Coding

The developers getting real leverage from AI treat it like a sharp junior engineer: brilliant with clear instructions, dangerous with vague ones. These prompts are written to that standard — they assign a role, demand a format, and force the model to show its reasoning. A code review that flags the race condition instead of praising your variable names. A debug session that forms hypotheses instead of shotgunning fixes.

How developers actually use these

Not to write whole features — to compress the slow parts of the job. Explaining a regex you'll otherwise burn twenty minutes decoding. Turning plain-English questions into SQL you can verify line by line. Writing the API docs nobody wants to write. Planning a refactor of legacy code before touching a single line, so the scary change becomes a sequence of safe ones.

Rules that make code prompts work

• Context is the prompt. Language, framework, version, and the constraint that matters ("must run on Postgres 14", "no new dependencies") change the answer more than any clever phrasing.
• Ask for the reasoning, then the code. A model that explains its approach first produces better code — and lets you catch a wrong assumption before it compiles.
• Demand tests or examples. "Include three example inputs with expected outputs" is the fastest way to expose a subtly wrong regex or query.
• Verify anything you'd ship. Run it, read it, test it. AI code is a draft from a fast colleague, not a merged PR.

Pair these with the planning prompts for project-shaped work, or the productivity collection for the non-code half of a developer's day.

Which prompt for which moment

Staring at a wall of red error text: Debug Detective, with the smallest reproduction you can build. Reviewing a pull request — yours or someone else's: the Senior Code Reviewer, before the human pass. Writing a query against a schema you half-remember: SQL from plain English, then read every line of the output before running it. Facing a regex in the wild: the explainer first, generator second. Shipping an API: the docs writer while the behavior is fresh in your head. Touching legacy code: the refactor plan before a single edit. And when the feature finally merges, the commit message generator writes the message your future self will thank you for.