Skip to content

TDD

Bucket: Engineering · Slash command: /zsl:tdd · Source: skills/engineering/tdd/SKILL.md

When this skill activates

Claude Code matches this skill against the trigger text below. You can also invoke it explicitly with the slash command.

Test-driven development with red-green-refactor loop. Use when user wants to build features or fix bugs using TDD, mentions "red-green-refactor", wants integration tests, or asks for test-first development.

Philosophy

Core principle: Tests should verify behavior through public interfaces, not implementation details. Code can change entirely; tests shouldn't.

Good tests are integration-style: they exercise real code paths through public APIs. They describe what the system does, not how it does it. A good test reads like a specification - "user can checkout with valid cart" tells you exactly what capability exists. These tests survive refactors because they don't care about internal structure.

Bad tests are coupled to implementation. They mock internal collaborators, test private methods, or verify through external means (like querying a database directly instead of using the interface). The warning sign: your test breaks when you refactor, but behavior hasn't changed. If you rename an internal function and tests fail, those tests were testing implementation, not behavior.

See tests.md for examples and mocking.md for mocking guidelines.

Anti-Pattern: Horizontal Slices

DO NOT write all tests first, then all implementation. This is "horizontal slicing" - treating RED as "write all tests" and GREEN as "write all code."

This produces crap tests:

  • Tests written in bulk test imagined behavior, not actual behavior
  • You end up testing the shape of things (data structures, function signatures) rather than user-facing behavior
  • Tests become insensitive to real changes - they pass when behavior breaks, fail when behavior is fine
  • You outrun your headlights, committing to test structure before understanding the implementation

Correct approach: Vertical slices via tracer bullets. One test → one implementation → repeat. Each test responds to what you learned from the previous cycle. Because you just wrote the code, you know exactly what behavior matters and how to verify it.

WRONG (horizontal):
  RED:   test1, test2, test3, test4, test5
  GREEN: impl1, impl2, impl3, impl4, impl5

RIGHT (vertical):
  RED→GREEN: test1→impl1
  RED→GREEN: test2→impl2
  RED→GREEN: test3→impl3
  ...

Flags

  • --no-ship — skip step 6 (Ship it) and switch step 5 (Review) into autonomous mode (/code-review --auto). After review and any auto-applied fixes commit, stop and report back: branch name (git rev-parse --abbrev-ref HEAD), last commit sha (git rev-parse HEAD), a one-paragraph summary of the changes, and a Deferred review findings section listing the 60–79 confidence findings that weren't auto-applied. Do not push, do not open a PR, do not update the project board's "in review" Status. Used by /tdd-parallel so the orchestrator can integrate slice branches locally and ship a single consolidated PR. Step 1's "in progress" Status update still happens — the work is real, only the ship step is deferred. Also engages Progress heartbeat (see below) — the orchestrator depends on it for live status.

Progress heartbeat

Under --no-ship (running inside /tdd-parallel), emit one JSON line to .tdd-progress.jsonl in the current working directory at each phase below. The orchestrator tails the file to render live progress; without these lines it sees nothing until you return, and a silent slice looks indistinguishable from a hung one.

Schema: {"slice": <num>, "ts": "<ISO-8601 UTC>", "phase": "<phase>", "note": "<short>"}. slice is the input issue number, passed through verbatim.

Phases (in order; red / green repeat per cycle):

  • started — immediately after pre-flight, before planning.
  • planned — once the behavior list is decided. note: "N behaviors: ...".
  • red — each failing test written. note: "<n>/<N>: <behavior>".
  • green — each test now passing.
  • refactor — after each refactor step that touched files. note: what was extracted/renamed.
  • committed — after each commit. note: short sha.
  • reviewed — after the /code-review pass completes. note: "auto-applied N, deferred M" (AFK) or "approved N fixes" (interactive).
  • done — final phase, just before returning to the parent.
  • escalating — instead of done, if you escalate. note: one-line question.
  • error — instead of done, if you halt. note: one-line cause.

Emit via bash -c "printf '%s\n' '<json>' >> .tdd-progress.jsonl" — one printf per line, no buffering, append-only. Without --no-ship, emission is optional and harmless.

