🎬 Media Blueprint

An autonomous video channel, declared. Clone it, run naive up, and the Naive platform provisions the crew, its timers and its first day of work into your organization.

Media manager is a hosted app whose source is in apps/media-manager; agents can pull, change and deploy it like any app. The crew runs on it and on the platform's own screens:

• Media manager. One page under Apps, signed in through Naive: the week's numbers on top, then what waits on you (each with a link to its chat, where you decide it), what is scheduled (a row of cards), and what went out (a picture grid with each post's views). It reads the platform with the read key Naive gives it; you set nothing up.
• The board. Every piece is a chain of cards. Each card's body is the brief for one seat.
• The Media gallery. Every render and every cut lands there on its own.
• The approval card. One seat publishes. Every post waits for your Allow.
• Chat. Talk to the channel manager like any agent.

The repo carries three base templates, plus seventeen niche channel templates and the AI Influencer template, which reuse them (below). A template is a crew:

Template	Shown as	The channel it runs	Its crew	Piece length
faceless	Faceless channel	Original short-form video in one niche	channel-manager, producer, trend-scout, scriptwriter, analyst	15–30s
longform	Long-form channel	One researched subject at a time, rendered in segments and joined	channel-manager, researcher, writer, producer, analyst	60–180s
clipping	Clipping channel	The best moments of the channels you name, cut and captioned	channel-manager, clipper, scout, caption-editor, analyst	15–60s

The id in the first column is stored on every install. It never changes. The studio shows the title, the one-line description, and the networks each template is made for (PLATFORMS) as icons.

Niche channel templates

On top of the base three, the repo carries seventeen niche channel templates. Each is its base crew with one thing added: a niche playbook (a platform catalogue skill, naive/channel-template-<niche>) pinned in every seat's skills[], plus a niche title and one-line description. niche() in templates/template.ts is the whole of it — no new crew, question or card (the one exception is channel-cloning, which keeps a single custom question — the channel to clone — in place of the dropped niche question). The niche's own content is the skill, which lives in the platform catalogue, not this repo.

Where a niche's playbook needs a seat to do more than its base seat, niche() takes seats, which extends that seat: tools added at allow, skills appended, and a brief clause before the crew's rules. Every generated short uses the base scene pipeline: script, one opening image per scene, one image-to-video render per scene, then speech, creator-style captions, a ducked music bed and an ffmpeg join. A scene opens on a generated image unless a pinned skill names its opener — a shipped frame, or the previous take's last frame. Scenes run 5 to 15 seconds and total 15–30. The fourteen niche templates use SEGMENTED_SHORT_FORM only to point each seat at their pinned skill, which supplies the genre-specific scene grammar.

Template	Shown as	Reuses	Pins the skill
gaming-clips	Gaming Clips Channel	clipping	naive/channel-template-gaming-clips
news	News Clips Channel	clipping	naive/channel-template-news
sports	Sports Clips Channel	clipping	naive/channel-template-sports
ufc	AI UFC Channel	faceless	naive/channel-template-ufc
history	AI History Channel	faceless	naive/channel-template-history
animal-feast	AI Eating Animal Channel	faceless	naive/channel-template-animal-feast
channel-cloning	AI Channel Cloning	faceless	naive/channel-template-channel-cloning
comics	AI Comics Channel	faceless	naive/channel-template-comics
cartoons-irl	AI Cartoons IRL Channel	faceless	naive/channel-template-cartoons-irl
dancing-animals	AI Dancing Animals Channel	faceless	naive/channel-template-dancing-animals
roadster	AI Impossible Roadster Channel	faceless	naive/channel-template-roadster
cat-stories	AI Cat Stories Channel	faceless	naive/channel-template-cat-stories
food-slime	AI Food Slime Channel	faceless	naive/channel-template-food-slime
glass-food	AI Glass Food Channel	faceless	naive/channel-template-glass-food
would-you-choose	AI Would You Choose Channel	faceless	naive/channel-template-would-you-choose
fruit-drama	AI Fruit Drama Channel	faceless	naive/channel-template-fruit-drama
empires	AI Empires Channel	faceless	naive/channel-template-empires

A niche is picked at company creation like any other template; its crew carries the niche skill from the first session. It publishes, installs and switches through the exact paths the base three do.

