Content Engine

Classify the drop, mine stories and people, inject proof, write in the voice stack, QA gate, park at Review. One human flip publishes.

Ruled 7/10

Any drop in: a rough note, a voice or meeting transcript, a brain dump, a photo set, a link Andrew liked, a testimonial screenshot, or a messy pile of all of it. Nothing is ever rejected at intake, capture is free. The engine turns the drop into a QA’d, on-brand piece and parks it at Status = Review in the DH Content Feed. Andrew reads, edits if needed, flips Status = Published. The AI never publishes on its own.

About 70% of this already exists: the Content Feed schema, the Publishing SOP, the AEO doctrine, the voice stack, meeting-intel, toolbox-extract, and dh-site-engine. The middle piece, the classifier, the writer harness, and the QA gate, is what the dh-content-engine skill builds.

The pipeline

Intake

Everything lands as a row in the Content Feed at Status = Draft, with Source Type set (Andrew / Client / Guest / Tool / Research).

Classify

The engine reads the drop and assigns exactly one content type from the 8-type taxonomy below, plus Category and Tags. A long transcript gets mined first and can spawn several typed drops from one source.

Extract & enrich

Stories, wins, testimonials, and quotable moments get mined into the Proof Library. New people go to DH People. Tool mentions run through toolbox-extract into DH Toolbox. Then the engine scans the other direction: is there an existing proof point, testimonial, or person that belongs IN this piece? Every article should carry at least one face and one proof point where honest. Teaching Packages and Listicles also get one bounded research pass for 2 to 3 high-authority adjacent citations. Drops skip research entirely.

Write, in the voice stack

Brain Candy teaching voice first (make smart people feel smarter, easy on-ramp then deep, copy-paste-usable takeaways). STORIES as the invisible structure whenever the piece connects or sells. Then the hard voice rules as the non-negotiable floor: no em dashes, no contrast-correction patterns, category-brand pairing in openings, banned-word list. Full detail in Voice Rules.

QA gate

An automated checklist the writer cannot skip, detailed below. Fail any gate and the engine fixes and re-runs, max twice, then flags the failure in the row’s Notes instead of shipping it to Review.

Approval

The drop sits at Review with a 3-line header the engine writes into the page: what this is, where it came from, what it links to. Andrew reads, optionally edits in Notion directly (that IS the feedback mechanism, no separate tool needed), flips Status = Published. Three gates never get bypassed: sensitivity, permission, editorial.

Publish

The Status flip fires the rebuild through dh-site-engine to Cloudflare Pages. The page lands at /updates/[slug], wired to People, Tools, Concepts, one CTA Offer, and its Proof relations. llms.txt regenerates on every build.

The QA gate, in detail

Voice regex sweep — em dashes, contrast patterns (“not X, it’s Y”), banned words
AEO structure check — answer-first opening in the first 150 words, FAQ block for long-form, category-brand pairing present
Anti-slop rule — at least one unique human observation per major section, pure AI filler fails
Sensitivity gate — anything client-specific, financial, or personal without clearance gets flagged, never silently included
Internal link check — 2+ links to existing pages, people, or tools, with descriptive anchors
Rights check — every injected proof point or photo has Permission Status resolved

The 8 content types

One transcript can spawn several of these. The classifier decides which; Content Type is a select property on the Content Feed.

Type	Trigger	Length	Special rules
Drop	quick insight, decision, observation	100-400 words	timestamp prominent, feed-native, can upgrade to a pillar later
Teaching Package (pillar)	a framework or method	1,500-2,500	what it is → who it’s for → steps → example → when NOT to use, never paywalled
Listicle (“Best X”)	comparison / tool intent	1,500-2,500	comparison title allowed, DH placed honestly, the headline AI-citation format
Proof Story	testimonial, client win	400-900	reverse-engineered from real proof, rights cleared, face required
Project Showcase	photo set + project	600-1,200	photos in, hyper-local long-tail article out
People Page	person worth featuring	entity page	approval required for customers, entity anchor for E-E-A-T
Tool Page	tool in the Toolbox	entity page	provenance from toolbox-extract only, how-DH-uses-it angle
Refresh	quarterly decay pass output	varies	updates timestamp, strengthens internal links both directions

Measure and improve

Fathom Analytics tracks traffic and conversion per piece. AI-referral traffic gets filtered by referrer (chatgpt.com, perplexity.ai, claude.ai, gemini, copilot); a rise in Direct traffic is treated as a citation proxy since Google AI Mode and mobile AI apps are untrackable. A monthly manual citation check asks each engine the target questions and logs yes/no per framework. A quarterly decay pass pulls the refresh shortlist: pieces ranking 4 to 20 with falling CTR, or anything sub-100-visit and over 6 months stale gets consolidated or pruned. Refreshes re-enter the pipeline at Stage 4 as a Refresh-type drop.

Repeatability (BBE packaging)

Everything above is engine, not brand. Brand-specific material lives in four swappable slots:

Voice pack

Brain Candy + STORIES for DH. A different voice doc for another brand.

Taxonomy weights

JSCG runs 90% Project Showcase. A coach runs mostly Teaching Packages.

Database bindings

Each brand’s Content Feed / People / Proof / Toolbox equivalents. The schema itself becomes the template.

Target-phrase registry

Per brand, per market, hyper-local for service businesses.

New brand, fill four slots, zero engine changes.