quill-builder-spec

name	quill-builder-spec
description	Partial Skill: invoke by name only — the Quill Builder bar at the spec gate — what a documentation spec must contain. Loaded by the Quill spec-producer to self-align and by the spec gate to grade. Not triggered by users directly.
user-invocable	false
metadata	{"actor":"builder","gate":"spec","compose":"union"}

Quill Builder-Spec Governance — the documentation contract bar

The Builder bar at the spec gate, specialized for documentation artifact-types (documentation, guide, tutorial, article, reference). It unions onto sdd:builder-spec-governance — the generic testability and coverage bar still applies; this adds what makes a documentation spec a complete contract. One merged bar read by both faces: the spec-producer (quill-spec-writer) reads it forward, the spec gate reads it backward.

A document is an implementation artifact with verifiable structure (design/doc-eval-model.md). Its spec therefore freezes what the document must land, never how it is written.

The shape of a documentation spec

The four SDD sections still apply (sdd:spec-format-governance). This bar says what fills them.

## What — seven required elements

1. Audience — a table, never a phrase.

Audience	Who they are	What the document gives them

• Name a role plus what they are trying to accomplish. "The reader", "users", and "developers" are not audiences — they name no goal, so nothing about the document follows from them.
• An audience with no entry point in ## Use Cases is not an audience. Either add the entry point or cut the row; an audience listed and never served is the most common way a document silently fails half its readers.
• Two audiences needing opposite things from one fact is a split signal. Record the decision either way — one document serving both, or two documents — but never leave it undecided.
• The audience list is what the coverage table and the use cases are derived from. Write it first.

2. Doc type — one of four, declared. The type sets what "good" means, and mixing types in one document is the most common structural defect (Diátaxis).

Type	The reader is	Success is
tutorial	doing something for the first time, learning by doing	it worked, and they now trust they can do it
how-to	accomplishing a goal they already understand	they got unblocked
reference	looking one thing up	they found it, and it was accurate
explanation	building understanding, not doing a task	they can now make a decision they could not before

3. North star — one sentence, plus a failure mode. The understanding or capability the reader leaves with. It must carry a concrete state a reader could be in that would mean the document missed — without that, the north star is unfalsifiable and grades nothing.

A reader finishes able to X. A revision that leaves a reader able to Y but not X has missed.

Its kind is fixed by the declared type. The Success is column above is not description — it is the form this element must take. Nothing else joins the two, so a spec can declare reference and write an explanation's north star with both elements looking correct on their own:

Declared type	The north star claims	The failure mode names
tutorial	the reader can now do the thing, and trusts they can	a reader who followed it through and still cannot
how-to	the reader got unblocked on the goal they arrived with	a reader still blocked, or unblocked on a different goal
reference	the reader retrieved the one thing they came for, accurately	a reader who found the entry and still cannot settle their case
explanation	the reader can make a decision they could not before	a reader who can restate the argument but not act on it

Reference drifting toward explanation is the common direction — "a reader finishes understanding how the bar works" grades nothing a lookup cares about, since comprehension does not fail when the entry the reader came for is missing. Every node in this corpus that declared reference made this join by hand; the bar is what stops it depending on the producer noticing.

4. Why it exists. The problem the document resolves, in the domain's own terms. If it cannot be stated without restating the title, the document may not need to exist separately from its parent — say so rather than papering over it.

5. Key points — required coverage. A numbered table: | # | Topic | Must convey |.

• Each row is a claim the document must land, not a section name. "Separating targets lets contradictory values coexist" is a claim; "the separation section" is a heading.
• The document is incomplete without each row. The test: if this row were dropped and the document unchanged, would a reader be worse off? If no, it is not a key point — cut it.
• Each row must be checkable by static inspection (the four checks in design/doc-eval-model.md: existence, structure, completeness, reader-path).
• The coverage list is complete when a document meeting every row cannot still trip the north star's failure mode. If it can, a key point is missing.
• Spend every row by its number. That completeness claim is an argument against the table, so make it row by row: cite each # and say what it rules out. A paraphrase — "the rows on splitting and on precedence cover it" — is what goes stale, because the argument keeps reading as sound while the row it names has moved underneath it. Three regressions in this bar's own corpus were exactly that, each surfacing only after a different, unrelated fix shifted what the summary referred to. A row the argument cannot spend is a row nothing depends on — cut it.