Step graph

stateDiagram-v2
    direction LR
    [*] --> setup
    setup: 1. Setup<br/>(branch, read sub-issue,<br/>scan tests)
    red: 2. RED<br/>(write failing test)
    green: 3. GREEN<br/>(minimal impl to pass)
    refactor: 4. REFACTOR<br/>(extract + delete<br/>aggressively)
    review: 5. REVIEW<br/>(/zsl:code-review,<br/>interactive or --auto)
    ship: 6. SHIP<br/>(PR or push or local mv)
    setup --> red
    red --> green
    green --> refactor
    refactor --> red: next behavior
    refactor --> review: all behaviors done
    review --> ship: approved (interactive)<br/>or ≥80 auto-applied
    review --> refactor: findings to address
    ship --> [*]

Heartbeat phases (for /zsl:tdd-parallel's orchestrator): setup, red, green, refactor, reviewed, shipped.

Workflow

1. Planning

Pre-flight: no-arg discovery (local-markdown trackers only). If no issue identifier was passed AND docs/agents/issue-tracker.md describes a local-markdown tracker (H1 reads # Issue tracker: Local Markdown, or the file describes the .scratch/<feature>/issues/ convention), discover what's ready to pick up instead of asking:

  1. List every .scratch/<feature>/issues/*.md file across all features. Exclude issues/_done/ and any feature already under .scratch/_done/ — those are archived.
  2. For each open issue, parse its ## Blocked by section. A reference is satisfied if the cited file (or the cited number under the same feature's issues/) resolves to a path inside issues/_done/. A reference that doesn't resolve to any file is a malformed issue — surface it to the user but don't crash.
  3. Present three buckets to the user:
  4. Ready to grab (numbered): feature slug, issue number, title, and a one-line summary (first non-heading line of the issue body). Sort by feature, then issue number.
  5. Pending future waves: blocked open issues + the Blocked by references each is still waiting on. Shown for context only — not pickable.
  6. Features fully archived but not closed: features whose issues/ directory contains only the _done/ subdirectory. Suggest running the feature-level close from step 6 against any of these before picking new work.
  7. Let the user pick by number or by path. Use that pick as the issue identifier for the remaining pre-flights and the rest of the workflow.

If the picked-from set is empty (no open issues anywhere under .scratch/), say so and stop — there's nothing to do.

If invoked with no argument against a GitHub/GitLab/other tracker, fall through to the next pre-flight, which will refuse for lack of an identifier. (Auto-discovery for remote trackers is out of scope here — use /triage to find work-ready issues.)

Pre-flight: refuse container issues. If you were given an issue identifier as input (or picked one in the previous pre-flight), fetch it and check whether it has open sub-issues. If it does, it's a tracking parent (likely a PRD), not a unit of work. Refuse and tell the user to run /to-issues to break it down first, then /tdd against one of the leaf children.

Pre-flight: signal "in progress" on the project board (if configured). If you were given an issue identifier and docs/agents/project-board.md exists, update the issue's project item Status to the option mapped to "work begins" (typically In progress). Use the same lookup-then-update procedure documented in triage/SKILL.md step 6 (including its hybrid-tracker rule: when the unit of work is a .scratch/ file, resolve the GitHub issue number from its github: frontmatter first): fetch the project item via gh api graphql filtered by the configured project node ID, then updateProjectV2ItemFieldValue with the mapped Status option ID. If the issue isn't on the configured project, log and continue. Best-effort — failure to update Status doesn't block the TDD work. Skip entirely if docs/agents/project-board.md doesn't exist or you weren't given an issue identifier.

When exploring the codebase, use the project's domain glossary so that test names and interface vocabulary match the project's language, and respect ADRs in the area you're touching.

Before writing any code:

  • [ ] Confirm with user what interface changes are needed
  • [ ] Confirm with user which behaviors to test (prioritize)
  • [ ] Identify opportunities for deep modules (small interface, deep implementation)
  • [ ] Design interfaces for testability
  • [ ] List the behaviors to test (not implementation steps)
  • [ ] Get user approval on the plan

Ask: "What should the public interface look like? Which behaviors are most important to test?"

You can't test everything. Confirm with the user exactly which behaviors matter most. Focus testing effort on critical paths and complex logic, not every possible edge case.

2. Tracer Bullet

Write ONE test that confirms ONE thing about the system:

RED:   Write test for first behavior → test fails
GREEN: Write minimal code to pass → test passes

This is your tracer bullet - proves the path works end-to-end.