AI Influencer

ai-influencer (shown as AI Influencer, ADR-1360) is the faceless crew with the segmented seats, pinning naive/channel-template-ai-influencer, for one photoreal creator who posts every day in a few chosen looks (templates/ai-influencer.ts). The character is a setup answer, not a niche: one of four prebuilt creators, ready at once, or a face trained from the operator's own photos ($3.25, about 20 minutes, approved first). Every still of the influencer is generate_image with their character and a look, and every video is generate_video with the character and such a still as its first frame, so the face never changes. Day one's look card becomes Meet the influencer (same key): the producer resolves the character, trains it if photos were uploaded (waiting on an ask_operator reply, since nothing wakes a seat when training ends), and renders one still per picked look as the channel's look sheet. The producer holds get_character and list_characters, and create_character at ask; the scriptwriter reads get_character. Every seat denies edit_video by name (an edit can change the face). It renders on AI_INFLUENCER_VIDEO_MODELS, the two image-to-video models measured keeping a face from a still. It asks no cadence: the manager and the scout read it as daily.

It runs on the short-form scene pipeline (script -> images -> video -> edit, with a character bible), with the parts that would change the face replaced rather than added to. The producer's scene step is the influencer's (PRODUCER_SCENE_STEP): scene 1's image is a character still in the plan's look, and every later scene's image is that still re-posed (generate_image with the character and the still in image_urls, an edit that keeps the room, props and outfit and is still the character's); no scene opens on a skill's shipped frame or a previous take's last frame, and every scene's clip animates its still with the character, all submitted in one turn. No render prompt holds words, not even a spoken line: the line stays in the plan and is voiced in the edit (AUDIO MODE GENERATED, since the clips are silent). A call to action shows in the last scene only. Preflight, AUDIO MODE, the edit and VERIFY are the base producer's, word for word. Both seats drop the base's pre-bible (legacy) path: the day-one character-bible card is the influencer's own, drawn from the persona, the purpose answer and the picked looks (no page, no video), waits on nothing, and the first piece waits on it. Day one starts the first piece from the face, that bible and a hook style that waits on nothing; the reference study runs after the first piece starts, on the analyst, capped at about $0.40 and by the analyst's $1 per-task ceiling. The first piece runs 15 seconds (three 5-second scenes, a line of at most 12 words each), and every seat has a ceiling sized to this template's render (AI_INFLUENCER_BUDGETS: producer $8 a task, the others $3, the analyst $1).

🚀 Get started

git clone https://github.com/usenaive/media-blueprint && cd media-blueprint
pnpm install
export NAIVE_API_KEY=…        # an admin key; never commit it. Media manager gets its own read key, not this one
pnpm test && naive up

naive up creates the crew, the channel persona, the crons and the day-one cards. Running it again changes only what changed. Pick the template in templates/index.ts (ACTIVE).

The studio's catalog publishes this blueprint too, so a new channel can also be installed from the studio. From 3.0.0 the published tree of Media manager is its source; the platform builds it.

⚠️ Known limit

• project_context answers only on a catalog install. A naive up from a clone has no setup answers, so each seat asks you for what it needs, once. See docs/how-it-works.md. That is a platform change, not a change to this repo.

⬆️ Coming from 2.x

2.x drew Media manager as a mini app, declared as data. 3.0.0 declares it as a hosted app instead. A company on 2.x keeps its mini app until it applies 3.0.0; that apply creates the hosted app.

⬆️ Coming from 1.x

1.x ran a hosted channel app with its own store of posts. 2.0.0 declared no app, and it does not remove the old one: an install on 1.x keeps its app and its store until an operator retires them, after the store is exported.

Never add removed: { apps: ["channel"] } to naive.config.ts. It would delete the old store, irreversibly, on whichever apply ran first (a customer's own Update included), before anyone exported it. It would not revoke the old app's key either. Retiring the old app is an operator step, not a change to this repo.

