Blend Labs Rx :: Content Engine Operating Manual

Contents

Open decisions and one thing that will not work
Roles, the approval rule, and where files live
The pipeline at a glance
Stage 1. Delivery intake from Esteban
Stage 2. Topic breakout and launch calendar
Stage 3. Transcription
Stage 4. Video copy build with Ahrefs
Stage 5. Upload and schedule
Stage 6. Editorial scoping and drafting
Stage 7. Figure build in Claude Design
Stage 8. Editorial publishing
Stage 9. Social amplification
The claims rule set
Measurement and reporting
Risk register

Section 01

Open decisions and one thing that will not work

Resolved: figures are built in Claude Design, not generated

Graphics here means charts, tables, decision paths, and comparison formats. Digestible containers for information. Claude Design is the correct tool and it closes the risk raised in the last version: a generative image model cannot compute, so asked for a chart it renders something chart shaped with invented numbers and garbled labels. Built figures carry real values.

The consequence nobody has drawn yet is about ownership. Your standing rule is that Kai writes no copy. A chart's axis labels, row headers, and cell values are copy. If Kai builds figures by prompting Claude Design, Kai is generating copy, which breaks the rule the whole governance model rests on. Jo builds the figures. Full reasoning in Section 10.

Still unanswered from the last round

#	Decision	Why it blocks work	Owner
1	Pet parent, professional, or both	The platform list is YouTube, Instagram, Facebook. That is a pet parent mix, and the graphics brief says pet parent focus, so the working assumption is consumer. Blend's revenue is prescription driven and the vet writes the script. LinkedIn is still absent and it is where the buyer is. If this is a consumer brand play, say so and judge it as one. If vet acquisition is the goal, the platform mix is wrong.	Serge
2	Cadence against actual runway	One topic per week is a commitment made before the topics have been counted. Count the delivered videos at Stage 1, divide, then commit. Publishing slower than capacity is recoverable. Going dark is not.	Serge
3	Compliance adjudicator and response window	Jo runs the approval gate but is not a pharmacist. The claims rule set in Section 13 covers the majority. Michael has to own the residue and commit to a response time, or assets sit.	Michael
4	One primary call to action per editorial	Pet parent editorials should end on the formulary or the prescription upload. Professional editorials on the Lunch and Learn. Pick one per track. A page with three equal actions converts on none of them.	Serge
5	Measurement baseline before week one	Four of five twelve month review metrics on this engagement already have no baseline. Snapshot the numbers, set the UTM convention, wire the GA4 events. Half a day now, or an argument in month six.	Jo
6	Scrub the prohibited phrase from the brand reference	The design system README lists "office-use formulations" as a stated value proposition. It is the document Kai will design from. Left in place it gets reproduced in good faith, forever.	Serge

Two arithmetic problems in the process as written

Editorial production outruns editorial publishing

One topic per week, editorial count set by what the transcript supports, publishing one to two per week. If a topic yields three editorials and two publish, the backlog grows by one a week. After eight weeks there are eight unpublished editorials and the oldest is two months behind its video, which breaks the whole point of pairing them.

Rule. Editorials produced per topic must equal editorials published per week. Cap at two. If the transcript genuinely supports a third, it goes into an _EVERGREEN_BACKLOG folder labelled as filler for a thin week, not into the schedule. Do not let a backlog form by accident and then discover it.

Kai's load has roughly doubled

Kai now owns transcription, all uploads, all scheduling, two to three graphics per editorial, WordPress formatting, and social posting. Transcription alone on a 25 minute video with veterinary terminology is not a fifteen minute task once it has been corrected.

Action. Time the first topic end to end and write down the hours. Set the cadence against the measured number rather than the assumed one. Nobody has stated Kai's weekly availability and the whole schedule rests on it. Confidence: high that this is the binding constraint

One correction to my previous version

I argued that Ahrefs should gate topic selection. Given the footage is already shot, that was overreach. You cannot select a topic you have no video for. Topic selection is correctly driven by what Esteban delivered, and Jo watching it. What Ahrefs should still do at Stage 2, in about twenty minutes per topic, is pick the angle within the topic and confirm the topic is not a dead search space before a week is spent on it.

Section 02

Roles, the approval rule, and where files live

Person	Owns	Does not do
Esteban AudioVision	Delivers finished videos as Google Drive links	Has no role after delivery. Is not a source of truth for what is current.
Jo Ambrose	Intake, topic breakout, launch calendar, Ahrefs, all copy, all editorial drafting, citation verification, figure build in Claude Design, and the sole approval gate before anything goes live	Does not cut video, upload, or format WordPress. Does not adjudicate clinical claims.
Kai	Transcription, uploads, scheduling, WordPress formatting, social sizing of Jo's figures, social posting	Writes no copy, and a figure label is copy. Chooses no titles. Builds no figures. Does not improvise when something is missing. Publishes nothing without a written approval line.
Michael Altman Blend	Clinical and compliance authority, on camera talent, approves the claims rule set and the boilerplate once	Is not in the weekly loop. Escalation only.
Serge Ambrose	Strategy, cadence, arbitration, the six open decisions	Does not review individual assets. If Serge is looking at thumbnails, something upstream has failed.

The approval rule

1. Kai marks an asset Ready for review in the topic's publishing doc.

2. Jo reviews against the checklist and writes a line: date, initials, version number, and either Approved or Return with the specific change required.

3. Only assets carrying an Approved line naming the version about to be published may go live.

4. Any edit after approval voids the approval. It goes back through the gate.

5. If Jo is unavailable, publication waits. No delegated approver, no auto approval on a timeout. A late post costs nothing. An unapproved post costs a regulatory conversation.

