hamza-ali-shahjahan /feature-slicing
Usage: /feature-slicing <the goal, or what "done" looks like>
A feature you can't evaluate is a guess. A feature you can't test end-to-end is a hope. This skill turns a goal into a short list of vertical slices — each a thin, shippable piece of the goal — and keeps only the slices that come with a way to prove them. It runs after the goal is pinned (/write-a-goal, or a one-line acceptance) and before /spec, so the spec describes provable features, not a wishlist.
The rule
If you can't name the eval and the end-to-end test for a slice, it doesn't get built yet.
That's the whole discipline. A slice with no eval is either too vague (sharpen it) or not actually part of the goal (drop it). The selection gate is provability, not enthusiasm.
The method
- Cut vertically, not by layer. A slice is a thin end-to-end thing a user can actually do — not "the database layer." Aim for the smallest slice that moves the goal's metric.
- For each candidate slice, write three things up front:
- Eval — the measurable check that says this slice meets its bar. Tie it to the goal's metric where you can (reuse the numbers from
/write-a-goal). Deterministic where possible; an LLM-judge rubric only when the output is genuinely subjective. This is the eval the build is written against, and it joins the project's eval harness. - End-to-end test — the real user journey that proves it works in the running app, start to finish (not a unit test of one function). For UI, a browser path; for an API, a request→response contract.
- Tracking — how we'll see real users use it: for a new screen/tab/mode, a virtual
page_viewpath; for a new interaction, an event name + low-cardinality params; or an explicit "not worth tracking — <why>". Deciding is mandatory, tracking isn't — but a slice whose usage question ("did anyone touch this?") has no answer planned is instrumentation debt the founder pays later (see/user-analyticsRecipe F). "Events fire" joins the slice's e2e check.
- Eval — the measurable check that says this slice meets its bar. Tie it to the goal's metric where you can (reuse the numbers from
- Apply the selection gate. Keep a slice only if both the eval and the e2e test are defined and runnable. If either is missing: sharpen the slice until they exist, or defer it (write down why) — don't smuggle an unprovable feature into the spec.
- Order by risk-and-value. Sequence the kept slices riskiest/highest-value first, so the goal's metric moves early and a dead end is found cheap.
- Hand off. Output the selected slices, each with its eval + e2e test, as the input to
/spec→/plan. In the per-task loop, the eval and e2e test are written first (TDD) and are the slice's acceptance gate.
Output shape
Goal: <the measurable outcome>
Selected (build these):
1. <slice> — Eval: <measurable check> · E2E: <user journey> · Track: <event(s)+params | page_view path | "none — why"> · Why first: <risk/value>
2. ...
Deferred (can't prove yet):
- <slice> — missing: <eval | e2e> — <reshape-or-revisit note>
Why this exists
Eval-driven development (see factory/playbooks/ai-native-2026/eval-driven-development.md) says the eval comes before the code. Slicing is where that happens at the feature level: the gate that keeps the build honest. You only commit to features you can prove, and "done" for each is a green eval and a passing end-to-end test — not an opinion. It's the feature-level twin of write-a-goal, which does the same for the objective as a whole.