What's the right way to write a prompt? It depends how much of the answer you already know
Chris Rupp
Boris Cherny, the creator and head of Claude Code at Anthropic, posted something recently that got a lot of people talking… “I am surprised that people are surprised this is how I prompt Claude. Talk to Claude the way you would a coworker. There’s no secret to prompting.” He added “There’s no need to be overly scaffolded or prescriptive for most tasks… give Claude a goal, and it will figure it out.” And then there are plenty of folks who insist detailed prompts are the answer, with every step spelled out… so which is it?
Honestly… both. Some of the prompts I write for Orbit (the AI orchestration program I’ve been building since back in March) have their own FAILURE CONDITIONS section, and plenty of others are a couple of casual sentences. The difference comes down to how much of the answer you already know, and I made a short video on it too.
No idea yet? Talk it out
If you don’t really know what you want yet, don’t try to write the perfect prompt… just start talking to the agent like a coworker. That conversation IS how the plan gets made… you see something, react to it, change your mind, and the agent keeps up. Then once you’ve got a plan, hand it to a fresh agent and say “go build this,” so all the wandering stays in the first chat and the builder just gets a clean plan (and read it yourself before you hand it off, cuz that’s the cheapest place to catch a bad idea). That’s honestly how Orbit started… at 1:19 one morning back in March I was brainstorming with Claude in chat, and that conversation turned into a spec that became Orbit’s very first build prompt for Claude Code.
Know the goal, but not the details? Let it loop
This is where Boris is dead on. If you know what you want but not how to get there, just say it and let it run. Say you want a little game… you could just tell it “Build a Tetris-style game in space. Sketchy art look. Two or three cool features, surprise me.” No file list, no step-by-step… then you’d have another agent critique it, and keep looping until it’s actually fun (that one’s just a made-up example for the video, I haven’t built it, but it’s exactly how I’d go about it). Boris’s own example works the same way, an interactive companion piece to the Acquired episode on Home Depot with a goal, a quality bar (“aim for New Yorker or nytimes data viz vibes but better”) and a push to keep going (“Make it really good, use lots of tokens and iterate till you’re proud of it”). You can’t write a spec for “fun” or “beautiful”, so you don’t try.
Know exactly what’s wrong? Spell it out
When I know exactly what’s wrong and what the fix should be, I have my agents write it all out… the context, the task, how to check it, and what not to do. Here’s a trimmed version of a real Orbit prompt (file paths and a bunch of lines cut):
Prompt #8072: Moving a prompt to another node is a human approval;
gate it on the sender, not the receiver
complexity: high
## Context
Operator moved a queued prompt from the Mac to the PC via "Run on <node>".
The PC refused it... An earlier change put a project-policy gate on the
RECEIVER. The hole is real, but it is on the SENDER.
Operator decision: moving a prompt to another node IS the operator approving it.
## Task
1. SENDER GATE (do this first). Moving must require exactly the same human
proof that desktop "Approve & run" requires. No proof: refused, row untouched.
2. Carry the proof (the approval receipt, not a free-text label).
3. RECEIVER. Remove the policy gate. Accept only from a known node that
carries a sender approval receipt.
...
This replaces the receiver-side project-policy gate; it does not add a new
mechanism alongside it.
## Verification
- Move without human proof is refused and the row status is unchanged.
- Name the exact log line that proves a live Mac->PC move succeeded.
## FAILURE CONDITIONS
- Deleting the receiver check without adding the sender gate. That reopens
agent self-approval through a move.
- Trusting approvedBy / sourceNode text as authority.
- Shipping only one node's half.
In plain words… Orbit can run across more than one computer, I moved a waiting prompt from my Mac to my PC, and the PC refused it. An earlier fix had put the safety check on the receiving computer, when the real gap was on the sending side, where an agent on my machine could move a prompt without me approving it (and that rule exists so agents can’t approve their own prompts). I already knew the fix, so the prompt says exactly what it is, how to prove it worked, and which wrong fixes to stay away from.
Why do I care?
Cuz each mode goes wrong in its own way. The model is often smarter than you here (it’s definitely smarter than me at a lot of this), and if you pile on rules, its work can actually get worse… it follows your half-baked first guess instead of finding something better, which is why a clear goal and a critic beat a page of instructions. But a loop with a fuzzy goal can wander down a rabbit hole and burn through your tokens. And when you already know the fix, a loose prompt just makes it guess, and it might take a shortcut and solve it the wrong way. With #8072, the quickest way to make that error go away was to just delete the check on the PC… the move would work, it would look fixed, and agents could approve their own prompts just by moving them, which is exactly why that’s the very first failure condition.
How do I pick?
Ask yourself how much of the answer you already have.
- Nothing yet? Talk.
- Just the goal? Let it loop.
- The exact fix? Spell it out.
And whichever one you pick, give it a way to check its work, or at least decide what “done” means before it starts… a test, a critic, a screenshot. That’s what keeps it out of the rabbit hole (and it’s in Boris’s post too, he says to tell the model what you want, how much effort to spend, and how it should check that it did the right thing).
If you want to go deeper, I put the whole thing in the free Learning Atlas too… the three modes, the tradeoffs, and how to pick.
Where Orbit fits
That “how do you know it’s done” part is a big piece of why I built Orbit (in Alpha now for free!). A prompt like #8072 waits in a queue until I approve it, with an optional auto-review first, and while an agent’s working I can watch its CLI session live and send it a message if it’s heading down a rabbit hole. Once it ships, it gets logged with its commit and the model that ran it, and Orbit checks the commit diff for ghost ships (the agent says it did something and the commit doesn’t back that up) and for changes to files that had nothing to do with the task. And when I put exact test commands in a prompt’s verification section, Orbit runs them itself in an isolated worktree, keeps the evidence, and flags it if a required check got skipped.
Orbit is a desktop app for Windows and macOS, in Alpha and free right now, and you can grab an invite here.
And I’d love to hear how you prompt… are you a “give it a goal” person, or do your prompts have failure conditions too?