---
name: "workflow-content-engine"
type: workflow
description: The three-stage content pipeline that turns raw work evidence into publishable content — truth capture, then case study, then article. Owned by Draft and run for each venture, with the voice supplied per domain rather than baked in, so the engine itself is voice-agnostic. Read before drafting content for any venture, when deciding which stage a piece is at, or when checking what evidence a draft must be grounded in.
status: proposed
license: CC BY-NC-SA 4.0
---

# workflow-content-engine

Takes raw work evidence and produces publishable content in three stages. It is the **shared drafting engine owned by Draft** (`DR-4FT`) and run for each venture; the **voice is selected per domain** — `skill-content-voice` for careerOS content, `convention-mrp-voice` for Mr Pritchard pieces and notes — so the engine is voice-agnostic and the domain supplies the voice.

**Stage 1 — truth capture** (`skill-content-truth-capture`): facts are extracted from real source evidence — transcripts, logs, the work itself — and recorded as a traceable `tc.*` document before any prose is written. The capture creates the evidence layer the later stages draw from.

**A direct ask is bound by the evidence layer whether or not it entered through Stage 1.** The engine binds content that enters through its stages, and a one-off ask — a cover letter, a role application, a note written to a brief in a chat session — does not pass through Stage 1, so it arrives with no capture behind it. The evidence rule still holds. A claim about the person's own record — what they did, where, for how long, with what result — comes from a `tc.*` capture or from a CV variant, and where neither carries the claim the draft says less rather than asserting it. The absence of a capture is not licence to reconstruct the record from the model's own reading of it. The failure mode is not inflated language but **unsourced specificity**, which reads as authoritative precisely because it is concrete, and the only reader who can catch it is the person whose record it is — which is the review burden the evidence layer exists to remove. "No inflation" is a voice rule: it governs register, not provenance, so a confident, well-written claim that is simply not in the evidence passes it. Where a direct ask has no capture behind it, write what the captures support and name the gap, or run Stage 1 first (`APP-2754`: a careerOS cover letter drafted from a chat brief opened by placing a decade in payments, where the record carries payments as one key area and fintech across the decade, and the subject was the only reader who caught it).

**The return path when the capture itself is the thing that is wrong.** The rule above binds a draft to the evidence layer, and where no capture carries the claim it sends the draft back to Stage 1. It does not reach the opposite case: a filed `tc.*` or `cs.*` is present, does carry the claim, and is **contradicted by a newer artefact the stage is drafting from**. The evidence layer is the source of truth *and is refined as it is used*, so a disagreement found downstream is a correction owed to the capture rather than a detail to settle inside one draft. So: **a stage that finds a filed capture contradicted by the artefact it is drafting from records the disagreement on the source document** — naming both figures and which artefact it read — **before it drafts from either**. Writing it down is the whole of the obligation: this rule does not rule the disagreement, and resolution order stays where it already is, in the commissioning ticket's source order and `A1-961`'s instruction that where sources disagree the truth capture and the deck outrank the CV and the snapshot. The failure mode is **a contradiction resolved privately inside one draft**: the draft reads clean, the capture still carries the losing figure, nothing on the source says anyone ever saw the clash, and the next drafting leg re-derives it from scratch. `skill-content-truth-capture` step 5 is the adjacent rule and is **cited rather than restated** — it binds a *capture leg* re-deriving its source ticket's `## Evidence status`, which is a different leg from a *downstream stage* finding a filed capture contradicted, and it writes to the source ticket where this writes to the source document. (`APP-3047`: the careerOS truth captures and case studies had drifted from the published JPMC deck across four measured disagreements. When this rule was authored, all four of the named documents — `cs.jpmc-transformation-snapshot`, `cs.jpmc-transformation-deck`, `tc.jpmc-measure` and `tc.jpmc-team` — were read, and not one of them carried a recorded discrepancy against its source: the drift had been found and resolved nowhere the next drafting leg would see it. That is the state this rule ends, and the condition it creates is the test of it — under this rule each of those four says which of its figures disagrees with which artefact, so the clash survives the draft that found it. The four figures and the Figma file id are the filing session's reported readings and were not re-derived; only the documents' metadata was read. The *what* here is the Owner's statement of 2026-11-11 that the careerOS documents must stay the source of truth and be refined as work uses them; only the *how* is authored. Correcting the content itself — the figure, the metrics line, the team size, the role title — is cOS.Content work whose lane is the Owner's call, not a canon edit.)

