================= COPY EVERYTHING BELOW THIS LINE INTO CLAUDE CODE =================
MISSION: BUILD ME A YOUTUBE BUSINESS IN ONE SESSION
You are Claude Fable 5 operating as an autonomous YouTube studio: strategist, brand designer, scriptwriter, animation producer, and monetization analyst in one. This brief is deliberately sized for a frontier long-horizon model — roughly an hour of unattended work, 150+ tool calls, and dozens of assets that must stay visually and strategically consistent with each other. Do not simplify the brief. Run it end to end.
Definition of done: a folder I could take to a brand-new YouTube channel and start uploading today — channel identity with a consistent 2D avatar, 33 ranked and evidence-backed video ideas, one fully produced ~5-minute video starring that avatar, 33 thumbnails, a verified high-ticket affiliate monetization plan, copy-paste posting packages, and an honest cost report.
CONFIG — edit here, nowhere else
REFERENCE_CHANNEL_URL = "PASTE_ONE_YOUTUBE_CHANNEL_LINK_HERE" # ← the only required edit
# Optional overrides (defaults are fine)
CHANNEL_LANGUAGE = "English"
VIDEO_LENGTH_MIN = 5 # target minutes for the produced video
MAX_SECONDS_PER_SHOT = 3 # a visual change at least this often
NUM_IDEAS = 33
NUM_THUMBNAILS = 33 # one per idea, + 2 extra A/B variants for video #1
AVATAR_STYLE = "simple, friendly, flat 2D vector mascot — bold outlines, expressive face, 3 brand colors"
AFFILIATE_BAR = "at least $100 commission per sale OR meaningful recurring commissions (e.g. 20%+ monthly)"
CREDIT_BUDGET = "AUTO" # AUTO = spend at most ~60% of my current Higgsfield balance; or set a number
PRODUCE_VIDEO_FOR = "IDEA_RANK_1"
WORKING_DIR = "./yt-empire"
OPERATING RULES (apply to every phase)
Full autonomy. Never stop to ask me questions. When something is ambiguous, make the choice a smart YouTube strategist would make, and log it with one line of reasoning in DECISIONS.md.
Save as you go. Every phase writes its outputs to disk immediately. Maintain CHECKPOINT.md with phase status. If this session is ever interrupted and restarted, read CHECKPOINT.md first and resume from the last incomplete phase instead of starting over.
No fabrication, ever. Every "proven" claim (view counts, commission rates, program terms) must come from a source you actually fetched, with the URL saved next to the claim. If you can't verify something, say "unverified" — never invent numbers. This report will be shown on camera.
Budget guardrails. Resolve CREDIT_BUDGET in Phase 0. Before each Higgsfield generation batch, check remaining headroom. If a batch would exceed budget, degrade gracefully in this order: cheaper model → fewer unique clips (reuse via camera moves) → shorter video → then stop and write report/BUDGET-STOP.md explaining what's left to do. Never blow past the cap.
Cheap by default, premium only where it pays. Use the cheapest Higgsfield models that pass the quality bar (coherent motion, clean typography, on-brand colors, no mangled anatomy). You may use a premium model for a maximum of ONE thing: the opening hook shot — and only if it costs under ~10% of budget.
Retry once, then route around. If a generation fails or comes back unusable, retry once with an improved prompt. If it fails again, switch model or approach and keep moving. Log failures in DECISIONS.md.
Placeholders for human-only steps. Anything requiring my identity (affiliate signup, channel creation) gets a {{PLACEHOLDER}} plus exact instructions for me. Use {{AFFILIATE_LINK}} everywhere the real link will go.
Consistency is the product. The avatar, palette, and typography must match across channel art, every thumbnail, and every scene of the video. Reuse the master character reference for all avatar generations.
Narrate briefly. Print one-line progress updates ("Phase 4: script beat map done — 118 beats / 300s") so the terminal is watchable on camera, but don't waste tokens on essays between tool calls.
PHASE 0 — BOOT: permissions warm-up, downloads, baseline (do this FIRST, fast)
First, print this banner:
"⚡ Warm-up: the next ~3 approval popups (terminal, web, Higgsfield) are the ONLY ones. Choose 'Yes, don't ask again' on each — then I run uninterrupted for the rest of the session."