6. Non-goals. What the document deliberately does not cover — and where that lives instead. A non-goal with no forwarding address reads as an omission rather than a decision.

7. Prerequisites. What a reader must already know, and which document supplies it. A document claiming to be self-contained declares that here explicitly.

## Use Cases — reader entry points, grouped by audience

One row per way a reader arrives, as trigger / what they bring / what they leave with. Group the rows under their audience. Every audience gets at least one; every entry point traces to a coverage row that serves it.

## Control Flow — the reader's decision path

Draw the questions the document must answer to route a reader, not its table of contents. A section list is not a control-flow graph: it records what was written, while the CFG records what a reader needs and in what order they need it. Where the document serves several audiences, the first branch is usually which audience the reader is, because that selects which part they need.

An outcome node holding a disjunction is an unmade decision. A leaf reading [do A, or do B] leaves the reader exactly where they arrived — needing to choose, with no criterion to choose on. Promote it to a decision node and put the discriminator on the edges:

Before — one outcome, two options, no criterion:

K -- yes --> K1[split into units, or scope by prose]

After — the criterion is the node, and each option is its own outcome:

K -- yes --> K1{would splitting copy more than it separates?}
K1 -- no --> K2[split into single-target units]
K1 -- yes --> K3[keep one file, branch by prose]

A disjunction in a decision node is fine — that is a question. A disjunction in an outcome is a decision the author deferred onto the reader, and it is the shape that lets a document ratify whichever branch its draft happened to take.

A route must reach every option the spec itself enumerates. Where a coverage row names a set — three mechanisms, four arrangements — and a node routes a case across that set, an option silently absent from the routing is a gap, not a simplification. Route to it or state why it is excluded.

Check this against the spec, never against the document: at the spec gate the document may not exist yet, and grading a contract by the draft it was written to catch is how the draft's omissions become the contract. The coverage table is the inventory; the CFG must spend it.

## Scenario map — 1:1 with the suite

Standard (sdd:spec-format-governance). Every coverage row (element 5) is reachable from at least one scenario; a row no scenario checks is unenforced.

A claim several passages could carry is asserted against the reader's paths, not against a count. Then it states X is satisfied by X appearing anywhere, so a coverage row whose claim is load-bearing in several sections is under-specified: the scenario passes on a document that lands the claim only where one reader will pass. Assert the paths instead:

Then it presents two values that cannot both be one house style
And it states that contrast on each path the control flow routes to it

The CFG is already in the spec and the passages are already in the document, so this is a comparison the reader-path check settles. It freezes no wording, no section order, and no count — only that a reader who needs the claim reaches it.

Do not assert a place count. An earlier revision of this bar prescribed in exactly one place, later passages referring back on the reasoning that an unquantified suite pays for restatement. Recurrence has no empirical warrant as a defect (.research/documentation-craft/), and the count gets the driving case backwards: a reader arriving at a later section from the sidebar has not read the lead, so the "redundant" restatement was that reader's only statement of the claim, and replacing it with a pointer made the document worse for exactly them.

A routing scenario asserts the discriminator, not the destination. Then it directs that case to prose matching names one member of a set and passes whether or not that member is the right one — so the suite ratifies the draft's route instead of testing it. Assert what decides:

Then it directs a target whose rules can stand as their own unit to description matching
And it reserves prose matching for variants that splitting would duplicate

A destination-only Then is worth writing only where the set has one member. Wherever the reader is choosing, the criterion is the behavior and the destination is a consequence of it.

Relations the suite still cannot reach, because they hold between passages no single scenario reads, are graded once per document against quill:quill-builder-impl.

Identifiers — one namespace per node

A doc spec keys three populations: coverage rows (element 5), reader entry points (## Use Cases), and CFG node labels (## Control Flow). Coverage-row IDs are cited from the other two — an entry point ends (G1, G3, G7), a scenario map row names the row it checks — so the keys are load-bearing here in a way they are not in a capability spec, whose use cases are named rather than numbered.

An identifier denotes one thing per node. Two populations reaching for the same audience-initial letter is how it breaks: an entry point keyed A1 for the adopter and a CFG node keyed A1 for the actor-name branch are two objects that a grader tracing a citation reads as one. This is measured, not anticipated — the website corpus shipped five such collisions across two nodes, four of them in a single file where A1 A2 T1 T2 were simultaneously entry points and graph nodes.