**And publication moves the source of truth, so the stage that publishes reconciles the draft document or marks it superseded — and re-checks every derived outline still inside it.** The two rules above bind a draft to the evidence layer and send a contradiction back to the capture. Neither reaches the moment the *published* artefact becomes the newest thing in the chain: a long-form draft is filed as a Linear document and named on its ticket as the source of truth (`skill-content-article`, *Hand-off — where the draft lands*), the Owner then makes his final edits in the publishing surface rather than in the document, and the document goes on reading as current while the published copy has moved. The same document is where that piece's **derivatives** are outlined, so one stale section feeds every note and post drawn from it afterwards. So publication is two obligations in one pass. **Reconcile or supersede**: bring the document up to the published copy, or state inside the document which surface is now canonical for that piece and that the draft is superseded for it — and say what the document is still source of truth for, because a multi-piece draft is superseded piece by piece and not all at once. **And re-check the derived outlines against the published copy and the capture**, because a correction the Owner makes to a body does not propagate to an outline sitting in another section of the same document, and the outline is what the next derivative is drafted from; a figure corrected in the body and left standing in the outline is the shape to look for, and where the capture itself is the ambiguous one the disagreement is recorded on it under the rule above rather than settled inside the derivative. (Worked case: when this rule was authored, the live document "Mr P — philosophy series (draft 1)" was attached as the source of truth to `MRP-110`, `MRP-119`, `MRP-7`, `MRP-8` and `MRP-111`, and piece 1’s body read *"Pausing the channel brought conversion back to 48% without any design or engineering work. A few quick fixes the week after took it to 52%"* while the derived "The proof" bullet in the same document still read *"Pausing it took no design or engineering effort, and conversion recovered to 52%"*. The Owner’s correction had reached the body and had reached `MRP-8`, whose description carries the corrected split and flags that outline as wrong, and had not reached the outline sitting between the two. `MRP-110` carries a hand-written *Source of truth after publishing* section naming the published Substack piece as canonical for piece 1 and holding the document as source of truth for the unpublished pieces 2–8 — the instance this rule generalises, and the shape the condition this rule creates tests: under this rule the document says which surface is canonical for each published piece, says what it remains source of truth for, and carries no derived outline still contradicting the published body. `tc.just-onboarding` is reported by the filing note to carry only "within two weeks registration conversion returned to 52%", which is ambiguous on the split, and was not read when this rule was authored; where the capture is the ambiguous one the disagreement is recorded on it under the rule above. No stage in this engine, in `skill-content-article` or in `convention-mrp-voice` named a post-publication reconciliation anywhere in the store. `APP-3137`.)

**Stage 2 — case study** (`skill-content-case-study`): the truth capture becomes structured case content, formatted per the output specs (snapshot / full / article / cv-entry) with the writing standard applied throughout.

**Stage 3 — article** (`skill-content-article`): the verified evidence is shaped into long-form argument for the writing surface and downstream syndication, grounded in the truth capture and avoiding the failure modes of thought-leadership prose.

Voice governs throughout all three stages, selected by domain: `skill-content-voice` for careerOS (calm, precise, structurally confident; outcome before adjective; British spelling; no inflation), `convention-mrp-voice` for Mr P (the person's editorial voice, argued from the capture's evidence spine). Whichever applies, the register mechanics hold: sentence case, British spelling, no inflation.
