Marketing OS Carousel

One filed asset in, one branded PDF out, and the OS records that the child exists.

This skill carries no brand of its own. No palette, no typography, no logo, no example carousel. All of it lives once in Context/brand/brand-kit.md, which is the executable design contract, and this skill resolves it there every run. The original carousel skill shipped its own copy of the palette and its own logo file, so a rebrand meant editing a skill folder and the two versions silently disagreed until somebody noticed.

And it does not guess its source. The original scanned an inbox for an email somebody described. This one repurposes an asset that is already filed, which is what makes the carousel and its parent provably the same argument.

Run from the OS root. Stay inside that root. Start from zero on identity: the name on the footer, the handle, the mark and every colour come from the OS. If the OS does not name one, you do not know it, and you stop rather than substituting a plausible value.

First, check what you have

State	Do
No Context/config.md	Not a Marketing OS. Point at marketing-os-setup and stop
No Context/brand/brand-kit.md	Stop. There is no brand contract to build against, and inventing one is how brand drift starts
Brand kit exists but names no logo file	Stop and ask for the path. The mark is composited by code and must be a real file
No source named, Channels/ populated	Propose sources from the OS, best candidate first
Higgsfield not connected	Say so and stop before Step 4. This skill cannot render without an image model

Step 1: resolve the brand, once

Read Context/brand/brand-kit.md and pull:

What	Used for
The audience-facing palette	Slide background, ink, and the single highlight colour
The background rule	The brand kit states whether black may be a background on an audience-facing surface. Obey it exactly as scoped
Typography	The heading and mono faces, passed to the footer script
The logo	Its markup for reference, and the pointer to the real raster the footer needs
The portrait pointer	Optional. Only if the brand kit names one

Read Context/config.md for operator_name and org_name, which are the two strings on the footer bar.

Use the audience-facing token set, not the data-surface one. A brand kit may carry two sets on one chassis, one for what an audience sees and one for internal dashboards. A carousel is audience-facing. Never blend them.

Never hardcode a colour. If you find yourself typing a hex that is not in the brand kit you have already made the mistake this skill exists to prevent.

Step 2: resolve the source from the OS

Accept a slug, or propose candidates. Priority order, because it follows the cascade the OS documents:

1. A filed newsletter edition. Channels/{email channel}/broadcasts/YYYY-MM-DD-<slug>.md. Read the ## Edition body and the ## The decisions block. The decisions block is the gift: the angle, the insight and the outcome were already argued out, so the carousel inherits them rather than re-deriving them.
2. A published pillar asset. Channels/{primary}/published/YYYY-MM-DD-<slug>.md. Read its angle, its snapshot and its ## Repurpose tree.
3. A pipeline brief. Channels/{channel}/pipeline/<slug>/brief.md, for something not yet shipped.

Check the parent's repurpose tree before you build. If a carousel child already exists, say so and ask whether this is a second angle or a replacement. Building a duplicate silently is how a cascade turns into noise.

Read alongside the source:

Path	For
Context/personal-brand/voice.md	The master voice register and the banned constructions
Channels/{target channel}/voice.md	The delta for the surface this posts to
Channels/{target channel}/strategy.md	The format contract and what this channel receives
Intelligence/research/swipe/hooks.md	Proven opening shapes for the cover slide
Intelligence/research/swipe/ctas.md	CTA phrasings and where each routes
Offers/{offer}/offer.md	The destination for the final slide, and its live facts
Analytics/what-works.md	What has measured well. Honour what each finding licenses

Step 2b: fan out the reads, in one batch

Steps 1 and 2 name a lot of files. Launch them as parallel readers rather than opening them one at a time, then plan the slides once everything has returned.

Agent	Reads	Returns
1	The source asset in full, plus its parent if it has one	The argument, the decisions already made, and whether a carousel child is already recorded in the repurpose tree
2	Context/brand/brand-kit.md	Every token needed to build the style block, the background rule at the scope it is written, and the logo and portrait pointers
3	Context/personal-brand/voice.md and Channels/{target}/voice.md	The register, the banned constructions, and the delta for this surface
4	Intelligence/research/swipe/hooks.md and ctas.md, plus Channels/{target}/strategy.md	Opening shapes, CTA phrasings with their destinations, and the format contract
5	Offers/{offer}/offer.md and proof/, plus Analytics/what-works.md	The CTA destination and its live facts, the real proof or the honest gap, and the measured patterns

Do not plan slides until every reader has returned. A slide built on a token you guessed is a slide that has to be re-rendered, and re-rendering costs credits.

