The ogilvy-writing-audit skill (v0.1, built and packaged) checks business writing against David Ogilvy's ten rules. v0.1 judges voice against a generic standard — natural, spoken cadence. Inside a Loomworks engagement that is the wrong standard, because the engagement has declared a voice in its seed (R-A7), and the right standard is the declared voice, not generic naturalness.
This note scopes v0.2: make the voice audit seed-aware. The change is surgical — only the voice rule and part of the clear-ask rule become seed-aware; the rest stays universal. The hard part: seed voice is per-consumer (a board briefing and a regulatory submission have different declared voices in the same engagement), so the audit must first establish which consumer the piece is for.
Decision needed: confirm the four defaulted decisions in §6, or amend. Then it goes to a CR-drafting session.
in scope seed-aware voice (rule 2), ask (rule 9), jargon caveat (rule 4). out of scope fetching the seed, changing rules 3/5/6, any rewrite capability.
v0.1 audits six observable rules (2, 3, 4, 5, 6, 9) and surfaces four process rules (1, 7, 8, 10) as prompts. It produces a severity-sorted diagnostic. It does not rewrite.
On voice specifically, v0.1 rule 2 ("write the way you talk") rewards natural, spoken cadence and flags stiff, written-not-spoken phrasing. This is correct for generic business writing.
Prior position (v0.1): voice is a single quality — conversational good, stiff bad.
Corrected position (v0.2): voice is declared and per-consumer. The seed's R-A7 names the posture the work takes toward each consumer, and that posture varies by consumer within one engagement.
Evidence from a live seed (geology engagement, v0.3, R-A7):
"Evidence-based and descriptive for technical consumers. Advisory for the board. Declarative for regulatory submissions."
Three consumers, three declared voices, one engagement. A board briefing written in a conversational-casual register is wrong against this seed even though v0.1 would praise its naturalness. v0.1's generic standard actively misfires inside an engagement that has declared a non-conversational voice for the relevant consumer.
This is the crystallisation: the audit gains a question it did not have — who is this piece for? — because the answer selects which declared voice to judge against.
Surgical scope. Only where declared voice legitimately changes the judgement.
| Rule | Seed-aware? | Why |
|---|---|---|
| 2 — Write like you talk | Yes | The declared voice replaces the generic naturalness standard. "Advisory for the board" means measured advisory phrasing is correct, not stiff. |
| 9 — Clear ask | Partly | An action-directing voice should have a clear ask. A neutral/descriptive voice may legitimately have none. The seed tells the skill whether a missing ask is a defect. |
| 4 — No jargon | Caveat | Terms of art are not jargon when the reader shares them. R-A6 + R-A7 tell the skill which reader shares which vocabulary. "Evidence-based for technical consumers" means domain terms are expected. |
| 3 — Short sentences | No | Concision is good regardless of declared voice. |
| 5 — Two-page limit | No | Length discipline is universal; the long-form note already handles long renders. |
| 6 — Check quotations | No | Accuracy is universal. |
Because declared voice is per-consumer, the audit must establish which consumer the piece targets before it can pick the right voice. Two approaches were considered.
Before auditing, present the seed's R-A6 consumer list and ask which one the piece is for. Clean and unambiguous. Cost: adds a step; assumes the writer can name the consumer.
Read the draft, infer the likely consumer from content and register, state the inference, audit against that consumer's declared voice. Faster. Cost: inference can be wrong, and auditing against the wrong voice yields confidently wrong findings.
Infer, state the guess, let the writer correct in one word. This matches the project's "best-guess default, confirm with a single word" discipline. The safety property: if the skill cannot infer the consumer with reasonable confidence, it does not guess — it falls back to v0.1's generic standard and says so. Wrong-voice findings are worse than generic findings, so the fallback is the floor, not an error.
The skill does not fetch the seed. Claude.ai cannot reach record.dunin7.com (behind Cloudflare Access). The skill therefore requires the seed's R-A6 and R-A7 content to be in context — either the writer pastes the consumer list and voice declaration, or the skill runs in a session where the seed is already loaded.
If neither R-A6 nor R-A7 is present in context, the skill degrades to v0.1 generic behaviour and states that it is doing so. It never guesses what the seed says. This preserves drift-surface discipline: the skill names the source of the voice standard it applies (declared-from-seed vs generic-fallback) so the judgement is visible and correctable.
One word per decision confirms or amends.
v0.2 brings the skill closer to the seed, not into conflict with it. The seed commits R-A7 voice as governing shaping and rendering across the engagement's lifetime; v0.2 makes the writing audit honour that commitment instead of overriding it with a generic standard. No conflict surfaced. The skill remains additive — a quality lens, not a structural element.
ogilvy-writing-audit v0.2. Inputs: this scoping note + the v0.1 SKILL.md and references file.references/ogilvy-rules.md (rules 2, 4, 9 gain seed-aware subsections). Bump package and reference to v0.2.ogilvy-writing-audit-v0_2.skill; place under the user-skills path via CC with halt-before-commit.