Coding Prompts for Cursor: Write a Spec, Not a Wish

“Add authentication” sounds clear to a human who already knows the repository. An AI assistant still has to guess the framework, identity provider, session model, persistence layer, protected routes, failure behavior, and migration constraints.

Every guess increases the chance of a plausible but incompatible implementation. A good coding prompt works like a small issue specification: it identifies the change boundary and the evidence required before the change can be accepted.

Start with the facts that constrain implementation. Mention the language and framework versions, package manager, relevant directories, persistence layer, and existing patterns that should be reused.

Implementation instructions tell the assistant what to type. Acceptance criteria tell it what must be true. Favor observable behavior: response status, UI state, persistence result, accessibility behavior, and test outcome.

In an established codebase, the assistant should inspect adjacent files before choosing an implementation. Tell it to identify the current pattern, state assumptions, and ask when a decision would materially change the architecture.

This reduces the common failure mode where the generated code is locally reasonable but duplicates an existing abstraction or violates a repository rule.

Ask for the smallest coherent change, not the fewest lines. A coherent change may include code, tests, documentation, and migration handling, but it should avoid unrelated cleanup.

For risky work, split the request into inspect, plan, implement, and verify phases. Review the plan before allowing migrations, dependency replacements, permission changes, or destructive commands.