Step 3: find the one point, then plan the slides

The carousel must make its point on its own. Assume the caption is one line and nobody clicks through. If the source only teases, reconstruct the real substance from it. If the source carries several ideas, pick the most useful one and say which you dropped.

Slide count 3 to 10, chosen from the content. Typical 5 to 8, and fewer lands harder.

The arc:

• Slides 1 to 3, high impact. Cover and hook, the problem, the first turn toward the answer. One large headline, one short highlight, minimal body.
• Middle slides, substance. The actual method: a smaller header plus two to four short bullets, numbered steps, or one labelled diagram. One idea per slide. This is what makes it self-contained.
• Final slide, a CTA that the slides earned. It routes to an offer named in Offers/. Never a CTA that leans on context the slides never gave.

Write the copy in the operator's voice per the register: eyebrow in caps, one headline carrying one argument, exactly one short highlighted phrase, then the body. No em dashes. No buzzwords, no hedges.

Step 4: approve the copy plan, before spending credits

Present the full slide-by-slide plan as text and stop. Per slide: eyebrow, headline, the highlighted phrase, body or bullets.

Offer at least these and wait:

1. Approve and render
2. Revise named slides
3. Change the slide count
4. Change the CTA framing on the final slide
5. Change which point the carousel makes

Rendering spends credits, so never render before the plan is approved. This is the one gate in this skill and it is not optional.

Step 5: render, then composite the footer

Render each slide with Higgsfield. Every prompt is the shared style block, built from the brand tokens read in Step 1, plus that slide's copy, and every prompt must tell the model to leave the bottom twelve percent empty in the background colour because the footer is added by code afterwards. Full call shapes and the style block in references/render-pipeline.md.

Never let the image model draw the logo, the name, or the page numbers. It cannot spell a name reliably and it cannot reproduce a mark. All three are composited:

MOSC_LOGO=<from the brand kit pointer> \
MOSC_PORTRAIT=<from the brand kit pointer, or omit> \
MOSC_NAME=<config operator_name> MOSC_HANDLE=<config org_name> \
MOSC_CREAM=<brand background hex> MOSC_BAR=<brand ink hex> \
python3 scripts/footer.py <in.png> <page> <total> <out.png>

The script exits with a one-line reason rather than guessing when any of those is missing. That is deliberate: a wrong mark on every slide is worse than a failed run.

QA every slide with a parallel agent, one per slide, launched in a single batch. Serial QA on eight slides is eight round trips, and it is where a run quietly gives up and ships a garbled bullet slide.

Each QA agent gets one rendered PNG and the copy plan for that slide, and answers four things:

1. Is every word spelled exactly as the plan specified
2. Is exactly one phrase highlighted, and is it the right one
3. Is the bottom twelve percent empty, with no model-drawn footer, page number or watermark
4. Is the type free of cursive or script anywhere

Prompt the QA agents to fail the slide, not to pass it. "Find what is wrong with this slide, and default to rejecting when uncertain" catches what "check this slide" does not. A plausible-looking misspelling is exactly the failure that survives a friendly check and then ships.

Re-render every rejected slide, then re-QA only those. Bullet slides fail most often.

Step 5b: the effort floor

Floor	Every run
Readers launched before planning	5, in one batch
QA agents	one per rendered slide, in one batch
Slides shipped without a QA verdict	zero
Re-render attempts per failing slide before asking the operator	2
Tokens or colours typed by hand rather than read from the brand kit	zero

A carousel that shipped without per-slide QA has failed even if it looks fine, because the one slide nobody checked is the one that gets screenshotted.

Assemble the PDF in order and save the slides and the PDF to a working folder outside the OS. Default ~/Downloads/<slug>-carousel-<YYYYMMDD-HHMM>/, or wherever the caller says.

Step 6: write it back into the OS

No media in the brain. The OS stores markdown and pointers. The PDF and the PNGs live on disk or in the DAM, and the OS records where.