The shared drive

Blend Labs Rx owns and pays for the shared drive. That is the right arrangement: the client owns their content library, and if the engagement ever ends the assets do not leave with the agency. Two consequences to handle deliberately.

Permissions are set once, at the top level. Jo and Kai get edit. Michael and Sam get view. Nobody gets link sharing set to anyone with the link, because a video of the CEO discussing medications should not be publicly reachable before it is published.
Ambrose keeps its own copy of anything it needs for its own records, particularly the publishing docs and the reporting sheet. Not the video masters, those belong on Blend's drive.

Folder structure

Blend Labs Rx Content   (shared drive, Blend owned)
│
├── _INVENTORY.gsheet                one row per delivered video, all topics
├── _CLAIMS_RULE_SET.gdoc            Michael approved, the operative document
├── _REPORTING.gsheet                one row per published asset
├── _EVERGREEN_BACKLOG/              editorials produced beyond the publish rate
│
└── 2026-W38_Cats-and-pills/         one folder per topic, per week
    ├── _TOPIC_BRIEF.gdoc            Jo. The thesis, the angle, launch dates
    ├── _PUBLISHING_SCHEDULE.gdoc    Jo to Kai. All copy, all dates, approval log
    │
    ├── 01_LONG_why-cats-refuse-pills/
    │   ├── BLRX_2026-W38_LONG_01_v1.mp4
    │   ├── TRANSCRIPT.gdoc                    Kai
    │   ├── EDITORIAL_01.gdoc                  Jo, includes graphic placement marks
    │   └── graphics/
    │       ├── BLRX_2026-W38_FIG_01_1x1_v2.webp    2000x2000, editorial and email
    │       ├── BLRX_2026-W38_FIG_01_4x5_v2.png     1080x1350, IG and FB feed
    │       ├── BLRX_2026-W38_FIG_01_9x16_v2.png    1080x1920, stories
    │       └── BLRX_2026-W38_FIG_02_...
    │
    ├── 02_SHORT_flavour-options/
    │   ├── BLRX_2026-W38_SHORT_02_v1.mp4
    │   └── TRANSCRIPT.gdoc
    │
    └── 03_SHORT_transdermal-gel/
        └── ...

Version numbers are the only version control you have

A Drive link does not change when the file behind it does. If a revised cut is uploaded over the same file, the link is identical, the filename may be identical, and every approval that referenced it is now meaningless with no record that anything happened. Version numbers only ever increment and nothing is overwritten. v3 that gets edited becomes v4. The approval line names a version for exactly this reason.

Section 03

The pipeline at a glance

Nine stages. The structural feature that makes the schedule work is the fork after Stage 4: Kai uploads and schedules the video while Jo drafts the editorial. Those two things happen at the same time, not one after the other, and the calendar assumes it.

Figure 1

Topic pipeline, delivery to published editorial

Two artifacts carry the whole process. The topic brief tells everyone what the week is about. The publishing schedule doc is the only briefing Kai receives, and if something is not in it, it does not happen.

Figure 2

Two week rhythm, prep week and live week

Section 04

Stage 1. Delivery intake from Esteban

Version control now runs through Drive links Esteban sends before anything is uploaded. That works only if there is an intake step, because a shared link is not custody and it carries no version number.

Two exposures this step closes

Custody. If the masters live in Esteban's Drive and the team works from his links, Esteban controls Blend's content library. He can move a file, revoke a share, or lose access to the account. The AudioVision commercial relationship is already flagged as a risk on this engagement. A soured vendor relationship should cost you a vendor, not the footage.

Silent replacement. A Drive link does not change when the file behind it does. If a revised cut is uploaded over the same file, the link is identical, the filename may be identical, and every approval that referenced it is quietly worthless with no record that anything happened.

Intake checklist, every delivery Jo

Open the link and watch the file start to finish. Not a scrub through. This is also the compliance pass, so it has to be watched.
Copy the file into the Blend shared drive. Make a copy, never add a shortcut. A shortcut still points at his file.
Rename to the convention with a version number, on intake, before it is used for anything.
Confirm it is a final export and not a watermarked review copy or a low resolution proxy. Review exports look finished and are not.
Add a row to _INVENTORY.gsheet: asset ID, duration, aspect ratio, who is on camera, provisional topic, compliance status, and one line on what changed from the previous version.
Assign compliance status now: Clear, Edit required with a timecode and the exact phrase, or Unusable.
Check rights: music licensed, no unreleased third party on camera, no identifiable clinic, client, or patient without consent. Any failure is Unusable pending resolution, not Clear.
Check audio. Bad audio is unfixable and disqualifies an asset regardless of how good the content is.
Reply to Esteban confirming receipt and naming the version now on file, so both sides are looking at the same number.
Never forward Esteban's original link to Kai. Kai works only from shared drive paths.

Settle this in writing while the relationship is functional

Confirm who owns the project files, not just the exports. Delivered MP4s are the output. The project files, graded sequences, caption files, and raw footage are what you need if someone else ever has to pick up the work. Asking now is a request. Asking after a dispute is a negotiation.

Section 05

Stage 2. Topic breakout and launch calendar

Jo watches the delivered videos and groups them: one long video plus its associated shorts becomes one topic, and one topic becomes one week. The folder is built, the dates go on, and from that point the week has a shape everyone can see.

The twenty minute Ahrefs check

Topic selection is correctly driven by what was shot. You cannot pick a topic you have no footage for. What Ahrefs should still do here, before a week is committed to a topic, is confirm the topic is not a dead search space and pick which angle within it leads.

