can do more.
to them. Similar with this. It's not that I am choosing to prompt instead of building, it's that I'm choosing to prompt the things I never would have bothered building. And then when I
realize how powerful that is, I go one step further where I focus my building on creating the primitives and creating the workflows, the tooling, the pipelines, and the data pipes and access that my agents need so that my prompts
Theo reads T3 Code's feature requests, issues and DMs. He says at least half of them are things users could get just by prompting the agent, without changing the product.
cannot tell you how often I'm scrolling through the feature requests for T3 code or the issues we have on the repo or seeing the things people DM me that they're requesting. At least half of it is stuff that they could prompt for themselves. Not to extend T3 code, but just to use it. And this idea that a
Three skills here are new to you.
Support triage as one prompt handed to an agent New
- What it is
- A single command passes a written playbook to Claude Code or Codex. The agent asks what went wrong, gathers reproduction steps, checks the source code, fixes the problem or files a clean bug report, and removes any secrets first.
- Why Theo cares
- He started building an elaborate multi-step flow. He then realized the models can run that flow themselves. It went from thousands of planned lines to five lines of code plus 30 lines of markdown, and he says it works far better.
- How to try it
- Write down, in plain steps, how you want a user's problem diagnosed and reported.
- Have your coding agent turn it into one prompt that starts by asking the user to describe the problem and share screenshots.
- Include rules: strip secrets before filing, and comment on existing issues instead of opening duplicates.
their issue. This feature when I first started working on it was going to be thousands of lines of complex code. And then I realized it could be five lines of code and then 30 lines of markdown, and now it works way better. I can't
Ask the agent what it can do, then clone it New
- What it is
- Theo asked the Codex desktop app to describe its computer-use abilities in detail. He then pasted that description into Opus in T3 Code and told it to clone as much as it could, and it did.
- Why Theo cares
- He wanted better computer use in T3 Code and knew Codex's was great. He thinks it is very cool that you can just ask an agent what it has access to. If it has less access than another tool, ask it to summarize the difference.
- How to try it
- Ask your agent to describe in detail every capability and tool it has access to.
- Paste that description into another agent and ask it to clone as much of it as it can.
- If one tool can do less than another, ask it to summarize the difference.
myself. And I was able to copy and paste its description of all the things it can do into Opus in T3 code and say, "Clone as much of this as you can." And it did.
plans, but like, do you know how cool it is that you can just ask it what it has access to? And if you notice it has less access than it does in a different tool, ask it to summarize the difference.
Connect your email inboxes to ChatGPT New
- What it is
- ChatGPT lets you link several email accounts at once, so you can ask an agent about your inbox instead of reading it yourself.
- Why Theo cares
- He manages five inboxes and resisted this for a long time. After a hand injury he tried it and went from checking email about 10 times a day to roughly once.
- How to try it
- Link your inboxes to your ChatGPT account.
- Ask it what needs your attention instead of opening the inbox.
- Ask it to watch for a specific email, like a large bill, and dig into why it is high.
I begrudgingly tried the feature where you could link your emails, multiple of them at a time by the way, to my ChatGPT account. I went from checking my inbox 10 times a day to like once-ish a day and now I just ask my agents instead.
Advice
Don't ask for a feature in your tool. Ask your agent for the outcome you want. Theo says this applies most to complaints people have about their agents.
of asking me to build a feature in T3 code, try asking for the outcome you want from your agent. And this applies best for the complaints people have
Don't wait for a menu to pick which sub-agents (helper agents) are used, because the model chooses them. Just say it. Theo's example: two Opus 5.5 helpers explore performance, then a new Grok 4.7 thread implements the findings.
to I don't know 61 Soul on medium fast, I can ask it spin up two Opus 5.5 delegates to explore performance wins in T3 code, summarize their thoughts, and then start a new thread using Grok 4.7
If you keep typing the same instruction, put it in your skills or CLAUDE.md files, the standing instructions your agent reads every time.
And if you find yourself saying the same thing over and over again, that's when you start adjusting your skills and your agent and Claude MD files. Oh look, the
Run slower requests in the background and let them pop up when they're done. How long a task takes stops mattering once you aren't waiting on it.
you get in this mindset of like run the thing in the background and then let it pop up when it's ready, you realize that it doesn't matter that much how long things take as long as you don't expect it or need it immediately. But, the main
Hot takes
The prompt-engineering hype was "crap." Engineers are bad at prompting because they over-engineer their prompts, not because they engineer them too little.
His reasoning: Over-engineering makes them miss the chance to step back and rethink how the computer can work for them.
I'm actually here to say the opposite. I think that engineers suck at prompting, but I don't think it's cuz they're not engineering it enough. I think it's cuz they're over-engineering it, and they're not taking the opportunity that prompting offers us to take a step back and rethink how the computer can work
Developers should get out of their terminals. Theo, once a big terminal person, says terminal lovers are "just wrong."
His reasoning: He dictates whole paragraphs by voice, and in a terminal he can't read what he just said.
I do not get how people live in their terminals. I just voice to text it a paragraph and I can't read it. Get out of your [ __ ] terminals, guys. Get over yourselves. You're just wrong. It like I I was such a big terminal person.
Managing your context window yourself barely matters anymore.
His reasoning: He had 70,000 extra tokens at the front of every prompt for two months and never noticed.
to think about these things. And I think the fact that I had a 70k token context thing at the front of all my prompts for the last 2 months and didn't even notice shows how little managing your context yourself matters at this point.
Developers who let AI write code aren't brain-rotted. Theo says he is engineering harder than ever, now on things like his own JavaScript runtime.
His reasoning: Problems that once weren't worth solving now matter more and are within reach of one person. So the problems he works on are more complex.
No. I'm building my own JavaScript runtime for my own cloud. I'm engineering harder than I've ever engineered and my brain is working harder than it's ever worked. If you see
How Theo thought about this
Theo was building a T3 Code dashboard to manage all running agent threads, and kept adding a button for every action someone might want. Then he saw he could instead make it a chat: a meta thread connected to all his T3 Code instances that just does whatever is asked.
kept adding more and more things to it. I kept adding new buttons for different things you might want to do, a button that would auto regen all of the titles,
you would want to use this, or I could make this a chat instead. I could have a meta thread in T3 code that connects to all of your T3 code instances on all of your environments and let's you just ask for the thing you want. Want a nice
For triage, he reasoned that a fixed flow only works if you can predict every failure. If he knew every way T3 Code could break, he would have fixed them all already. Because he can't, an open-ended agent handles the unknown cases better than a fixed feature.
supports everything they might need. If I knew all the ways T3 could fail, then I wouldn't need this feature at all cuz I would have just fixed it all. The reality is I don't. T3 code can and will
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.
