/define-done

name: define-done
description: "Pin a measurable, outcome-based Definition of Done for a diamond — a behaviour-change that creates value, not a build-list. Problem-first Socratic sequence; writes definition_of_done to the diamond at birth or retrofits it when missing."
metadata:
  instruction_budget: "48"
  framework_dependency: "mycelium"
  framework_dependency_note: "This skill is designed to run within the Mycelium framework (https://github.com/haabe/mycelium). Standalone use will skip the canvas state, theory gates, and harness behavior the skill assumes. Install: /plugin install mycelium@haabe-mycelium."

Define Done

Pin an explicit, outcome-based **Definition of Done** for a diamond. "Done" is *a change in human behaviour that creates value* (Seiden), not "the feature shipped" — you're done when it ships **and** has the intended impact (Amplitude). The diamond's implicit done defaults to the harshest, least-controllable bar ("a user shipped a product"), which is both wrong for validating *purpose* and a demotivation engine. This skill makes the bar explicit, measurable, and pre-committed.

This is distinct from `/mycelium:definition-of-done` (the per-feature agile *quality checklist* run at `close`). This skill sets the **outcome bar a whole diamond is judged against**; that one verifies a feature meets quality criteria. Both can apply.

When to use

• **At diamond birth** — `/mycelium:interview` (L0) and child-diamond spawn (`/mycelium:diamond-progress`) call this before a diamond is "live."
• **Retrofit** — when `/mycelium:canvas-health`, `/mycelium:diamond-assess`, or the SessionStart nudge flags a diamond with no `definition_of_done`.
• Never silent-fill. The *question* is what produces a real bar ("fits, not ships"); a back-filled field is theatre.

Preflight: Read target file before any Write/Edit

**Hard rule.** The `definition_of_done` field lives on the diamond in `.claude/diamonds/active.yml`. Before `Write`/`Edit` on it, use the **Read tool** on that file this session (`cat`/`head`/`grep` via Bash do NOT satisfy Claude Code's check). `Edit` with `limit:1` suffices for a partial update; full Read before a `Write`. See `${CLAUDE_PLUGIN_ROOT}/engine/agent-operating-contract.md` *Canvas writes — Read before Write*.

The sequence (Socratic, problem-first)

Ask these in order. The teaching is in the **sequencing** (problem → signal → maybe-number → kill) and the good/bad contrast — not a help doc. Reject build-lists at step 1; that is the whole point.

**1. "When this is done, what's different — and for WHOM?"** → `outcome` + whose behaviour. Lead with the problem and the person, never the deliverable (Cagan: "most measurement problems are clarity problems").

• ✗ *"the onboarding flow ships."* (a build-list — reject it)
• ✓ *"non-fluent users get through the brief without hitting the vocabulary wall."*

**At the learning scales (L0 to L3) what changes is a decision, not yet a behaviour (v0.317.2).** Those diamonds exist to find something out (see the build-mode gate below), so what is different when one is done is what its owner can now decide: what gets built, resumed, rewritten or dropped. Write the outcome as what they will know plus the decision the answer unlocks, and make the "for whom" whoever takes that decision. A behaviour change in the users comes later, from what gets built once the answer is in.

• ✗ *"builders who shipped and heard nothing are shown to be stuck the way opp-080 says."* (a description of the need: nothing in it changes, and the problem appears only as an id)
• ✓ *"you know, from builders' own stories, whether 'shipped something, nobody came, and cannot tell whether people passed or never arrived' is real and repeated. That decides what gets built next: the parked ideas resume against it, or the problem is rewritten or dropped before anything is built for it."*

**Name the problem in words, never by an id alone.** The owner reads the bar weeks later without the canvas open, and an id carries nothing they can check against. Dogfood 2026-10-08: the owner could not interpret an L2 outcome that cited only its opportunity's id ("I cannot interpret the 'what changes', because I don't know what opp-080 says"), and once it was rewritten in words: "this kind of clarity must be a given in the way that these definitions are developed."

**2. "What's the ONE thing you'd SEE them DO that proves it?"** → `signal` (One Metric That Matters — one, not many).

• ✗ *"it feels better."*
• ✓ *"they came back for a second session."*

**3. *(optional)* "Is there a point where it's ENOUGH?"** → `threshold`, only if one genuinely fits. A number is **optional, not mandatory** — directional outcomes are legitimate. When you do set one, pair it with a qualitative guard ("3 warm bodies who bounce ≠ done") so it can't be gamed (Goodhart).

**4. Pre-mortem for the invalidation criterion** (schema key `kill_criterion`; the name is the worst reading of what it is — a pre-committed state and date at which the goal is shown WRONG, which is a completion path, not a termination; see `diamond-rules.md` on `close`). Ask in the past tense: *"It's [review date]. This diamond failed. What happened?"* (premortem finds ~30% more real failure reasons than "might it fail?" — Klein). From the answer derive a concrete **state + date** that means *kill, not finish*:

• `state` — an objective benchmark that means this goal is **wrong**.
• `date` — the review date by which the state must hold. Pre-commit BOTH, before the data (anti-HARKing).

Invalidation-with-evidence at that date is a legitimate "done" — routed through `/mycelium:diamond-progress kill` + `dogfood-mode`, not silently declared.

• `on_expiry` — **what happens when the date passes and NEITHER the outcome nor the kill has fired (v0.227.0).** Most results land between the lines, and a bar that is silent there stays open for good. The default is **stop**, or re-pitch as a NEW bar