Run matching terms and related terms on two or three seeds from the video.
Look for the operational and formulation long tail, not the head term. "Flavoured methimazole for cats" is worth more than "hyperthyroidism in cats" even at a tenth of the volume, because the intent is a purchase and the competition is nobody.
Run SERP overview on the strongest candidate. If Merck Veterinary Manual, VCA, PetMD, or Cornell own the top five, that term is not winnable and the angle changes. This is not pessimism, it is a fact about domain authority in veterinary informational search.
Write the chosen angle and the primary keyword into the topic brief. Stage 4 builds on it rather than starting over.

Topic brief template

_TOPIC_BRIEF.gdoc

Topic ID        2026-W38_Cats-and-pills
Live week       Week of 15 September 2026
Videos          01_LONG_why-cats-refuse-pills      18:42
                02_SHORT_flavour-options           0:38
                03_SHORT_transdermal-gel           0:44

THE WEEK IN ONE LINE
[What this topic is about and why a pet owner should care. One sentence.]

TRACK          Pet parent / Professional
PRIMARY KEYWORD
SECONDARY      x2
ANGLE          [Which framing leads, and what Ahrefs said about why]
SERP NOTE      [Who owns the head term, and what is being targeted instead]

COMPLIANCE
[Anything flagged during intake, with timecodes, that must not survive into a cut,
a title, a caption, a graphic, or an editorial.]

EDITORIAL COUNT   1 or 2. Never 3 into the schedule.
LAUNCH DATES      Long video, each short, each editorial, with times.

Stage 2 checklist Jo

Every delivered video is assigned to a topic. Nothing is left unallocated in the inventory sheet.
Topic folder created in the shared drive, named to convention, with one sub folder per video.
Video files moved into their sub folders. Not linked, moved.
Ahrefs check run. Primary keyword, two secondaries, and the SERP note written into the brief.
Head terms owned by authority sites explicitly ruled out, with a note saying so, so nobody re-litigates it in Stage 4.
Editorial count decided and capped at two. Anything beyond goes to _EVERGREEN_BACKLOG, deliberately, not by drift.
Launch dates and times set for every asset, long video, shorts, and editorials.
Compliance flags from intake carried into the brief with timecodes so they cannot be forgotten downstream.
Runway checked: how many topics remain unallocated. If fewer than four, raise it with Serge now rather than in the week you run out.

Section 06

Stage 3. Transcription

Kai transcribes every video in the topic folder and returns each transcript to its own sub folder. The transcript is the input to all copy and all editorial, so an error here propagates into everything downstream.

Standard

Verbatim. Filler words can go. Meaning cannot be adjusted or tidied.
Timecoded every fifteen seconds or tighter. A transcript without timecodes cannot be used to cut anything or to write chapters.
Speaker labelled where more than one person appears.
Saved as TRANSCRIPT.gdoc inside the video's own sub folder. One per video, not one per topic.

The drug name pass

Automatic transcription reliably mangles veterinary drug names. Every one of these comes back wrong on a regular basis:

pimobendan methimazole mirtazapine cyclosporine gabapentin ondansetron enrofloxacin trilostane maropitant

A drug name that is wrong in the transcript will be wrong in the title, the description, the editorial, and the graphic. Kai searches the transcript for every drug mentioned and confirms the spelling before handing it back.

Stage 3 checklist Kai

One transcript per video. Every video in the topic folder, not a sample.
Timecodes present at fifteen second intervals or tighter.
Speaker labels applied where needed.
Every drug name checked and corrected by hand.
Saved into the correct video sub folder with the correct filename.
Flagged to Jo if anything in the audio is inaudible, with the timecode, rather than guessed at.

Then, before Stage 4 Jo

Read every transcript against the claims rule set in Section 13. Mark the timecode of anything that reads as an efficacy, safety, comparative, turnaround, or capacity claim.
Update the compliance status in the inventory sheet from what the transcript actually says, not from the Stage 1 estimate.
Anything the rule set does not clearly cover goes to Michael and the asset waits. Update the rule set with his answer so the same question is never escalated twice.
Load all transcripts for the topic into Claude together, as one context, not one at a time.

Section 07

Stage 4. Video copy build with Ahrefs

Jo produces every word that will appear on every platform, plus the dates and times, in one document. That document is the only briefing Kai receives. If something is not in it, it does not happen, and Kai stops and asks rather than filling the gap.

The reason for that rigidity is not distrust. It is that every improvised decision, however sensible, is an unapproved decision, and the value of an approval gate is that it has something fixed to check against. A brief with gaps produces a gate that cannot fail anything.

Copy rules by platform

Platform	Title	Description or caption
YouTube long form	Primary keyword in the first four words. Under 60 characters or it truncates. Sentence case. No hook the video does not pay off, because a mismatched title costs retention, which costs more than the extra click was worth.	First 157 characters are what shows above the fold, so the keyword and the reason to watch go there. Then the editorial link with UTM, then chapters, then the standing about block, then the approved boilerplate. Three hashtags maximum.
YouTube Shorts	Under 40 characters.	Short. Editorial link with a distinct UTM. Abbreviated boilerplate, approved separately by Michael.
Instagram	Not a field. The first line of the caption is the title in practice and it is the only line most people read.	Hook on line one. Links are not clickable, so point to the link in bio and make sure the link in bio is actually current. Three to five hashtags at the end, not fifteen.
Facebook	Same, the first line does the work.	Links are clickable here, unlike Instagram. Use the editorial URL with a distinct UTM. Shorter than the Instagram caption.

