Incremental Implementation

Incremental Implementation

Execute a plan as a series of small, verified, committed steps. The codebase is green after every step — there is no "it'll compile again on Thursday".

When to use

• A task plan exists and it's time to build (see task-breakdown)
• "Implement this feature" where the change spans multiple files/layers
• Large refactors or migrations that must remain reviewable
• Resuming work across sessions — each committed step is a safe resume point

Principles

• One task per cycle. Pick the next task, finish it completely, commit, then look up.
• Green to green. Tests pass before the step and after it. A red intermediate state never gets committed.
• Smallest diff that completes the task. Resist drive-by fixes — note them, don't do them.
• Follow the house style. Match existing patterns in the codebase, even when you'd personally do it differently.

The cycle

pick task → read context → implement → verify → commit → report → next

1. Pick

Take the next unblocked task from the plan. If no plan exists, make one first — even three bullet points.

2. Read context

Read the files you're about to change and one example of the pattern you're imitating (a similar controller, component, test).

3. Implement

• Stay inside the task boundary; out-of-scope discoveries go to a deferred list, not into the diff
• Write/adjust the test for this task as part of the step, not afterwards
• If the task turns out bigger than expected, stop and split it rather than pushing through

4. Verify

• Run the task's "done when" check
• Run the affected test suite (full suite if cheap)
• Run linter/formatter/type-checker if the project has them

5. Commit

One commit per task, message describing the outcome (see conventional-commits). The commit is the checkpoint — if step n+1 goes wrong, reset to it.

6. Report and continue

After each task, state: what was done, how it was verified, what's next. Then proceed — don't wait for permission unless the next step is risky or ambiguous.

When things go wrong

• Verification fails → fix forward only if the cause is obvious; otherwise git checkout . back to green and re-approach
• Plan turns out wrong → stop, amend the plan visibly, then continue. Don't silently improvise
• Blocked on a question → pick another unblocked task; park the question, don't guess on irreversible decisions

Output format (per step)

### Task n:  ✅
**Changed:** files + one-line summary
**Verified:** 
**Deferred:** 
**Next:** Task n+1: 

Anti-patterns

• ❌ Implementing tasks 1–6 in one giant diff, then debugging the pile
• ❌ "Tests are failing but it's because of the next step" — then the steps are wrong
• ❌ Drive-by refactoring inside a feature task
• ❌ Committing "WIP" — if it's not a completed task, it's not a commit
• ❌ Pushing through a too-big task instead of splitting it
• ❌ Silently deviating from the plan when reality disagrees with it

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: kwhorne
• Source: kwhorne/elyra-skills
• License: MIT
• Homepage: https://elyracode.com

Install and usage instructions live in the source repository linked above.