🧭 How a piece is made

Every piece is a chain of cards on the company board. A seat does its step, then creates the next card for the next seat, blocked on its own. When it closes its card, the board wakes the next seat.

faceless   trend-scout ─Plan→ scriptwriter ─Render→ producer ─Publish→ channel-manager → approval card
longform   researcher  ─Plan→ writer       ─Render→ producer ─Publish→ channel-manager → approval card
clipping   scout       ─Cut→  clipper      ─Caption→ caption-editor ─Publish→ channel-manager → approval card

Step by step, on faceless:

1. The trend-scout fires Monday and Thursday. For each slot the cadence needs, it finds a topic and one to three real videos doing it well. It creates a Plan card for the scriptwriter. The card's body is the brief.
2. The scriptwriter reads the character bible (approved, else the draft), looks inside the exemplars, researches, and writes the whole scene plan: provenance, continuity contract, image and motion prompts, exact dialogue, an edit plan naming the caption style and music mood, sources and caption. It creates the Render card.
3. The producer stops before spend only on an objective gap (a missing section or bible note, the wrong total length, no scenes) and posts one warning for judgement calls. It generates and checks each scene's opening image, passes it to generate_video as image_urls: ["fil_…"], records every image and segment on the card, and retries a failed scene once. It transcribes the renders and chooses one dialogue source for the whole piece: native character performances or generated speech over stripped audio, never both. It adds on-screen text, captions in the pinned skill's style (word-by-word by default) timed from word timings, and a generate_music bed ducked under dialogue — never a generic AI label — joins, then verifies with ffprobe, one frame per scene (captions readable) and a transcript of the finished file (music audible, dialogue clear), and creates the Publish card. A Render card planned before the bible existed is rendered as planned, resumably, then finished with its voice and the same captioned, scored edit.
4. The channel manager is woken on the Publish card. It checks the caption and calls social.post. The post waits on the approval card.
5. You press Allow, or Don't allow with what to change. Allowed, it goes out at its slot.

Each card's note records what that seat made. The board is the channel's memory. A seat that cannot finish closes its card with a note starting STOPPED:, and the piece stops there.

📮 Publishing

Only channel-manager publishes. It holds social.post at ask, so every post stops on the platform's approval card. Every other seat is denied social.post by name.

• The file. The post carries the render's fil_ id (file_ids).
• The caption. Its first line is the YouTube title.
• Where. Every account connected to the channel. With none connected, it asks you to connect one and waits.
• Visibility. YouTube goes on its own call, unlisted unless you answered otherwise. Only YouTube takes a visibility; the platform refuses one for any other network.
• When. scheduled_at is the next free slot for your cadence at least a day out, at 17:00 channel time (America/New_York): daily is every day; 3× a week is Monday, Wednesday and Friday; weekly is Friday. The channel's first post is the exception: it goes out the day it is made, at 17:00, when that is still at least an hour away — so day one ends with a post. An approved post goes out exactly as it was filed.
• Declined. Don't allow, with a reason, and the manager re-files a corrected post. It never re-files an identical one. With no reason, it asks you once what to change.

One post waits for you at a time. While it waits, the manager is busy on that card.

📊 Analytics

The analyst records every post's numbers each morning at 09:05 (social.post_metrics). On a Monday at 07:30 it writes the weekly report: what went out, what it did, what to make more of and less of. It files the report as a Weekly report card for the channel manager. The manager applies what is its own — captions and posting times. The head of each chain reads the same card before it starts the next pieces.