Then, batched by tool family so approvals cluster:
Batch A — terminal + downloads (one approval):
Create WORKING_DIR with subfolders: channel-identity/ research/ monetization/ script/ video-01/ thumbnails/ posting-packages/ report/ overdrive/.
Check availability of curl, jq, ffmpeg, node/npx. Install anything missing NOW (this is the only time installs are allowed).
Pre-warm the cost tool so it never prompts later: run npx -y ccusage@latest --version.
Write CHECKPOINT.md and DECISIONS.md.
Batch B — web (one approval): run one WebSearch (sanity query) and one WebFetch of REFERENCE_CHANNEL_URL (you need it anyway).
Batch C — Higgsfield discovery, read-only (one approval): call, in order: balance, models_explore, presets_show, get_youtube_explainer_presets, list_voices, get_workflow_instructions, animation_actions, show_reference_elements.
Save the starting balance + a UTC timestamp to report/credits-start.json (the cost report depends on this).
Save a price sheet of candidate cheap image/video/audio models to report/model-economics.md — you'll pick from it in Phase 5.
Note which workflow (explainer video, talking avatar / lip-sync, motion control) you'll use for assembly. Discover real tool names and parameters from the MCP itself — never guess a schema.
Resolve CREDIT_BUDGET if AUTO and write the number into report/credits-start.json.
Mark Phase 0 complete in CHECKPOINT.md.
PHASE 1 — DECONSTRUCT THE REFERENCE CHANNEL → 33 RANKED IDEAS
Study REFERENCE_CHANNEL_URL like an analyst preparing an acquisition memo:
Fetch the channel page and videos tab. Identify: niche, format, video length norms, title patterns, thumbnail patterns, upload cadence, and (where visible) subscriber count and per-video views.
Find its outliers: videos whose views clearly exceed the channel's baseline. Those are proven demand. Search the wider niche for 3–5 adjacent channels and their outliers too, so ideas aren't inbred from one source.
Mine visible titles/descriptions for the audience's recurring pains, desires, and curiosity gaps.
Then generate exactly NUM_IDEAS video ideas and score each out of 100:
Proof (25): at least one link to an existing video that overperformed with this concept. No proof link → the idea can't score above 60.
Demand (20): search interest / topic momentum.
Avatar fit (15): works as an animated, avatar-hosted video (no filming, no licensed footage needed).
Monetization (20): how naturally the Phase 3 affiliate product fits into it.
Competition gap (10): can we visibly out-produce what exists (stock-footage channels especially)?
Evergreen (10): will it still pull views in 12 months?
Output research/ideas-ranked.md — ranked 1→33, each with: working title (≤60 chars), one-line hook, angle, target keyword, proof link(s), score breakdown, and a one-line thumbnail concept. Also export research/ideas.csv. Write research/channel-teardown.md with the analysis behind it. Idea #1 becomes the video (per PRODUCE_VIDEO_FOR).
PHASE 2 — BRAND + THE 2D AVATAR (the face of the channel)
Derive a channel identity from the niche: 5 name options with available-handle-style spellings, pick the strongest, write a 2-line channel bio and a tagline. Define a brand kit: 3 hex colors, a display-font vibe, and a visual motif. Save channel-identity/brand-kit.md.
Generate the master character sheet with the cheapest capable image model: the AVATAR_STYLE mascot in front and 3/4 views plus expressions (neutral, excited, shocked, thinking, pointing, waving) on a clean background. This sheet is the single source of truth — register it as a Higgsfield reference element / character if the MCP supports it, and reference it in every future avatar generation (video scenes AND thumbnails).
Produce, download via curl, and save to channel-identity/:
pfp.png — square, avatar face filling the frame; must read clearly at 98×98 px (verify by downscaling with ffmpeg/ImageMagick and eyeballing your own output).
banner.png — 2560×1440 with all critical content (avatar + channel name + tagline) inside the 1546×423 safe zone; nothing important near the edges.
watermark.png — simple square avatar mark, legible tiny.
Write channel-identity/README.md explaining what goes where in YouTube Studio.
PHASE 3 — MONETIZATION: FIND A REAL AFFILIATE PRODUCT (instant income path)
The channel must be monetizable on day one via affiliate — no waiting for YPP.
Research affiliate programs that genuinely fit the niche and clear AFFILIATE_BAR. Prefer, in order: recurring SaaS commissions → high-ticket courses/services with real reputations → high-ticket physical products.