Two rules that are easy to break and expensive to break

The editorial slug is locked here and never changes. The video description contains the final editorial URL before the editorial exists. That is what lets the video publish first without a broken link that matters. If Kai adjusts the slug in WordPress because it read better, the description link dies permanently and nobody notices for a month.

The compliance boilerplate is never edited per video. Michael approves one block and it is pasted verbatim. The moment somebody tailors it to fit a specific video, it stops being an approved block and becomes new copy nobody approved.

The publishing schedule doc

_PUBLISHING_SCHEDULE.gdoc          Jo to Kai. The only briefing.

1  ASSET MANIFEST        Asset ID | Type | Drive path | Destination | Status
2  COPY BLOCKS           COPY AND PASTE ONLY. DO NOT EDIT.
      2.1  YouTube long form title
      2.2  YouTube long form description, with chapters and boilerplate
      2.3  YouTube Shorts title and description, per short
      2.4  Instagram caption, per asset
      2.5  Facebook caption, per asset
      2.6  Editorial title, meta description, and LOCKED SLUG
      2.7  Alt text for every image
3  THUMBNAIL BRIEF       Frame timecode to pull, text to set, colour pairing
4  FIGURE MANIFEST       Per editorial. Type, native-table flag, alt text, placement
                        mark, and the three export paths (1x1, 4x5, 9x16).
                        Built by Jo, not briefed to Kai.
5  SCHEDULE              Date | Time ET | Asset | Platform | Confirmed by Kai
6  APPROVAL LOG          Asset ID | Version | Decision | Date | Initials | Change
7  PUBLISHED LOG         Asset ID | Live URL | Verified by | Time

Stage 4 checklist Jo

Primary and two secondary keywords written in with volume and difficulty from Ahrefs, not from memory.
SERP overview run on the primary. If the top five are unbeatable, the target changed rather than attempted anyway.
A separate copy block written for every platform. Not one block reused three times with the hashtags swapped.
Editorial slug locked and marked in bold as immutable.
UTM strings built per platform, each with a distinct utm_source and utm_medium.
Chapters written with accurate timecodes taken from the transcript.
Compliance boilerplate pasted unaltered, long and abbreviated versions.
Every block scanned for prohibited claims, the phrase "office use", em dashes, and en dashes.
Every date and time in section 5 filled in. No "TBC" anywhere in a document handed to Kai.
Doc shared to Kai with comment access, not edit access. Kai comments, Jo changes.

Section 08

Stage 5. Upload and schedule

Kai uploads everything and schedules it, unpublished, to the times in the doc. Nothing goes live at this stage. The gate sits between staging and publication for a reason.

Stage 5 checklist Kai

Long video uploaded to YouTube as Private or Scheduled, never Public.
Each short uploaded natively to each platform. Four uploads of the same file, not one cross post. Cross posting through Meta's share tools produces a watermarked, downranked, lower quality post on the second platform. It saves five minutes and costs most of the reach.
No watermark from any editing tool on any file.
Every title, description, and caption pasted from the doc, character for character. Nothing retyped, because retyping introduces changes.
Thumbnail uploaded and rendering, not left as an auto generated frame.
Chapters entered with the timecodes from the doc.
Every scheduled time matches section 5 of the doc exactly, including time zone.
Captions burned in on every vertical and legible on a phone.
Section 5 marked Confirmed by Kai for each row.
Asset marked Ready for review in the approval log with the version number.

Gate 1. Jo, before anything publishes

Published copy matches the doc word for word
No efficacy, safety, comparative, turnaround, or capacity claim anywhere
The phrase "office use" does not appear in any form
Dr. Lake-Bakaar, where named, is attributed as an independent board certified cardiologist and not as Blend staff or an endorser
Nothing flagged in the transcript triage has survived into a cut, a title, or a caption
Boilerplate present and unaltered
No em dashes and no en dashes
Drug names spelled correctly, checked one by one
Editorial URL in the description matches the locked slug exactly
UTM strings present, correct, and distinct per platform
Aspect ratio correct per destination, captions accurate, audio levelled, no watermark
Thumbnail readable at 15 percent zoom, contrast on the approved pairing list
Scheduled times correct on every platform
Approval line written naming the version number about to publish

Section 09

Stage 6. Editorial scoping and drafting

This runs at the same time as Stage 5, and it runs straight into Stage 7, because the figures are built in the same sitting as the draft. The editorial is the only asset in this system that Blend actually owns. YouTube, Instagram, and Facebook are rented distribution that can change their terms tomorrow. The editorial sits on blendlabsrx.com, accrues authority, and is still working in three years. Weight the effort accordingly.

How many editorials a transcript supports

The honest answer is usually fewer than it feels like. A twenty minute video has three or four genuinely distinct arguments in it and a lot of connective tissue. Slicing it into five articles produces five thin articles that compete with each other for the same keyword, which is worse than one good article, because now the site is cannibalising itself.

The test. An editorial earns its own page if it targets a keyword no other Blend page targets, and if it can be read by someone who has not watched the video. If two candidate editorials would both rank for the same query, they are one editorial.

Broader than the video, as specified. The video answers a specific question. The editorial takes the wider query around it and uses the video as the evidence inside. Video: "why cats spit out pills." Editorial: "what to do when your cat will not take medication," with the video embedded at the point where it answers part of it.