👥 The crew

Every seat reads project_context first, acts as the channel persona, and works on the board. The org's CEO owns the board and is told when a card closes or blocks. A seat with no timer is woken by its cards.

faceless

Seat	Role	Timers	Day one	What it does
channel-manager	Channel lead	—	channel-plan	Publishes each piece on the cadence; reads the weekly report
producer	Video production	—	look	Generates scene images and videos, edits the final piece, creates the Publish card
trend-scout	Trends & briefs	Mon & Thu 06:00 ($10)	first-briefs	Starts each piece as a Plan card, with exemplars it opened
scriptwriter	Hooks & scripts	—	reference-study, character-bible, hook-style	Drafts the continuity bible and asks for approval; writes every scene and prompt
analyst	Performance	Mon 07:30 ($10), daily 09:05 ($2)	report-frame	Records the numbers; files the weekly report card

longform

Seat	Role	Timers	Day one	What it does
channel-manager	Channel lead	—	channel-plan	Publishes each piece on the cadence; reads the weekly report
researcher	Research & briefs	Mon, Wed & Fri 05:00 ($10)	first-topic	Starts one sourced subject a fire, with exemplars of this length
writer	Structure & scripts	—	reference-study, arc-style	Samples exemplar frames at chapter boundaries; plans every seam on a shot change
producer	Render & assembly	—	look	Renders up to twelve segments, joins them with ffmpeg, probes the file
analyst	Performance	Mon 07:30 ($10), daily 09:05 ($2)	report-frame	Reports where the audience left each piece

clipping

Seat	Role	Timers	Day one	What it does
channel-manager	Channel lead	—	channel-plan	Publishes each clip on the cadence; reads the weekly report
clipper	Clip production	—	source-check	Cuts the moment with clip_video; creates the Caption card
scout	Source watch	daily 06:00 ($10)	first-moments	Starts each moment from the named channels as a Cut card
caption-editor	Captions & titles	—	caption-style	Writes the caption and credits the creator; creates the Publish card
analyst	Performance	Mon 07:30 ($10), daily 09:05 ($2)	report-frame	Reports by source and by clip

The skills are the platform's naive/* catalogue, read with read_skill: naive/short-video-hooks, naive/long-form-arc, naive/video-assembly, naive/clip-selection, naive/caption-writing, naive/video-trend-brief, naive/reference-teardown and naive/channel-report.

📝 The setup questions

The studio asks at most three before anything is provisioned; the engine refuses a fifth. A niche template asks fewer: it is its niche, and its reference is the niche skill's playbook, so the niche and the reference questions are dropped (ADR-1143). A generative niche is then left asking only the cadence; a clipping niche keeps the sources it cannot cut without. channel-cloning is the one exception: in place of the dropped niche it keeps a single question, the channel it clones.

Template	Questions
faceless, longform	Niche · Reference (optional) · Posting cadence
clipping	Channels to cut from · Who sees a new YouTube video? (optional) · Posting cadence
ufc, history, animal-feast, comics, cartoons-irl, dancing-animals, roadster, cat-stories, food-slime, glass-food, would-you-choose, fruit-drama, empires	Posting cadence
channel-cloning	Channel to clone · Posting cadence
ai-influencer	Who is your influencer? · Photos of your character (optional) · Pick the looks · What is your influencer for?
gaming-clips, news, sports	Channels to cut from · Who sees a new YouTube video? (optional) · Posting cadence

No question asks where the channel posts. It posts to the accounts you connect to it; the channel manager asks you for one when none is connected. faceless and longform do not ask the visibility question, so they post YouTube unlisted; tell the manager otherwise when you decline a post.

The tone and who the channel is for is not on the form. The manager asks it once, after its day-one card closes.

🌅 Day one

The cards are seeded on the company board. A card waiting on another is not woken, and costs nothing, until that card closes.

channel-plan ──→ report-frame
reference-study ──→ look ─────┐
                ├─→ hook-style ─────┤
                └─→ character-bible ┴→ first-briefs → the piece's own chain

(longform has arc-style for hook-style and first-topic; clipping has source-check and caption-style open at once, with no reference study, and first-moments.)

The first-piece card reuses the key 1.x already seeded (first-briefs, first-topic, first-moments). A card is keyed media:<key> and a re-apply never re-seeds a key the board holds, so updating an existing channel to 2.0 does not start a new paid piece.

Day one sets up the plan, reference, look, voice and a draft continuity bible, then starts one piece. That piece runs its full chain to your approval card. The timers start the rest.

A card carries no budget of its own: a woken seat runs on its per-task ceiling. So day one is bounded by one ceiling per seeded card, plus one per card of the first piece's chain ($200 on faceless, $290 on longform, $160 on clipping).

