The right way to write a prompt: talk, loop, or spell it out
Published Updated
This is the Learning Atlas from Orbit. Get your invite
AI agents and harnessesConcepts and methodologyOrbit-internal concepts
-
Step 1: Prompt
Why this matters here
Start with the prompt itself, because the real question isn't short versus detailed. It's how much of the answer you already have, and that decides how much you need to write.
In short
The message you give an AI model or agent describing what you want. How much detail it needs depends on how much of the answer you already know.
-
Step 2: Spec
Why this matters here
When you have no idea yet, talk it out like you would with a coworker. That conversation is how the plan gets made, and the spec is what it leaves behind for a fresh agent to build.
In short
A written plan for what to build, including the goal and important details, that a fresh agent can use to carry out the work.
-
Step 3: Agentic loop
Why this matters here
When you know the goal but not the details, say the goal and let the agent loop. The model is often smarter than you at the details, and piling on rules can make its work worse.
In short
The continuous cycle where an AI agent observes its environment, makes a plan, executes a tool, evaluates the result, and decides what to do next.
-
Step 4: Critic agent
Why this matters here
A loop needs someone to judge each round. A second agent as critic, checking against a clear goal, usually beats a page of instructions.
In short
A second AI agent that reviews each round of work against a clear goal and sends it back with notes until the result is good enough.
-
Step 5: Tool loop exhaustion
Why this matters here
This is the catch with loops. With a fuzzy goal, the agent can wander down a rabbit hole, repeating work that never gets closer to done.
In short
A failure state where an AI agent gets stuck repeatedly executing the exact same incorrect tool actions until it runs out of its allowed token budget or context window.
-
Step 6: Token budget
Why this matters here
Rabbit holes cost real money, because every lap burns tokens. A clear goal and a way to check are what keep the budget from draining.
In short
The strict allocation and monitoring of how many API tokens an AI agent or application is allowed to consume during a specific task or workflow.
-
Step 7: FAILURE CONDITIONS
Why this matters here
When you already know the exact fix, spell it out: the context, the task, how to check it, and what not to do. A loose prompt here just makes the agent guess, and it might take a shortcut, like deleting a failing test instead of fixing the bug.
In short
A required final section in Orbit prompts listing specific, verifiable, binary criteria that constitute prompt failure. Catches incomplete or wrong implementations.
-
Step 8: Definition of done
Why this matters here
Whichever way you prompt, decide what done means and give the agent a way to check its work: a test, a critic, or a screenshot. That's what keeps it out of the rabbit hole.
In short
The agreed check that says a task is finished, such as a passing test, a critic's sign-off, or a screenshot that matches the expected result.