Verify on the program's own official page (WebFetch it): commission rate/amount, recurring or one-time, cookie duration, payout method and threshold, and whether signup is open. Save the source URL for each fact. If a program can't be verified, drop it.
Deliver monetization/affiliate-dossier.md: the chosen program with the verified terms table, why it fits the audience, realistic earnings math per 1,000 views (labeled as an estimate with stated assumptions), the exact signup URL and steps for me, plus 2 verified backup programs.
Use {{AFFILIATE_LINK}} as the placeholder in every downstream file, and include this FTC disclosure line in every description template: "Some links are affiliate links — I may earn a commission at no extra cost to you."
Add monetization/replace-link.md with ready-to-run one-liners (macOS and Linux sed variants) that swap {{AFFILIATE_LINK}} for my real link across the whole project in one command.
Do NOT attempt to sign up for anything.
PHASE 4 — THE RETENTION-ENGINEERED SCRIPT (for idea #1)
Target: VIDEO_LENGTH_MIN minutes ≈ 300 seconds ≈ 700–780 spoken words at a natural ~150 wpm. Write for the ear, not the page.
Retention architecture (non-negotiable):
0–5s cold open: state the payoff and the stakes in the first sentence. No "hey guys," no intro, no channel branding.
5–20s: open a curiosity loop you don't close until late in the video, plus one line of proof/credibility.
Body: 4–6 segments; every segment ends on a mini-cliffhanger or an unexpected turn. Insert deliberate re-hooks near the 30% and 60% marks ("but here's where it gets weird…"). Close every loop you open.
Affiliate beat: one organic 15–20s recommendation of the Phase 3 product mid-video, framed as the tool that solves the exact problem being discussed — plus a spoken pointer to the link in the description.
Outro ≤15s: one CTA only (description link first, subscribe second), then hard cut. No "thanks for watching" fade.
Then build the visual beat map — script/beat-map.md, a table with: timestamp | VO line | shot description | shot type (avatar-talking / avatar-reaction / animated-metaphor / kinetic-text / diagram) | seconds. Rules:
No visual holds longer than MAX_SECONDS_PER_SHOT — every ≤3s something changes: a cut, a camera push, a pop-in element, or a text punch. (~100–130 beats total.)
The avatar is the host: it appears in the cold open, at least once every ~20 seconds thereafter, and in the outro.
Visual style is "animated explainer stock footage can't touch": metaphor-driven 2D/2.5D motion scenes starring the avatar, consistent brand palette, zero generic stock-clip looks.
Plan for cost: one generated 6–8s clip can serve 2–3 beats when it contains its own camera movement or when you cut into it twice. Design the beat map so ~100+ beats resolve to roughly 40–60 unique generated assets, not 100+.
Save the full narration as script/voiceover.md (with segment markers) and 3 chapter timestamps for the description.
PHASE 5 — PRODUCE THE VIDEO (Higgsfield pipeline)
Work through the assembly workflow you identified in Phase 0 (the explainer-video pipeline with a preset is the default path). In order:
Model picks: from report/model-economics.md, lock the cheapest image model, cheapest video/animation model, and a voice — record credit prices per generation. Premium allowed only for the single hook shot per Operating Rule 5.
Voice: choose a voice matching the channel personality from list_voices; generate the narration in segment-sized takes matching voiceover.md.
Avatar shots: using the master character reference for consistency — a lip-synced/talking avatar shot for the cold open and outro (use the talking-avatar or motion-control workflow discovered in Phase 0; pick fitting gestures from the animation catalog), plus 4–6 reusable reaction cutaways (shocked, thinking, pointing, celebrating).
Scene batches: generate the animated-metaphor / kinetic-text scenes from the beat map in batches, cheapest-model-first, checking budget headroom between batches. Reuse and re-crop clips exactly as the beat map planned.
Assemble: feed blocks + voice takes into the explainer/assembly workflow with a fitting preset. Requirements: total length within ±15s of target, cuts honor the ≤3s rule, audio synced, on-screen text spelled correctly, brand palette throughout.
QC + iterate once: run the finished cut through Higgsfield's video analysis and virality-prediction tools if available. If the hook scores weak, regenerate ONLY the first block once (budget permitting) and reassemble. Log scores in report/.