Structure

Element	Specification
H1	The question the reader typed, or as close as reads naturally. Contains the primary keyword.
Lede	Two to three sentences that answer the question immediately. Do not build up to it. Someone with a cat that will not take a pill has no patience for a preamble.
Video embed	High on the page, after the lede. Lazy loaded so it does not destroy page speed.
H2 sections	Three to five. Each one is a thing the reader can do or understand, not a heading like "Introduction."
Figure marks	Two to three, written into the draft as `[FIG 01: what it shows and why it goes here]`. Jo builds the figure at Stage 7. Kai places it at the mark and nowhere else.
Pull quote	One, set in Fraunces. That is the only place Fraunces appears anywhere in the brand. Use it once or not at all.
Citations	Primary sources only. Peer reviewed veterinary journals (JAVMA, JVIM, AJVR, Veterinary Record) or institutional sources (AVMA, ACVIM, FDA, USP, NABP, teaching hospitals). Never a competitor's blog, which hands them a link and cites marketing as evidence.
Internal links	Two to four, placed contextually, to the formulary, a relevant product page, or the prescription upload. Not a link dump at the bottom.
Call to action	One primary, per Decision 4. One styled block.
Length	900 to 1,400 words. Word count is not a ranking factor and padding is visible.

The failure that will eventually happen if this rule is not enforced

Every citation produced by an AI tool must be opened, read, and confirmed to exist and to say what the article claims it says. Every time.

Language models generate citations that are plausible in every respect and entirely fictional. Correct journal, correct author names, correct formatting, no such paper. A veterinary compounding pharmacy publishing a fabricated citation to a clinical study is materially worse than publishing nothing, and it is exactly the kind of thing a competitor notices and keeps. Jo verifies at the gate. This does not get skipped when the week is tight.

And the distinction that carries the whole compliance position: cited literature supports the disease or drug class context. It never supports Blend's preparation. No sentence may imply that a study evaluated something Blend compounded. Dr. Lake-Bakaar's work is cited as her independent research, presented by her, attributed by name and credential, and never as Blend research or a Blend claim.

Stage 6 checklist Jo

Editorial count matches the publish rate. One or two. A third goes to _EVERGREEN_BACKLOG, deliberately labelled as filler.
Each editorial targets a keyword no other Blend page targets. Cannibalisation checked against the site's existing organic keywords in Ahrefs.
Written to be readable by someone who never watches the video.
Primary keyword in the H1, the lede, and at least one H2, used naturally rather than stuffed.
Every claim checked against the rule set in Section 13.
Every citation opened and read, with a note in the publishing doc confirming it.
No citation framed as having evaluated a Blend preparation.
Two to three figure marks written into the draft, then built at Stage 7 in the same sitting.
Two to four internal links placed contextually.
One primary call to action.
Meta description under 155 characters with the primary keyword.
Alt text drafted for every graphic, descriptive rather than keyword stuffed.
Read out loud. If a sentence is hard to say, it gets rewritten.
Scanned for em dashes and en dashes.
Saved as EDITORIAL_01.gdoc in the relevant video's sub folder.

Section 10

Stage 7. Figure build in Claude Design

Charts, tables, decision paths, comparison formats. Digestible containers for information. Claude Design is the right tool, and it closes the risk from the previous version: a generative image model cannot compute, so it renders chart shaped pictures with invented numbers. A built figure carries real values.

Jo builds the figures, not Kai

This follows from your own rule rather than from a preference. Kai writes no copy, and a figure's labels are copy. Axis labels, row headers, column headers, cell values, the text in a decision node: all of it is published wording that carries the same claim risk as a caption. If Kai builds figures by prompting Claude Design, Kai is generating copy through a chat box, which is the exact thing the governance model exists to prevent, just harder to notice because it looks like design work.

Jo builds them in the same pass as the editorial draft. She has the transcript loaded, she is already writing the article, and the figure content comes from the same source. Stage 6 and Stage 7 are one sitting.

This does not add to Jo's load the way it looks like it does

Writing a specification complete enough for someone else to build a chart, every label, every value, every row, in a document, takes longer than building the chart. Moving the build to Jo removes the spec writing, which was most of the work. Net effect on Jo is close to neutral. Net effect on Kai is a clear reduction, which matters given her scope already includes transcription, uploads, WordPress, and social. Confidence: moderate to high

Export specification

Ratio	Pixels	Used for	Minimum type size	Notes
1:1	2000 x 2000	Editorial, newsletters	75px on the source canvas	Export as WebP or compressed JPG under 300 KB. PNG only if transparency is genuinely needed.
4:5	1080 x 1350	Instagram and Facebook feeds	36px	The densest shape that still works. Design here first.
9:16	1080 x 1920	Stories	44px	Keep content inside the central band: 250px clear at the top, 320px clear at the bottom, or platform UI covers it.

Where the 75px minimum comes from, because it will look absurd in the editor

The editorial's narrow container is 880px, so a 2000px figure displays at roughly 0.44 scale on desktop. That is not the constraint. On a phone at a 390px viewport with 16px gutters, the figure displays at about 358 pixels wide, which is a scale of 0.18. Type set at 40px on the source canvas renders at 7px on a phone. It is unreadable, and it will look completely fine to whoever built it on a monitor.

The practical consequence. A 2000 x 2000 figure that has to survive a phone can hold roughly eight to twelve lines of real content, and no more. Anything denser than that is not a figure, it is a table, and it goes into the article as a native table that reflows rather than as an image that shrinks.

Design once, reflow twice

Three ratios per figure, two to three figures per editorial, is six to nine layouts a week if each one is built from scratch. It does not need to be.

Build at 4:5 first. It is the tightest useful shape, so content that fits there fits everywhere.
1:1 is the 4:5 layout on a wider canvas, with the type scaled up to clear the 75px floor and the spacing opened out. Not a stretch, a re-set on a bigger stage.
9:16 is the 4:5 layout with vertical breathing room, pushed into the central safe band. If a figure has more than four branches or rows, split it across two or three story frames rather than compressing it.