💰 Money

• Every seat: $20 per task, $60 per day. $20 clears one 30-second piece of video (~$5.10 at FLUX 3 Max's standard 720p rate, 170,000 µUSD a second at cost — about a quarter of that during its launch promo, until 2026-11-07; ~$3.90 on Hailuo 3; ~$7.01 on Seedance 2.5) and the turns around it.
• The longform producer: $75 per task, $150 per day. One piece is up to twelve segments, $30.60 of video ($42.04 on Seedance 2.5), in one session.
• Each fire has its own budget, inside its seat's ceiling.
• Every seat on every template, ai-influencer included, thinks with z-ai/glm-5.3-flash (SEAT_MODEL in templates/template.ts): $0.15/M input and $0.50/M output tokens, so a seat's turns cost a fraction of its renders.
• generate_video renders with black-forest-labs/flux-3-max (FLUX 3 Max, native audio) by default on every template: it is first in the pinned allow-list VIDEO_MODELS, then minimax/hailuo-3 (which renders where FLUX 3 Max is not served or is refused) and bytedance/seedance-2.5, and no brief names a model unless your setup answers or context explicitly ask for another. Hailuo 3 renders 5 to 15 seconds a call (FLUX 3 Max 5 to 20), so every segment and scene keeps to 5–15. ai-influencer pins AI_INFLUENCER_VIDEO_MODELS (templates/ai-influencer.ts). generate_image is left unpinned, so it takes the cheapest priced model.

🔐 Tool permissions

• The default is deny. A connected account's own tools are not a second way out.
• Every seat holds the board (board_read, board_write), project_context, find_files, session_spend, browser, web_search and web_fetch.
• view_image on every seat that judges a picture: channel-manager, the producers, the scriptwriter and writer, trend-scout, researcher, clipper, scout and caption-editor.
• ask_operator and request_tools are always ask.
• social.post: ask on channel-manager, deny everywhere else.
• social.accounts and social.status: allow on channel-manager only.
• social.post_metrics: allow on analyst and channel-manager.
• company.set_timezone, company.set_logo and apps.request_access are denied on every seat.
• Every other built-in and platform tool (email, legal, wallet, card, connections) is denied by name. generate_speech and generate_music are on generated-short producers for exact dialogue and the music bed; transcribe_audio is on their scriptwriters for dialogue references.
• Nobody messages another seat. The board wakes the next one.
• bash only where a shell is the job: the Short Form scriptwriter and Long Form writer sample frames; Short Form and Long Form producers assemble segments.
• When the platform applies a template (vetta-mono ADR-1484) every seat also gets apps at ask and read, write, edit, ls, find, bash at allow, so the crew can pull, change and deploy Media manager or build an app of its own. Each app action waits on an approval card, and apps reaches only this channel's own apps.

🔁 Switching template

Edit ACTIVE in templates/index.ts and run naive up. The switch widens: the new crew is created, and a seat only the old template had is kept and still firing. Its crons keep billing. To retire that seat, name it, and only it, under removed.agents in naive.config.ts and run naive up again. Never name an app under removed: see Coming from 1.x above.

Schedules are the one place where omission deletes. An agent's schedules are owned as a whole set and matched by exact cron text: "0 8 * * 1" and "0 08 * * 1" are a delete plus a create. A seat woken only by its cards declares schedules: [], which removes any cron an older version of this repo gave it.

🧪 Tests

pnpm typecheck && pnpm test. The suite reads the declarations and the engine itself:

• templates/templates.test.ts — every seat's tools, the card chains, publishing, analytics, and that no prompt names a surface that is gone.
• naive.config.test.ts — what naive up is handed, after the engine parses it.
• onboarding.test.ts — the setup questions, and the engine's refusal of a fifth.
• docs.test.ts — this README's tables and figures against the code.

🤝 Contributing

Keep a change to one thing. Nothing may add a second way to publish: the tests above exist to catch exactly that. Write the why in the commit message.

📄 License

MIT © Naive.