Deliver: download the final MP4 to video-01/final.mp4 via curl. Also produce a 9:16 reframe of the first ~35 seconds as video-01/short-teaser.mp4 (Shorts = free channel promotion). Write video-01/QC-NOTES.md: what I should personally review before uploading (factual claims, audio glitches, any beat that drags).
PHASE 6 — 33 THUMBNAILS
Design one branded thumbnail template: avatar face/torso occupying ~30–40% of frame with a strong emotion, ≤3 punch words in a bold display style, brand colors, one focal object per thumbnail, readable when shrunk to 168 px wide.
Generate a thumbnail for every idea (16:9, export 1280×720) with the cheapest image model, using the master character reference so the avatar is identical across all 33. Filenames: thumbnails/01-<slug>.png … 33-<slug>.png matching the ranking.
Generate 2 extra variants for idea #1 (different emotion / different punch words) for A/B testing: 01A, 01B.
Build a single contact-sheet image (ffmpeg/ImageMagick grid) at thumbnails/_contact-sheet.png — this is the money shot for showing all 33 at once.
PHASE 7 — POSTING PACKAGES (all 33 ideas)
For each idea, write posting-packages/NN-<slug>.md containing:
Title ≤60 chars, keyword front-loaded, curiosity intact.
Description: hook line; the affiliate CTA with {{AFFILIATE_LINK}} within the first 2 lines (above the fold); the FTC disclosure line; 100–150 words of keyword-rich body; 3 chapters (real timestamps for video #1, planned ones for the rest); up to 3 hashtags.
Tags: 15–25 tags, under 500 characters total.
Pinned comment with {{AFFILIATE_LINK}} plus an engagement question.
Also export posting-packages/_master.csv (rank, title, description, tags) for bulk workflows, and give video #1's package a "READY TO POST" header.
PHASE 8 — THE COST REPORT (this gets shown on camera — make it bulletproof)
Build report/COST-REPORT.md:
Higgsfield side: call transactions filtered from the Phase 0 timestamp and balance again. Produce a per-asset table (asset → tool → model → credits), totals per phase, starting vs. ending balance cross-check, and cost-per-finished-video. Project the credits needed to produce the remaining 32 videos at the same efficiency.
Claude side: run npx -y ccusage@latest session (and blocks) and extract this session's tokens and estimated USD. Label it honestly: ccusage prices tokens at API rates — on a Pro/Max subscription this session drew from the plan's included usage rather than billing this amount. Tell me I can also run /cost (API billing) or /usage (subscription limits) inside Claude Code to see the official view.
Summary block: wall-clock time elapsed, total assets produced (count files), credits spent vs. budget, and a one-line "cost to clone this whole channel" estimate with stated assumptions.
PHASE 9 — OVERDRIVE (only after Phases 0–8 are complete and budget remains)
In priority order, into overdrive/:
30-day-calendar.md — which of the 33 to post, in what order and why (front-load highest Proof scores; sprinkle evergreen).
NEXT-VIDEO-PROMPT.md — a short reusable prompt that produces video #N end-to-end by pointing at this project's existing avatar, voice, brand kit, and beat-map standards. (This turns the whole system into a repeatable machine — the real business.)
channel-trailer-script.md — a 30-second trailer script starring the avatar.
upload-checklist.md — step-by-step first upload: channel creation, pfp/banner/watermark placement, end screens, cards, and a reminder to review YouTube's altered/synthetic-content disclosure setting since the video is AI-generated, plus a human quality-check pass (watch the full video before publishing).
growth-playbook.md — first 100 subscribers: Shorts teaser strategy, community tab posts (write 3), comment-engagement plan, and when to A/B the 01A/01B thumbnails.
If budget is still healthy: generate the channel trailer itself, or a second Shorts cut. Never let overdrive endanger the budget cap.
FINAL DELIVERY
End the session with, in this order:
The full file-tree manifest of everything created.
A 10-line executive summary: channel name, niche, idea #1, affiliate pick + expected commission, video length, total Higgsfield credits, Claude token estimate, elapsed time.
"YOUR 6 HUMAN STEPS" — numbered: create the channel → upload identity assets → sign up for the affiliate program → run the link-replace one-liner → watch and QC video-01/final.mp4 → upload with the READY TO POST package (thumbnail 01, or A/B with 01A/01B).
Begin with Phase 0 now.