1. File the asset. Channels/{target channel}/pipeline/<slug>/brief.md for something not yet posted, or published/YYYY-MM-DD-<slug>.md once it is. Frontmatter and body shape in references/os-contract.md. It records the slide count, the one point, the CTA destination, the PDF path as a pointer, and the parent as a wikilink.
2. Update the parent's repurpose tree, naming this child as a wikilink in a sentence. Do not touch the parent's ## Performance block.
3. Log it. Intelligence/logs/YYYY-MM-DD.md, naming the file written and the specific change. This is a hybrid: it produces a deliverable and it changes what the OS knows, so it logs the knowledge change only.

Never ask permission to save. Write and report.

Posting is not your job. You produce the PDF and record the asset. Publishing is the operator's step or a routine's.

Hard rules

1. Every colour, face and rule comes from Context/brand/brand-kit.md. Never a hex typed by hand.
2. The mark is composited by code, never drawn by the image model.
3. The footer avatar sits on a disc in the brand background colour, never on a black circle.
4. Obey the brand kit's background rule as it is scoped. If it forbids black backgrounds on audience-facing surfaces, that includes every slide here.
5. One highlighted phrase per headline. No second accent colour.
6. No em dashes, and none of the constructions the voice register bans.
7. Self-contained. It makes its point without the caption.
8. No cursive or script anywhere. Eyebrows are uppercase, letter-spaced, sans.
9. Never invent proof. A result, a name or a number comes from Offers/{offer}/proof/ or Analytics/metrics.md with its date, or it does not appear on a slide.
10. Never write a price on a slide unless you read it from Offers/{offer}/offer.md this run. Prices step on dates and a slide is hard to correct once posted.

If something you expect is missing

Missing	Do
Context/brand/brand-kit.md	Stop. This is the one file the skill cannot proceed without
The logo pointer or the file it names	Stop and ask for the path
The portrait pointer	Continue. The footer skips the avatar and the name shifts left
Channels/{channel}/voice.md	Continue on the master register alone and say so
The swipe files	Continue. Draw shapes from the source and the real published assets instead
Higgsfield	Stop before Step 5 with the plan already approved, so the run can resume later
Pillow, for the footer script	Say the one install command and stop. Do not skip the footer

Create only the files this skill owns: the channel asset, the parent's tree line, the log.

Step 7: render the one-pager. Always

Every run closes by rendering an HTML page. Not "if asked".

The PDF is the asset that gets posted. The page is how anyone sees what the run actually did without opening a PDF and a markdown file side by side. This is the plugin's convention and it is the reason instant-ui sits in this plugin at all.

Invoke the instant-ui skill with this output path:

Analytics/dashboard/runs/YYYY-MM-DD-carousel-<slug>.html

That is the same folder every routine writes its run report into. Create the folder if it does not exist.

The page carries the one point, the slide-by-slide copy table, every slide image embedded as a data URI so the page is self-contained and shows what actually rendered, each slide's QA verdict, the caption, the CTA and where it routes, the PDF path, and the OS paths written. A run report that names eight slides without showing them cannot be checked, and checking is the whole reason it exists.

Never build the page yourself and never hardcode a colour. instant-ui owns the design language and its tokens trace to Context/brand/brand-kit.md. Note that the run report is a data surface rather than an audience surface, so if the brand kit carries two token sets, this page takes the data-surface one while the slides take the audience-facing one.

Record the page path in the channel asset, alongside the PDF pointer.

The run is not complete until the page exists. If instant-ui is unavailable, say so plainly, note it in the log line, and never hand-roll a page.

Self-improvement

This skill is never finished. Improve it as you use it.

• When the user corrects a render or footer step, update references/render-pipeline.md so the correction sticks.
• When a correction is a hard rule ("always X", "never Y"), add it to the hard rules above.
• When a brand detail turns out wrong, fix Context/brand/brand-kit.md, not this skill. The brand lives in the OS. A correction landing here instead is exactly the duplication this skill removed.
• Never start a local examples folder. Good carousels are recorded as OS assets with a pointer to the PDF. That is the example library.
• Keep the skill small: when you add something, run the deletion test and cut anything that no longer changes behaviour.

Files

File	Contains
references/os-contract.md	Every OS path read and written, the channel asset frontmatter and body shape, the log line format
references/render-pipeline.md	The Higgsfield connector and CLI calls, how the style block is built from brand tokens, footer, PDF assembly, the failure modes
scripts/footer.py	Code-composited footer. Identity-free: mark, portrait, name, handle and both colours are injected
assets/fonts/mono.ttf	A mono face for the page numbers, with a system fallback ladder if the brand kit names its own