Only re-lay-out properly when the content genuinely will not survive the reflow. Most decision paths and comparison tables will.

Reels is a video surface, not an image surface

A static figure cannot be posted as a Reel. Instagram Reels requires video, and Facebook treats a static upload the same way. If a 9:16 figure is intended for Reels rather than Stories, it has to become video: a simple build on, one element appearing at a time over four to six seconds, exported as MP4.

Recommendation. Send figures to Stories only and keep Reels for actual video cut from the long form. Animating a table to satisfy a slot is effort spent on the wrong thing, and a Reel that is visibly a still image with a fade gets swiped in under a second, which teaches the algorithm the wrong lesson about the account. Confidence: high on the platform constraint, high on the recommendation

Which figures become images at all

The three ratios are the social and newsletter deliverables. Inside the editorial itself, the format depends on the figure type, and this is the distinction that decides whether a figure works on a phone.

Figure type	In the editorial	Exports
Comparison table	Native WordPress table. Responsive, indexable, readable, selectable, and it reflows instead of shrinking.	All three ratios, for newsletter and social.
Chart	SVG if the theme accepts it, otherwise the 1:1 export at full content width.	All three ratios.
Decision path or process flow	SVG preferred. These break worst as raster because the type is small by nature.	1:1 and 4:5. Split across frames for 9:16.
Checklist graphic	Native list markup. It is text.	All three ratios.

Page weight

Three 2000 x 2000 PNGs in one article can add several megabytes and will undo the compression work already done on the site bundle. Export WebP or compressed JPG, target under 300 KB each, and confirm the total page weight after the first editorial rather than after ten of them.

Six figure formats that are safe, useful, and pet parent shaped

Most of what a pet owner needs is orientation, not statistics. These carry no claim risk because they describe what exists and what happens, never what a preparation does to an animal.

Format	What it shows	Why it is safe
Decision path	"Your cat will not take the pill" branching to the dosage forms that exist: flavoured liquid, transdermal gel, tiny tab, soft chew, with one line on how each is given.	Describes available forms. Says nothing about outcomes.
Dosage form comparison	Form, how it is given, what kind of animal or situation it suits. Three columns.	Formulation facts only.
How compounding works	Four steps from prescription to delivery. Steps only, no durations.	Turnaround time claims are prohibited. Show the sequence, never the clock.
Flavour grid	Species down one axis, available flavours across the other.	A menu, not a claim.
What your vet needs to send	A checklist the owner can hand to their clinic.	Operational, and it removes friction from the actual conversion path.
Mechanism or anatomy diagram	Only where drawn from a cited source, with the source named on the figure itself.	Attributed education, not a Blend assertion.

If a figure has a number on it, it has a source on it

Any figure carrying a statistic shows where the number came from, on the figure, in small type. If the number is Blend internal data that cannot be independently verified, it does not go on a figure. Note that five of the six formats above carry no numbers at all, which is the cleanest way to avoid the problem entirely.

Tokens to give Claude Design

Type          Inter. 700 for headings and labels, 400 and 500 for body.
              Negative tracking on large sizes. Sentence case, never title case.
Colour        #2e9381 accent   #1e3f38 dark surface   #cfe6df borders on tint
              #e3f1ec tint fill   #000000 text   #555855 secondary text
              #ffffff paper   #f4f5f2 alternate surface
Contrast      White on #1e3f38, or black on #ffffff or #e3f1ec.
              Never dark text or icons on green.
Radius        4px chips and cells, 8px cards and panels.
Prohibited    Circular icon badges. Shadows. Gradients. Illustration.
              The brand is flat and has no illustration system.
Logo          Supplied PNG only. Never re-set the wordmark in another typeface.

Where Higgsfield sits now

Nowhere in this pipeline. If figures are charts and tables built in Claude Design, and thumbnails pull a real frame from the video, there is no remaining job for a generative image model here. Do not find one for it. Having a tool is not a reason to route work through it, and in a regulated category the downside of a generated image that reads as clinical is much larger than the upside of a nicer background.

Stage 7 checklist Jo builds and approves

Two to three figures per editorial, built in the same sitting as the draft.
Every value and label taken from the transcript or a cited source, not invented to fill the shape.
Every number has its source shown on the figure.
No durations, no turnaround times, anywhere.
Claims rule set applied to every label. A cell in a table is published copy.
Colour values are exact tokens. Contrast on the approved pairing list, checked on every figure.
No circular badges, no shadows, no gradients, no illustration.
Opened on an actual phone at actual size and read. Not zoomed out on a monitor, which is how every illegible figure in history got approved.
Exported at all three ratios: 2000 x 2000, 1080 x 1350, 1080 x 1920. Reflowed, not stretched.
Type clears the floor for each ratio: 75px at 2000, 36px at 1080 x 1350, 44px at 1080 x 1920.
The 9:16 version keeps 250px clear at the top and 320px at the bottom for platform UI.
Editorial exports are WebP or compressed JPG under 300 KB. Not PNG unless transparency is needed.
Tables are marked in the publishing doc as rebuild as native table, so Kai does not paste an image.
Alt text written for every figure, describing what it shows rather than repeating the title.
Saved to the video's graphics/ folder, named with the aspect suffix and versioned: BLRX_2026-W38_FIG_01_1x1_v2.webp.
Placement mark written into the editorial draft as [FIG 01: what it shows and why it goes here].

