/purpose-properties

name: purpose-properties
description: "Derive the binding properties your why/how/what impose on everything below them, so no solution can silently contradict your own product definition. Retrofit entry point for projects that started before v0.120.0; also the re-derive path when why/how/what change and every stance below them goes stale."
metadata:
  instruction_budget: "21"
  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."

Purpose Properties Skill

Retrofit and re-derive entry point for `purpose.yml#purpose_properties`. **For new projects this runs inside `/mycelium:start` Step 3b.** Use this skill when:

• The project existed before **v0.120.0** and has no `purpose_properties` (the retrofit case — every

project created before that release, since `/mycelium:start` Step 4 exits for already-initialized projects and never reaches the extraction).

• `why`, `how` or `what` changed, so `derived_from_hash` no longer matches and **every stance below

it is superseded** (the drift case — this is the one that matters most and the easiest to skip).

• A property list exists but the user wants to revisit which properties are `binding`.

Full contract: https://github.com/haabe/mycelium/blob/main/docs/purpose-stance.md (it lives in the repository, not in the installed plugin).

Step 1: Read the current state and say which case this is

Read `purpose.yml`. Then say plainly which of these you are in, because the three have different consequences and the user should not have to infer it:

• **No `purpose_properties`** → retrofit. Nothing below has ever been checked against the purpose.
• **Hash mismatch** → drift. Report it as *"your purpose changed after these were derived, so every

`purpose_stance` below them was reasoned under a definition you have since changed."* This is the state that quietly invalidates work; name it in those terms.

• **Hash matches** → nothing to do. Report the property count and stop. Do not re-derive for tidiness.

Step 2: Attempt extraction, and classify what fails

For every element of `why`, `how` and `what`, ask: **can I name a specific solution that would CONTRADICT this?**

• **Yes** → candidate property. Record `property`, `verbatim` (their words, not yours), `source`, and

the contradicting solution in `contradicted_by`.

• **No** → **do not force one.** Quality adjectives — *secure, accessible, fast, simple, delightful,

robust* — yield nothing checkable, because every candidate solution claims to satisfy them.

**Measured 2026-08-23, blind:** *"accessible and secure"* produced **no** checkable property, while five abstract framework principles produced clean ones. **The axis is not abstract vs concrete — it is quality adjective vs structural claim.** Do not treat concrete-sounding words as safe.

Step 2a: Screen for quality adjectives BEFORE trusting your own refusal

**DO NOT RELY ON NOTICING.** Step 2 says a quality adjective yields nothing checkable and asks you to decline. The evidence that an extractor actually declines is **one blind run, one model family** — the 2026-08-23 test, where the extractor returned NONE for *"accessible and secure"* unprompted. That result is real and it is thin, and this skill runs under model families that test never touched, in runtimes where no hook fires to catch the difference. **A model that does not refuse produces a confident property, marks it binding, and creates a stance field that every solution answers `preserves` forever — populated, green, and meaningless.** The mechanism's own failure mode, at its entry point.

So the screen is a rule, not a judgement:

**TRIGGER LIST — if an element contains any of these words, it routes to Step 3 REGARDLESS of whether you extracted something from it:**

> secure · safe · private · accessible · usable · intuitive · simple · easy · fast · performant · > scalable · reliable · robust · flexible · modern · seamless · delightful · high-quality · > best-in-class · efficient · powerful · lightweight

**Extracting a property from one of these words is not evidence the word was checkable.** It is the predicted failure. Route it, let the builder supply the standard, and record what they say.

**AND THE GENERAL TEST, for words the list does not carry** (it cannot be complete, and treating it as complete is the next version of this bug):

> **Would every plausible competing solution ALSO claim to satisfy this?** > If yes, it is a quality adjective, whatever it sounds like. Route it.

*"Secure"* fails that test — nobody proposes an insecure login. *"Anonymous"* passes it — a solution requiring an account plainly does not claim anonymity. **Note that the concrete-sounding word is the one that fails.** The axis is never how technical the word looks.

**WHAT IS STILL UNTESTED, stated so nobody reads the list as coverage:** whether a builder actually answers the Step 3 prompt. The 2026-08-23 run exercised extraction only — no builder was asked anything — so the recovery path has evidence for its necessity and none for its effectiveness.

Step 2b: Check the ALTITUDE of every candidate before you record it

**A property must sit at the altitude where violation is meaningful.** Ask, of each candidate:

> **What would have to be true for this to be violated — a single move, or a design choice?**

• **A design choice** → it is a property of the product. Record it.
• **A single move** → **you have extracted a rule about BEHAVIOUR, not a property of the product.

Do not record it.** A property that can be violated by one sentence turns the stance field into a footnote requirement, and a check people must satisfy per action is one they mute.

**WORKED EXAMPLE, and it is why this step exists.** From *"Theory-guided dec