Prefix reader entry points UC. UC1, UC2, UC3 — two alpha characters, so it cannot collide with the single-letter labels a mermaid graph uses, and it is the same key in every node. An audience-initial prefix loses nothing by going: the rows are already grouped under their audience's own heading, so the letter restates the heading while colliding with the graph. Twelve authored nodes currently use twelve schemes, which costs a reader moving between them a fresh key each time.

After an edit — reconcile what references the claim, and audit what freezes it

A node is a web of restatements: one coverage row is cited by an entry point, described in the CFG, and frozen by a Then. Change the claim and the copies stay where they are. Four of four remediated nodes regressed in a single round, and every finding was introduced by the remediation itself — so this is a duty of the edit, not of the next review.

Reconcile every passage that references the claim. Having changed a coverage row, a north star, or an audience row, walk what cites it — entry points, the CFG, non-goals, prerequisites, the scenario map — and correct whatever still describes the old claim.

Then audit every Then you wrote or touched. The sweep above is necessary and not sufficient, which is measured rather than supposed: run on its own it found no staleness — a cold judge confirmed that — and the same node still shipped a defect in each of two consecutive rounds, both a Then that mistranscribed the clause it freezes. Confirming that a reference still points somewhere does not confirm it still says the same thing. So quote the source clause beside each touched Then and classify the pair: same / narrower / wider / different. Only same survives; the other three are the finding.

Where a Then's only home is a CFG node label, say so. One of those two defects existed precisely because a label was a claim's sole home — a claim with one copy has nothing to disagree with, and reads as consistent for exactly that reason.

A sweep is not self-certifying. Both halves reduce rounds; neither replaces the cold judge.

What a documentation spec must never freeze

Freezing any of these produces a spec that breaks on every honest revision while catching no real defect — the failure mode that makes teams abandon doc specs:

• Section order. A reordering that serves readers better must not fail the gate.
• Wording, phrasing, or headings verbatim. Assert the claim, never the sentence.
• The specific examples used. Assert that an example of the required kind is present.
• Length, tone, or voice. Not statically checkable, and not the contract's business (design/doc-eval-model.md).

The rule in one line: freeze the claims and the reader's path; leave the prose to the author.

Fit — when this bar does not apply

A subject with no inspectable document surface is outside Quill's lens and recuses to the SDD-default chain rather than being graded here (design/doc-eval-model.md, Fit). A spec whose artifact is code, config, or an agent definition is not a documentation spec, however much prose it carries.

Gaps are CONTENT_GAP, never guesses

A missing audience, doc type, north star, or coverage table is returned as a CONTENT_GAP for the conductor to put to the user. Do not infer an audience from the document's existing prose — that launders whatever audience the draft happens to serve into the contract, which is exactly the defect the audience element exists to catch.

Key points (read-check)

1. Audience first, as a table — a role plus a goal; an audience with no use case is not an audience; opposite needs on one fact is a split signal.
2. Declare the doc type — tutorial / how-to / reference / explanation; mixing them is the common structural defect.
3. The north star carries a failure mode, or it grades nothing — and its kind is the declared doc type's Success is cell: retrieval for a reference, a decision for an explanation.
4. Key points are claims the document is incomplete without, each statically checkable, complete when meeting them all rules out the failure mode — argued row by row, citing each #, since a paraphrased argument goes stale silently. A row the argument cannot spend is cut.
5. The CFG is the reader's decision path, not the table of contents — branch on audience first when there are several; an outcome node reading A, or B is a decision deferred onto the reader, and a route must reach every option the document names.
6. Never freeze order, wording, examples, or tone — freeze the claims and the reader's path.
7. Assert a load-bearing claim against the reader's paths, never a place count — states X passes wherever X lands, so require it on each path the CFG routes to it. Recurrence is not a defect.
8. A routing scenario asserts the discriminator, never the destination alone — otherwise it ratifies the route the draft took rather than testing it.
9. One namespace per node — an identifier denotes one thing; entry points are keyed UC1, UC2, so they cannot collide with a single-letter CFG label.
10. After an edit, reconcile the references and audit the Thens you touched — quote each against its source clause and classify same / narrower / wider / different. The reference sweep alone measurably misses mistranscription.