| name | incremental-implementation |
| description | Drives incremental development through thin vertical slices. Use when: implementing features, building new functionality, executing sprint tasks, or when the agent tries to build everything at once. Enforces one-feature-at-a-time discipline with test-verify-commit cycles.
|
| allowed-tools | Read, Write, Edit, Bash, Glob, Grep |
Incremental Implementation
Overview
Build in thin vertical slices — one feature end-to-end before starting the next. Each slice is independently testable, verifiable, and committable. This prevents the "build everything then debug everything" anti-pattern that wastes hours on cascading failures.
When to Use
- Implementing any feature from a plan or spec
- Building new functionality with multiple parts
- Executing sprint tasks
- Agent is attempting to build multiple features simultaneously
- Large implementation that tempts a "do it all at once" approach
When NOT to Use
- Single-file config changes or one-line fixes
- Pure refactors with no behavioral change (use dedicated refactor workflow)
Process
1. Pick ONE Task
Select a single task from the plan. If no plan exists, stop and create one first.
2. Read Acceptance Criteria
Know what "done" looks like before writing code. If criteria are vague, ask for clarification.
3. Load Relevant Context
Read existing code, patterns, types, and conventions. Match the codebase style — do not introduce new patterns without discussion.
4. Write a Failing Test
If a testing framework is present, write a test that captures the expected behavior. Watch it fail. This confirms the test is actually testing something.
5. Implement Minimum Code
Write the minimum code to make the test pass. No extras, no "while I'm here" additions, no premature abstractions.
6. Run Full Test Suite
Run all tests, not just the new one. Check for regressions. A feature that breaks existing features is not a feature.
7. Run Build
Verify the project compiles and builds cleanly. Type errors, lint errors, and build warnings are bugs.
8. Commit with Descriptive Message
One commit per working slice. The message describes what the slice adds, not how it was built.
9. Mark Complete, Move to Next
Update the plan, mark the task done, pick the next slice. Repeat from Step 1.
Key Principles
Vertical slices, not horizontal layers. Build one feature end-to-end (UI + API + DB) rather than all features one layer at a time. A vertical slice is usable; a horizontal layer is not.
Each slice must be independently testable. If you cannot test it in isolation, the slice is too coupled. Break it smaller.
Never leave the codebase broken between slices. After every commit, the project must build, tests must pass, and existing features must work.
Commit after each working slice. Small commits are reviewable, revertable, and debuggable. Large commits are none of these.
If a slice is too big, break it into sub-slices. Signs a slice is too big: more than 5 files changed, more than 200 lines added, or estimated time exceeds 30 minutes.
Common Rationalizations
| Rationalization | Reality |
|---|
| "Let me build the whole backend first" | Untested code is unknown code. Build and verify slice by slice. |
| "I'll commit everything at the end" | Large commits are unreviewable and un-revertable. Commit per slice. |
| "This feature is too interconnected to slice" | Every feature can be sliced. Start with the simplest user story. |
| "Tests slow me down" | Tests slow you down now. No tests slow you down forever. |
| "I'll just scaffold everything then fill it in" | Scaffolds without tests rot. Build real slices, not empty shells. |
| "It's faster to do it all at once" | It feels faster. It is not. Debug time on monolithic changes dwarfs slice overhead. |
Red Flags
- Agent modifying more than 5 files without a commit
- Agent building multiple features before testing any
- Agent skipping test runs between slices
- Agent writing "scaffold" code with TODO comments instead of real implementations
- Agent committing hundreds of lines in a single commit
- Agent building all API routes before any UI (horizontal, not vertical)
- Agent saying "I'll test it all at the end"
Verification
