# Implement With Tdd > Implement a SPEC, plan, or feature outside-in, with TDD. - Skill: `stevepolitodesign/implement-with-tdd` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add stevepolitodesign/implement-with-tdd` - Raw SKILL.md: https://api.skillmd.com/api/skills/stevepolitodesign/implement-with-tdd/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: stevepolitodesign (https://skillmd.com/u/stevepolitodesign) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/stevepolitodesign/implement-with-tdd --- # Implement with TDD Work from the outside in and let each failure name the next move. Stop at shameless green; a later step refactors. The discipline exists because you're fluent enough to write plausible code for almost anything, and plausible isn't correct. A failing test tells them apart. ## What you were handed The thing to build: `$ARGUMENTS` If that's empty, ask. If it's a path, read it. Take the acceptance criteria as the scope. Those are behavior, and behavior is what you can test. A suggested design, a test path, an ordering: that's a guess from a session that never ran the suite. Yours will run it, so your evidence is better. Where the repo contradicts the plan the repo is right — say so rather than diverge quietly. A `## Where to look` section is the exception to that distrust, not to the scope rule. Those paths are where a prior session already found the relevant code, so they save you the search and not the judgment: open them before you go looking for anything else, and treat nothing on the list as work you owe. The criteria are still the whole scope. If a path is gone or turns out to be irrelevant, skip it and say so in the report — don't edit the SPEC. The list names the commit it was written at. When that isn't `HEAD`, the paths usually still hold and the reason beside each one may not, so read the file before you act on its description. ## Before the first test Learn the suite: - The command for one test file, and the one for everything - Which layers exist here (end-to-end, request or API, unit) — you can only start at the outermost layer the repo actually has - The local idiom: factories or fixtures, helpers, naming Read the two or three tests nearest your area and copy their style, even where you'd write it differently. `references/test-style.md` covers how to write the test itself and what belongs at each layer; read it before the first one. If the suite can't be made to run using the repo's documented setup, stop and report what's missing — a missing database or service isn't something to work around. If there's no test framework at all, stop and ask which to use. ## Size the feature Roughly 200 lines of implementation is the budget for one feature. Count every line you typed into a file that ships — views, routes, config, locales, `end`s and blank lines included. Free: tests and the factories or fixtures feeding them, comments, and generated files you didn't edit — schema dumps, lockfiles, snapshots, scaffolding output. A migration you generated is free; the body you wrote inside it isn't. Tests stay free here even where your usual diff budget counts them. Count what you added, not net. Deleting doesn't buy room, and code you moved unchanged is free. `git diff --stat` minus your test files is close enough; don't script it. Estimate now, with the neighboring code read. Per criterion, name the files you'd touch and whether the layer already exists — several criteria each needing a new object is the shape that blows the budget. If the estimate is over, stop before the first test. Report the number, the criteria carrying the weight, and the smaller first cut you'd take instead — the one or two criteria that stand on their own. Then wait. What to drop is a scope decision, and starting is a way of making it for them. ## The loop One criterion at a time, and only one test you wrote red at once. 1. **Write the outermost test for the criterion.** Drive it the way a user reaches the behavior. It should fail on the thing you haven't built, not on a typo. 2. **Run it and read the failure.** Literally — not the failure you expected. 3. **Make the smallest change the failure asks for.** Not the change that finishes the feature. `references/failures.md` maps each kind of failure to its next move, including which ones send you down a layer and which don't. 4. **Run it again.** A new message means you moved. An unchanged message means you did something necessary but insufficient, or you guessed — say which, and drop the change if you can't name what it unblocks. 5. **Repeat until green**, then run the tests around what you touched. A failure in something you never touched may predate you — check it against a clean checkout before you spend the session on someone else's broken test. Check the count against the budget, then take the next criterion, and run everything once at the end. Work through every criterion without checking in. Four things earn a stop: a suite you can't run, the same failure surviving two different fixes, anything needing a new runtime dependency, and the budget running out. The last two are project decisions, not implementation details. When the count crosses, finish the criterion you're on and stop there rather than opening the next one — a half-built criterion is worse than a missing one. Report the count, which criterion crossed, and what's left unbuilt. ## Shameless green Reach green with explicit, duplicated, readable cases. Spell the value out. Copy the neighboring case and change what differs. Write one branch per case the tests name, and none for a case no test named, since that branch is a guess you'd have to maintain. A passing test is the signal to stop writing code. You're trading changeability for clarity, not clarity for speed. Duplication is cheap to remove later; a wrong abstraction drawn from one example spreads into every caller first. The duplication is also the evidence the refactor runs on — it shows where the seam really is. Fakes aren't shameless, they're false. A hardcoded return value that satisfies the test is fine; an accessor standing in for a column that ought to exist, or a stub where the real call belongs, makes the test lie about what works. If the criterion needs schema, add the schema. Duplication counts against the budget. If spelling the cases out is what blows it, keep the duplication and stop at the boundary above: that means the feature is too big for one slice, not that it needs an abstraction. No comments apologizing for the shortcut, no TODO promising to fix it, no half-built abstraction for later. ## When you're done Report the criteria you covered, any you left unbuilt and why, the tests you added and where, the suite result, and the lines you spent against the budget, including whether your estimate held. Then name the ugliest thing you left standing, because that's where the refactor starts. Leave everything uncommitted.