CREATOR PLAYBOOK
Daily AI YouTube Workflow
A repeatable system for finding, scoring, and packaging the strongest AI video opportunities every day. One primary video, two backups, three Shorts, and one evergreen idea — in 45 to 60 minutes.
6Stages
12Candidates
75Score gate
60Min target
PRIMARY · score 96
Mistral's 1-Trillion-Param "Le Chonk" Just Became Free — Here's What to Run
Preview API is live today; full open weights drop 27 Oct. Once the weights land, this becomes pure "how to run" instead of "what is it."
- Target viewer
- Developer / maker with 1 GPU planning hardware for 27 Oct
- Viewer promise
- Know what Le Chonk is, what it can do, and whether your rig can run it
- Format
- News explanation + live comparison test
- 30-sec opening
- One-shot prompt: build a Python data pipeline that de-duplicates Hacker News posts. Show output live. Then: "That's 1 trillion parameters — but only 52 billion were active for that answer."
5 titles + 3 thumbnails
- Mistral's 1-Trillion-Param Model Runs With Only 5% Active
- Le Chonk vs Claude Haiku 5.5 — Same Code, Half the Compute?
- I Asked Mistral's Biggest Model to Build a Web App
- What a Trillion-Param Open-Weights Model Actually Means
- The 1T Model That Anyone Can Run on 27 Oct
Thumbnails: (A) "1T params" over a chonky cat; (B) Mistral vs Claude split-screen; (C) one huge bar + 49 tiny ones.
Evidence: mistral.ai/news/mistral-large-4 (primary) · The Register · BetaNews pricing
BACKUP · score 88
Claude Code Mods Let You Rewrite Claude Itself — I Made One in 10 Minutes
TypeScript plugins ship inside plugins and can rewrite prompts, customize tool calls and permission decisions, change the interface, replace built-in features.
EXPERIMENTAL · score 89
I Tested GPT-6's "Intelligent UI" — It Builds Tools Inside Chat
OpenAI rolled out GPT-6 (Sol + Luna) globally with a feature that lets ChatGPT answer with fully interactive screens — buttons, forms, charts, tools in the conversation.
SHORTS · 3
- RESULT Le Chonk just wrote a web app in one prompt. Here's the diff.
- WARNING OpenAI's Dots agents can browse. Here's the one setting to flip.
- TUTORIAL Enable Claude Code Mods in 60 seconds.
EVERGREEN LIBRARY
Decision Models Are the New Stack Layer. Cloudflare Just Open-Sourced One.
Clef (27B) and Clef-flash (9B) — Apache 2.0, $0.24 / $0.09 per million input tokens, runnable on Workers AI. Not news-bounded.
Show the full score table (12 candidates, 8 passed)
| # | Angled title | Raw | Penalty | Final |
|---|
| 6 | Mistral Le Chonk — what to run | 96 | 0 | 96 |
| 3 | GPT-6 Intelligent UI tested | 89 | 0 | 89 |
| 9 | Claude Code Mods in 10 min | 88 | 0 | 88 |
| 14 | POCKET-Darwin-180B on a 16GB Mac | 84 | 0 | 84 |
| 8+17 | Claude voice stack + Haiku 5.5 | 81 | 0 | 81 |
| 12 | Decision models as new stack layer | 81 | 0 | 81 |
| 16 | Germany's Kolibri runs on a laptop | 81 | 0 | 81 |
| 18 | GLM-5.3 Flash beat GPT-5.6 on real bugs | 79 | 0 | 79 |
| 10 | Claude Dashboards + Motion | 79 | 0 | 79 |
| 4 | OpenAI's 372 math proofs | 84 | −10 setup | 74 |
| 1+2 | OpenAI Dots safety net test | 83 | −10 saturated | 73 |
| 7 | Gemini 4 Argon 1M output tokens | 86 | −20 unverified | 66 |
Archive
One brief per workday. Each row links to that day's full production brief. The first row below is the active run; future days append here automatically.
| Date | Primary topic | Score | Brief |
|---|
| 2026-10-10 |
Mistral Large 4 / "Le Chonk" — what to run before 27 Oct |
96 |
This page |
| 2026-10-11 |
— scheduled — (cron appends at 14:00 UTC) |
| 2026-10-12 |
— scheduled — (cron appends at 14:00 UTC) |
The 6-Stage System
Run all six, in order, every workday. Skipping a stage to save time is the single most common reason a published video underperforms.