Section 11

Stage 8. Editorial publishing

Kai builds the editorial in WordPress as a draft, places the graphics at the marks Claude specified, and stops. Gate 2 sits between the draft and the publish button.

WordPress build Kai

Title and slug entered exactly as specified. The slug is not adjusted for any reason, including if it reads wrong.
Body pasted from the doc. Nothing retyped.
Figures placed at the [FIG] marks and nowhere else. If a mark seems wrong, Kai comments and asks, she does not move it.
Any figure flagged as a table is rebuilt as a native WordPress table, never pasted as an image. Values copied from Jo's figure exactly. An image of a table is unreadable on a phone and invisible to search.
Charts and diagrams inserted as SVG where the theme accepts it, otherwise the 2000 x 2000 WebP or JPG export at full content width. Never a PNG over 300 KB.
Page weight checked after insertion. Three large figures can undo the compression work already done on the site.
Alt text entered for every figure from the doc.
Meta description entered in the SEO plugin field, not left to auto generate.
Heading hierarchy correct. One H1, H2 for sections, H3 only nested under an H2.
Video embedded with the theme's embed block, lazy loaded.
Internal links inserted where the doc specifies, not wherever they happen to fit.
External citation links open in a new tab with rel="noopener".
Category and tags assigned to the topic taxonomy.
Article and video schema applied if the SEO plugin supports both.
Linked from the Knowledge Center index. No orphan pages.
Mobile preview checked before scheduling, not after publishing.
Left as a scheduled draft. Draft URL posted into the approval log.

Gate 2. Jo, before the editorial goes live

Body matches the approved draft word for word
Every citation link opened and confirmed to resolve to the right paper
No claim, no "office use", no em dashes, no en dashes
No citation implies evaluation of a Blend preparation
Lake-Bakaar attribution correct where present
Tables are native tables, not images. Confirmed by trying to select the text
Charts render at full width and are legible on a phone
Every figure with a number shows its source
Video embed plays
Slug matches the locked slug. Load the video description link and confirm it lands here
Internal links resolve. External links open in a new tab
Meta description present and under 155 characters
Alt text on every image
Appears in the Knowledge Center index
Primary call to action present, and only one
Approval line written naming the version

Within 30 minutes of go live Kai

Page loads on a phone, on cellular data, not only on wifi
Every link on the page clicked
Video embed plays on mobile
Graphics legible without pinch zoom
Call to action clicked and the destination loads
Live URL written into the published log
Jo confirms the version published matches the version approved, and sends the log to Sam

Section 12

Stage 9. Social amplification

Jo's figures already exist at 1080 x 1350 from the Stage 7 dual export. This stage costs almost nothing because the work was done, which is the whole reason every figure is exported twice.

Format	Size	Use
Carousel	1080 x 1350 4:5, 5 to 8 frames	The editorial's argument, one point per frame. Highest saved format on Instagram. Frame one is the hook and it carries the whole cost of the post. Final frame is the call to action.
Single figure	1080 x 1350 4:5	A Stage 7 figure posted on its own with a caption pointing to the editorial. Cheapest post in the system and often the best performing, because a useful table gets saved.
Quote card	1080 x 1350 4:5	One line from the video or from a named independent expert. Attribution exact and mandatory.
Story frames	1080 x 1920 9:16, 3 to 5	Drives to the editorial or the video. Link sticker where the account is eligible, placed inside the safe band. Figures go here, never to Reels.

Every frame is published copy and goes through the same rule set as everything else. A carousel frame is not exempt because it is short.

Stage 9 checklist Kai builds Jo approves

Frame copy taken verbatim from the publishing doc. Kai does not edit any text on a figure, including to make it fit.
Frame one works as a standalone image with no context.
Final frame carries one call to action.
Claims rule set applied to every frame independently.
Contrast pairing checked on every frame. No dark on green.
No circular badges, no shadows, no gradients, no illustration.
Exported at correct dimensions, sRGB, under platform file size limits.
Native upload per platform. Never a cross post share.
Scheduled to the times in the doc.

Section 13

The claims rule set

The operative document. It governs transcripts, titles, descriptions, captions, editorials, graphics, and social frames. Michael approves it once, and after that Jo applies it without escalation for anything it clearly covers. It lives at the root of the shared drive and is updated every time Michael answers a new question, so the same question is never escalated twice.

Permitted. Formulation and operational.

What dosage forms Blend compounds, and in what flavours
What compounding is and how it works as a practice
How to order, refill, transfer, or submit a prescription
The practical problem in plain language: pets refuse pills, cats hide medication, dosing a small animal accurately is hard
General disease and drug class education, attributed to published literature or to a named independent expert
Documented licensure and accreditation: NABP VPP, USP 795 and 800, state licensure
Dr. Lake-Bakaar's findings, attributed to her by name, as an independent board certified cardiologist

Prohibited. No softened versions.

Any statement that a compounded preparation treats, cures, prevents, manages, or improves a condition
Any comparative efficacy, bioavailability, or bioequivalence claim against an approved product
Any safety claim. "Safe for", "gentle on", "no side effects", "well tolerated"
Turnaround time, shipping speed, volume, or capacity claims that are not independently verifiable
The phrase "office use" in any construction
Any framing implying a cited study evaluated a Blend preparation
Any statement attributed to Dr. Lake-Bakaar she did not make, or any framing of her as Blend staff, advisor, or paid endorser
Language referencing lulls, slow periods, or dead time. Momentum framing only
Em dashes and en dashes anywhere in published copy

The test Jo applies to any sentence she is unsure about

Rewrite the sentence with the subject changed from the condition to the preparation. If it now reads as a statement about what the medicine does to an animal, it is a claim and it does not go out. If it reads as a statement about what the pharmacy does to the medicine, it is a formulation fact and it is fine. "Compounded into a chicken flavoured suspension so it can be given by mouth" passes. "Helps cats with hyperthyroidism get their medication" fails, because the second half is doing clinical work.