3. Incremental Loop

For each remaining behavior:

RED:   Write next test → fails
GREEN: Minimal code to pass → passes

Rules:

  • One test at a time
  • Only enough code to pass current test
  • Don't anticipate future tests
  • Keep tests focused on observable behavior

4. Refactor

After all tests pass, look for refactor candidates:

  • [ ] Extract duplication
  • [ ] Delete aggressively — every refactor pass should remove something. Hunt for: dead code from over-eager early cycles, unused imports, debug statements (console.log, print, dbg!), commented-out blocks, comments that explain what the code does (rewrite the code instead). Bring net additions down where you can — TDD that only adds is half-done.
  • [ ] Deepen modules (move complexity behind simple interfaces)
  • [ ] Apply SOLID principles where natural
  • [ ] Consider what new code reveals about existing code
  • [ ] Run tests after each refactor step

Never refactor while RED. Get to GREEN first.

5. Review

After refactor lands and tests are still green, run /code-review against the slice diff to catch what the author (you) missed.

  • Interactive (default): runs with the approval gate. Apply the fixes you accept, re-run tests, commit fixes via /commit (same discipline as step 6).
  • AFK (--no-ship): runs as /code-review --auto. High-confidence (≥80) fixes auto-apply as a separate commit (subject review: <summary>) — /code-review commits directly here rather than going through /commit, so the AFK contract isn't broken by /commit's interactive prompts. Mid-confidence (60–79) findings get captured for the return summary under a Deferred review findings section with file:line refs.

If /code-review halts (lint or tests fail after auto-fix and the review commit was reverted), escalate per the AFK contract — return early with the failure cause.

Emit a reviewed heartbeat after the review pass completes.

6. Ship it

Skip this entire step if --no-ship was passed. Stop after the review step's final commit and report back: branch name, last commit sha, one-paragraph summary, plus any deferred 60–79 review findings. The orchestrator (/tdd-parallel) will integrate the branch and ship a consolidated PR.

Once tests are green and refactored, ship the slice. The repo's workflow is defined in docs/agents/ship-style.md (written by /setup-zsl-superpowers) — read it before doing anything.

  • Always commit via /commit — never craft commits yourself. The commit body must reference both the sub-task and the parent issue so git history is navigable. Use #<num> for GitHub/GitLab (auto-linked in the UI), paths for local-markdown trackers (e.g. .scratch/auth/issues/03-add-oauth.md), or full URLs for other trackers:
<subject>

Sub-task: #<sub-task>
Parent: #<parent>
  • Close the sub-task on merge. How depends on docs/agents/issue-tracker.md (check its H1 — # Issue tracker: <Name>):

  • GitHub / GitLab:

    • PR-style — put Closes #<sub-task> in the PR body, alongside the same sub-task/parent references.
    • Direct-push — put Closes #<sub-task> in the commit body itself (replaces the Sub-task: line).

    The tracker auto-closes the parent once its last child closes (the sub-issue link was set at /to-issues time). - Local markdown (issue is a file under .scratch/<feature>/issues/): - Update the Status: line near the top of the issue file to its final state — typically shipped. Use wontfix only if the maintainer is closing without delivery; since you just got the tests green, default to shipped. - git mv .scratch/<NNN>-<feature-slug>/issues/<NN>-<slug>.md .scratch/<NNN>-<feature-slug>/issues/_done/<NN>-<slug>.md. The filename stays the same — NN is permanent and unique across both folders. - Include both the Status: edit and the git mv in the same commit as the slice's code change so the close is atomic with the work that earned it. (The commit ref convention above already accommodates path-style sub-task/parent references.) - Feature-level close (optional, prompt first): after the git mv, list .scratch/<NNN>-<feature-slug>/issues/ and check whether any .md files remain outside the _done/ subdirectory. If none, tell the maintainer every issue in the feature is now archived and ask whether to archive the feature itself: git mv .scratch/<NNN>-<feature-slug>/ .scratch/_done/$(date +%Y%m%d)-<NNN>-<feature-slug>/. The date prefix orders archived features chronologically (ls .scratch/_done/ shows close order); the feature number stays embedded so number-based lookup keeps working. Don't move it without confirmation — the maintainer may want to spin up a follow-up issue first. - Confirm with the user before opening a PR or pushing to the default branch. - Signal "in review" on the project board (PR-style only, if configured). If docs/agents/project-board.md exists and the ship style is PR, update the issue's project item Status to the option mapped to "PR opened" (typically In review) once the PR is open. The project's existing Auto-close issue workflow will move Status to Done automatically when the PR merges and closes the issue. In direct-push mode, no skill-driven Status update is needed at ship time — the closing commit triggers the same Auto-close workflow directly.