-
1
Collect
~15 min. Gather 20 candidates from the previous 24–48 hours. Three source classes: official/primary (OpenAI, Anthropic, Google AI, model cards), audience/trend (YouTube Studio Inspiration, Google Trends, Hacker News, Hugging Face Trending, Product Hunt, GitHub Trending, Reddit), and cautionary signals that need primary-source confirmation.
Signal vs. proof. Google Trends and Hacker News surface momentum, not truth. Confirm every claim against a primary source before scoring.
-
2
Filter
~5 min. Drop a candidate when any of these is true: no trustworthy primary source; only a rumor; nothing demonstrable; no meaningful audience effect; saturated by larger channels; off-positioning.
-
3
Angle
~10 min. Replace company-centered framing with a viewer-focused promise. Every survivor must answer: Who cares? What changed? What can be shown on screen? Why today?
-
4
Score
~10 min. Score each factor 0–5, multiply by weight, sum, normalize to 100. Ship only candidates at 75 or higher.
| Factor | Weight | What a 5 looks like |
|---|
| Search / conversation momentum | 25% | Clearly rising across more than one signal source |
| Audience relevance | 20% | Direct match for the channel's proven viewers |
| Viewer benefit | 15% | Saves time, money, effort, or helps avoid a mistake |
| Visual demonstration | 15% | Clear before/after, test, or surprising result |
| Novelty | 10% | Result or angle viewers probably haven't seen |
| Competitive opening | 10% | Demand exists, strong coverage still limited |
| Evergreen value | 5% | Remains useful after the news cycle ends |
Penalties: −20 unverified −10 weak visuals −10 saturated −5 setup tax
-
5
Package
~15 min. Daily mix: 1 primary, 1 backup, 1 experimental, 3 Shorts (one result, one warning, one tutorial), 1 evergreen. For the primary, fill the 9-field production brief: topic, why now, target viewer, viewer promise, format, 30-sec opening, structure, packaging, evidence, risks.
-
6
Learn
24–48h after publish. Record impressions, CTR, first-30s retention, average % viewed, views vs. channel average, subscribers gained, comments. Weekly: identify the topic class with the highest CTR, the format with the best retention, the title language that repeatedly succeeds, and the subjects that draw views but weak subscriber growth. Adjust next week's audience-relevance and competition scores accordingly.
Pitfalls
- Announcement-itis — publishing because a company published. Always run Stages 3 and 4; if it can't be re-angled as viewer-focused, drop it.
- Trend = truth fallacy — Google Trends and Hacker News show momentum, not correctness. Never skip primary-source confirmation.
- Score inflation — a topic that hits all 5s on momentum but takes −20 for unverified claims lands below 75 by design.
- Setup tax — the penalty exists because retention collapses after a slow opening. If you can't reach a screen-worthy moment in 30 seconds, pick a different topic.
- Saturated coverage — even a strong story loses priority if three large channels shipped near-identical videos in the last 48 hours. Pivot the framing or skip.
- Shorts selection — the three Shorts must be one of each (result, warning, tutorial), not three results. Diversifying the format teaches the algorithm and the audience what to expect.
- Prompt injection — do not paste the full reusable prompt into every chat. Use it once per day for Stage 1; do Stages 2–6 in the current chat to keep context tight.
Reusable Daily Research Prompt
Use once per day to drive Stage 1 (Collect). Paste it into a fresh chat with an LLM that has web access. Then run Stages 2–6 on the response in your current chat.
You are my AI YouTube topic researcher. Find the most consequential AI developments from the previous 24–48 hours. Prioritize new tools, model releases, surprising demonstrations, meaningful product changes, controversies with verifiable evidence, and practical developments affecting creators, businesses or everyday users. Search official company announcements, YouTube Studio signals, Google Trends, Hacker News, Hugging Face Trending Papers, Product Hunt, GitHub Trending, relevant Reddit communities and competitor comments. Use primary sources whenever possible. Treat trend sources as discovery signals rather than proof. Do not include rumors as facts. Merge duplicate coverage of the same event. For every candidate, provide: topic; announcement time; original source; why viewers care; best YouTube angle; demonstration opportunity; competition estimate; timeliness, audience relevance, usefulness and visual-potential scores; total weighted score; one title; and one thumbnail concept. Return the ten strongest candidates in ranked order. Then select the best long-form video for today, best backup video, best contrarian or experimental video, three Shorts ideas, and one evergreen topic. For the winner, create a production brief containing the viewer promise, opening hook, video outline, five titles, three thumbnail concepts, evidence links and important uncertainties. The intended audience is people who want practical, understandable AI news and demonstrations. Favor useful consequences over corporate announcements.