The prompt is the spec

Watch what happens when you ask Cursor "add an orders endpoint" with nothing else. It picks a router file at random. It invents a database helper that doesn't exist. It adds three error cases you didn't ask for and skips the one that matters. You spend twenty minutes reading code that fights your codebase.

Now watch what happens when you ask the same question with four sentences of structure. Cursor lands a diff in one shot. You read it, maybe tweak two lines, and ship.

The difference is not the model. The difference is the prompt. The scaffold I lean on has four parts — context, goal, constraints, format. It's not a canonical framework from Anthropic or Cursor; it's the shape that, in practice, fixes the most one-shot prompts. Run the editor on the right and read it out loud.

Part 1: context

What file am I in. What stack. What constraints exist already.

I'm editing app/api/orders.py. It uses fastapi and a postgres pool.

This is the part most people skip, and it's the part that does the most work. Without it, the model is guessing what kind of project you're in, and a guess takes ten generated tokens to wrong-foot you for the rest of the session.

In Cursor, Agent mode pulls workspace and open-file context automatically (other modes are mode-dependent — @-reference explicitly when in doubt). In Claude Code, what's in the conversation IS context. Either way, make the relevant file or pattern visible before you ask the question. Paste a snippet, open the file, or add it with @.

Part 2: goal

The single concrete thing I want.

Add a GET /orders/{id} endpoint that returns one order.

One sentence. One outcome. If you find yourself writing "and also" or "while we're at it," stop. That's two prompts pretending to be one. Send them separately. The model handles one well-scoped task per turn much better than two half-scoped ones.

Part 3: constraints

What you must use. What you must not do. What edge cases matter.

Use the existing get_pool() helper. Return 404 if missing.

This is where you save yourself the rewrite. Every constraint you state up front is a wrong path the model never goes down. "Don't add new dependencies." "Match the style of the existing endpoints." "No try/except for happy-path errors — let them bubble." Each line is a fence.

Part 4: format

How I want the answer back.

Show me the diff only. No explanation.

This is the most overlooked part. AI tools default to explaining what they did, which is helpful when you're learning and noise when you're shipping. Tell it: "diff only," "code only," "list of file paths," "two sentences max." The format constraint is free token savings on every turn.

What the four parts do together

When you read AI-generated code that's wrong, almost every time you can trace the wrongness back to a missing piece:

• Wrong stack assumed? Missing context.
• Did six things you didn't ask for? Missing goal scope.
• Used a banned library? Missing constraint.
• Wall of unwanted commentary? Missing format.

Get all four in. The model has nothing to invent.

What this lesson covers

Eight more steps. By the end you'll be able to look at a vague prompt and rewrite it on the spot, you'll know when a long-running session has gone toxic, and you'll have one tool — the show me your plan first prefix — that turns every prompt into a checkable two-step.