If docs/agents/ship-style.md doesn't exist, run /setup-zsl-superpowers first.

Checklist Per Cycle

[ ] Test describes behavior, not implementation
[ ] Test uses public interface only
[ ] Test would survive internal refactor
[ ] Code is minimal for this test
[ ] No speculative features added

Bundled book rules

Do not hand-edit content between the BEGIN/END markers — scripts/sync_book_rules.py overwrites it from vendor/agent-rules-books/.

Rules from "Refactoring" by Martin Fowler

OBEY Refactoring by Martin Fowler

When to use

Use as a compact always-on rule set for changing existing code under tight context.

Primary bias to correct

Cleanup must preserve behavior and move in small verified steps, not become a rewrite.

Decision rules

  • Preserve observable behavior; isolate feature changes, migrations, redesign, and cleanup.
  • Work in small buildable, testable, reviewable steps; split changes that are too large to reason about locally.
  • Get a safety net or record the verification gap before risky structural edits.
  • Refactor the smell that blocks the current change, not every smell nearby.
  • Prefer simple named moves: rename, extract, inline, move, split phases, encapsulate mutation, decompose conditionals, and remove duplication.
  • Put behavior, state, and validation with the concept that owns them; avoid vague utilities, pass-through layers, and just-in-case abstractions.
  • Keep mutation and call contracts clear: avoid boolean flags, parameter reassignment, public mutable data, unnecessary setters, and hidden side effects.
  • Stop when the change is easy, the code is clearer, and further cleanup would be speculative.

Trigger rules

  • When behavior is unclear or tests are weak, characterize current behavior before broader refactoring.
  • When adding a feature is awkward, make the smallest preparatory refactor that makes it straightforward.
  • When the same edit appears for a third time, centralize ownership instead of copying again.
  • When conditionals or type codes grow, decompose intent before reaching for polymorphism, state, strategy, or lookup tables.
  • When a patch mixes cleanup with behavior or broad code motion, split it where practical.
  • When tempted to rewrite, choose the next small behavior-preserving transformation.

Final checklist

  • Same behavior?
  • Safety net or verification gap?
  • Small runnable step?
  • Clearer names, ownership, and control flow?
  • No speculative abstraction or mixed patch?

Rules from "Working Effectively with Legacy Code" by Michael Feathers

OBEY Working Effectively with Legacy Code by Michael Feathers

When to use

Use when changing poorly tested or poorly understood code under a tight context budget.

Primary bias to correct

Legacy work starts with control, not cleanup, rewrite, or elegance.

Decision rules

  • Treat code without trustworthy tests as legacy code: state what changes and what must remain.
  • Characterize uncertain current behavior before changing it, including ugly behavior consumers may rely on.
  • Use the legacy loop: find the change point, find an observation point, create or exploit a seam, break the blocking dependency, test, change, then refactor locally.
  • Prefer fast focused tests; use broader harnesses only as temporary first coverage when no narrow test point exists.
  • Create the narrowest useful seam for sensing or separation, and break only dependencies that block feedback.
  • Use sprout, wrap, parameterize, inject, extract, or override moves when direct edits would be unsafe.
  • Keep behavior changes, structural refactorings, and cleanup separate and small.
  • Do not leave test-only seams, hidden dependencies, wrappers, globals, subclass tricks, or link/preprocessor tricks without a cleanup plan.

Trigger rules

  • When behavior is unclear, characterize first.
  • When constructors, globals, statics, frameworks, I/O, clocks, randomness, environment, or deep object graphs block testing, break one dependency at the narrowest point.
  • When a large method or class defeats local reasoning, sketch effects and create a seam before semantic edits.
  • When rewrite or broad cleanup feels tempting, choose the next smaller verified move.

Final checklist

  • Behavior characterized?
  • Feedback fast enough?
  • Dependency isolated?
  • One kind of change?
  • Safer and clearer now?