code

Step 6 of the pipeline (WORKFLOW.md). Build the minimum that makes each test pass.

When

After tdd (red tests written), during /BUILD.

Steps

0. Climb the ladder before writing, after reading the task and the code it touches. Stop at the first rung that holds: does the need exist at all; is it already in this repo; does the standard library do it; does the platform do it natively; does a dependency already installed do it; can it be one line; only then the minimum code. The ladder shortens the solution, never the reading. It never removes input validation at trust boundaries, error handling that prevents data loss, security measures or accessibility basics. The long form, with the guards, is skills/reuse-before-write.
1. Take one task_item, write the minimum code that makes its test pass.
2. Run that test scoped, per skills/run-generated-tests — the new/changed test, at widest the file it lives in, never the project's bare full-suite command. toggle_task_item when the increment is done, commit.
3. Blocked / unexpected error → invoke the debug block before proposing a fix.
4. Repeat until the task_items are exhausted.

Output / checkpoint

build_done.

Guardrails

No comments in the code. Never mark a test passes: true by hand, the GATE (step 7) decides, on evidence. We don't validate ourselves.

A failing test gets the implementation fixed, never the test loosened. Editing a pre-existing test file is for adding the case this task's new behaviour actually introduces, never for deleting, loosening or retargeting an assertion that was already there to make it match what the code currently does — that specific move is how a regression ships behind a green suite (see debug §3.4/Guardrails for the full statement, and route through tdd/dozer if the test itself is genuinely the thing that's wrong).

Origin

Native Claude Code + internal, rewritten. The no-test-tampering guardrail added 2026-08-11, same change as debug/tdd and the implementer agents: named directly by the operator, not sourced from a catalogue.

The ladder (step 0) is rewritten from the ponytail skill (MIT), read 2026-10-02. Its marker comment, debt ledger and benchmark figures are not carried over: comments are forbidden here and the author's numbers are not measured on this repo.