Specification
Tight specs in. Shipped code out.
An LLM doesn't know what you actually want. It knows what you said. The gap between the two is where most AI projects die. The work is closing that gap before any code gets generated: by tightening the brief, surfacing the unknowns, and putting words to the things that would otherwise be assumed.
The intern's loop
A junior's prompt is the feature description: 'add a billing page.' The model produces a plausible billing page, with the wrong currency, the wrong tax logic, no proration, calls to a service that doesn't exist. The next prompt is 'fix the currency.' Then 'fix the tax.' Three days later the codebase has six billing pages and none of them work. Speed without intent is just generating bugs faster.
The Before pass
Before any code: the LLM acts as a gatekeeper. What problem are you actually solving? What's the data shape? What breaks at the edges? What's the simplest version that ships? The brief gets challenged before the model is asked to execute it. It's a conversation, not a writing exercise. One or two extra prompts, hands-in-hand with the model, until the unknowns become decisions. For high-stakes work, the conversation lands in a spec.md the team can re-reference. For smaller work, it lands in a sharper prompt. Either way, the guesswork happens before generation, not after.
Intent that's been decided, not assumed
High-stakes specs live in the repo and version with the code, visible to the next engineer, the next team, the next model. Smaller decisions don't always need a file, but they've been made deliberately, in conversation, before generation. Intent gets recovered, not reconstructed.
- 01Before generating code, does something push back on the brief: a teammate, a checklist, a model?
- 02Could the intent of the last feature you shipped be reproduced by a stranger without asking anyone?
- 03When scope changes, what changes first: the brief or the code?
- 04For your highest-stakes work, is the spec written somewhere the team (and the model) can find it?