AI Tool Compass
Search

AI Workflow for Building a Small Feature

Use AI coding assistants across planning, implementation, review and documentation for a small, well-scoped feature.

Who it's for
Developers working solo or on a small team.
Difficulty
Intermediate
Time
Varies by feature size
Cost
$10-$40/month for AI coding tools
Goal: Ship a small, well-tested feature faster without skipping review.
  1. 1

    Clarify the spec

    Write a short, explicit description of what the feature should do, including edge cases.

    Open Claude →
    Prompt
    Help me write a clear spec for a feature that {{featureDescription}}. Ask me clarifying questions about edge cases first.

    Input

    Rough feature idea

    Output

    Clear written spec

  2. 2

    Scaffold the implementation

    Use an AI coding agent to generate an initial implementation from the spec.

    Open Cursor →

    Input

    Spec

    Output

    First-pass implementation

  3. 3

    Review the code

    Request a focused review for correctness, security and readability before merging.

    Open Claude →
    Prompt
    Review this code for correctness, security and readability: {{code}}

    Input

    Generated code

    Output

    Review notes

  4. 4

    Write or generate tests

    Generate tests covering the spec's edge cases, then verify they actually fail without the fix.

    Open GitHub Copilot →

    Input

    Implementation + spec

    Output

    Test suite

  5. 5

    Write the documentation/changelog

    Draft a short changelog entry or doc update describing the change.

    Open ChatGPT →

    Input

    Final implementation

    Output

    Changelog entry

Use cases

Frequently asked questions

Is it safe to merge code an AI agent generated?
Only after the review step — treat the scaffolded implementation the same way you'd treat a pull request from a junior contributor, checking correctness, security and edge cases before merging, not on trust.
Why use Claude for both the spec and the review, but Cursor for implementation?
Claude's strength here is careful, step-by-step reasoning — useful for clarifying requirements and critiquing code. Cursor is built specifically for generating and editing code across a codebase, which is a different job.
What if the AI-generated tests all pass immediately?
That's a warning sign, not a good sign — the workflow explicitly says to verify tests actually fail without the fix first. Tests that pass regardless of whether the fix exists aren't testing the right thing.
#coding#software-development