Section 14

Measurement and reporting

Four of five twelve month review metrics on this engagement already have no baseline behind them. That is about to be repeated on the content programme unless the instrumentation is built before week one rather than reconstructed in month six.

Before week one

Baseline snapshot. Record every number as it stands today: subscribers, average views per video, average view duration, follower counts, blog organic sessions per month, average engagement time. Date it and file it. A baseline reconstructed later is not a baseline, it is an argument.
UTM convention documented and applied. Every platform gets a distinct utm_source.
GA4 events on the primary call to action on the editorial template. Formulary click, upload click, Calendly click, phone tap.
HubSpot attribution from editorial form fills through to contact records, so a booked session can be traced to a specific article.

Views are to video what open rate is to email. It is the number the client asks for, it is the number that feels like progress, and it predicts nothing.

Metrics that matter

Long form: click through rate, average view duration, average percentage viewed, description link clicks
Shorts: average view duration and the viewed versus swiped away ratio. Retention in the first three seconds is the whole game
Editorial: organic sessions, average engagement time, scroll depth, primary call to action clicks, positions on the targeted long tail terms
Business: prescriptions or enquiries attributable to content, sessions booked from an editorial, content assisted deals in HubSpot

Anti signals. Report, never lead with.

Raw view count
Follower and subscriber totals
Impressions and reach
Likes
Total published asset count

These are the numbers that will get asked about. Report them honestly and put them second. The same discipline that established reply rate over open rate on the outreach programme applies here without modification.

Cadence	Owner	Question it answers
Weekly, 15 min	Jo	Did we ship what we said we would ship, and is anything blocked. Not a performance review.
Day 7 and day 28	Jo	One reporting row per topic. Day 28 matters more than day 7 for anything search driven.
Monthly	Serge	Which topics and formats are working. Does the cadence hold. Is the editorial backlog growing.
Quarterly	Serge	Is this producing commercial outcomes or only producing content. Where Decision 1 gets re-argued with evidence.

Section 15

Risk register

Risk	Likelihood	Impact	Mitigation
A figure is illegible on a phone	High	Moderate	A 2000 x 2000 figure displays at roughly 358px on a phone, a scale of 0.18. It looks correct to whoever built it on a monitor. Type floors are specified per ratio and every figure is opened on an actual phone before approval.
A static figure is posted as a Reel	Moderate	Low	Reels requires video. A still with a fade gets swiped in under a second and teaches the algorithm the wrong thing about the account. Figures go to Stories. Reels stay for video cut from the long form.
A table is published as a flat image	High	Moderate	Unreadable on a phone, invisible to search and screen readers, and it fails silently because it looks fine on the desktop where it was made. Tables are rebuilt as native WordPress tables. Flagged per figure in the publishing doc and checked at Gate 2 by trying to select the text.
Figure labels drift from approved wording	Moderate	High	A cell value is published copy and carries the same claim risk as a caption. Jo builds figures rather than briefing them, and Kai alters no text on a figure for any reason, including to make it fit.
A fabricated citation is published	Moderate	Severe	Every citation opened and read before publication, confirmed in the publishing doc. Manual, non negotiable, and not skipped in a tight week.
A claim survives into a published asset	Moderate	Severe	The rule set is applied at three points: transcript triage, copy drafting, and the gate. Three passes because one pass fails.
Kai's load exceeds her capacity	High	High	Transcription, graphics, uploads, WordPress, and social is a large scope and nobody has stated her weekly availability. Time the first topic end to end and set cadence against the measured number.
Editorial backlog grows silently	High if uncapped	Moderate	Cap editorials per topic at the publish rate. Extras go to the evergreen backlog folder deliberately, labelled as filler.
Masters stay in the vendor's Drive	High until intake runs	Severe	Copy every delivery into the Blend shared drive on receipt. Get written confirmation of who owns the project files and raw footage, not only the exports.
A file is silently replaced and prior approvals become meaningless	Moderate	High	Drive links do not change when the file behind them does. Incrementing version numbers and the intake log are the only defence. Every approval line names a version.
Jo becomes the bottleneck	High	Moderate	Structural, not personal. Jo holds two gates and produces all strategy and all editorial. If gates slip more than twice in a quarter, the answer is a lower cadence or a second reviewer trained to the same checklist, not a relaxed gate.
Content runs out	High if cadence is set before counting	High	Count the delivered topics at Stage 1. Raise it when fewer than four remain unallocated, not in the week you run out. A dark channel decays, it does not pause.
Q4 commitments collide with the calendar	High	Moderate	Michael's CE course launch in October and New York Vet on 29 and 30 October consume the same capacity. Reserve or skip those weeks now rather than discovering the collision in week six.

What to do first, in order

1. Confirm what Claude Design exports and at what fidelity, before the first topic rather than on a deadline. Specifically: does it export SVG, and does it export WebP. If SVG is available, use it for anything in the editorial that has type in it.

2. Set up the Blend shared drive with permissions, and run the intake protocol on everything Esteban has already sent.

3. Michael approves the claims rule set and the boilerplate. One meeting, one document, done once.

4. Serge answers Decision 1 and Decision 4. Track and primary call to action.

5. Time the first topic end to end. Set cadence against the measured hours, not the intended ones.

6. Snapshot the baseline numbers before anything publishes.

The Content Engine