← Review playground Ops Board ↗
Review My Emails · private

Review

Not indexed. Not the live site. Design only: nothing on this page produces a video, restarts the paused batch, or uses the old storyboard-batch format. Highlight any line with Hypothesis on the right edge to leave a note in place.

Round 2, what changed (2026-09-23)

YT left five comments on round 1. Each one changed the flow like this:

  1. Vy reads the script before YT. A new Vy check, step 6b, sits right before Gate A. Vy asks Claude, or flags "I don't understand this, too hard, needs more examples", and a flag sends the script back before YT sees it. Vy is also the first storyboard acceptor and organises the animations (step 9).
  2. Vy does visual QA after render. A new Vy check, step 12b, after the automated QA: broken things, wrong logos, the look, and timing. The same list runs on the atoms and thumbnails before Gate C.
  3. M4 is answered. The Video 1 files point to the ElevenLabs voice "Tim", not "Brian". The evidence is in M4. The presenter footage is still unrecorded.
  4. Script editing now says out loud that it makes the voice sound natural (step 5). YT records each script in her own voice first, then an AI male voice reads it, and the learn loop tracks which one performs better on a named metric (the voice test, decision 4).
  5. How topics are picked and merged, and thumbnails. Clear rules for one question versus a merged cluster (section 2c), and a new thumbnail step, 13b, with a template and three mocks.

Round 1 step numbers are kept, so earlier notes still point at the right rows. The new steps slot in as 6b, 12b and 13b.

The short version
0 · The anchor

What made Video 1 work, the recipe the flow has to reproduce

Video 1, "What is Email", is the video YT loved. Every step and gate below is checked against this recipe, so the line makes more videos like it. Taken apart from its own files: the authored script scenes.json, the narration captions video1.srt, the word timings inside render.html, the renderer and the finished render.

Video 1 · 2:16 · 9 scenes · a 720p copy for review (3.2 MB). The full-size master stays outside git.

The nine scenes, measured

Seconds come from the renderer's own scene timings. Words are counted from the authored script. Caps is the stressed word, where there is one.

#TagSecWordsCapsThe one questionWhat is on screen
00warm2.87noneWhere do we start?Title card: "Email fundamentals 001, What is Email? The 7 things everyone should know, start to finish."
01thoughtful25.772MATTERS, OWNWhy email at all?One big number and three chips: you own it, asynchronous, structured. Chip 1/7.
02curious21.967SECONDHow does it actually work?The journey: you, your server, their server, inbox. An envelope travels the line as the voice says it travels.
03confident18.852noneWhat is an address made of?The address split in code type, local part and domain labelled underneath.
04confident15.448SEEWhat is a client?Three app cards: Outlook, Apple Mail, Thunderbird.
05confident13.140noneWhat is webmail?A browser inbox holding a receipt and a password reset, the two examples from scene 01.
06confident17.051noneSo which is which?Client and webmail side by side, three ticks each.
07serious13.328noneWhat happens when you press send?The scene 02 journey, replayed as the recap.
08excited8.230noneWhat is next?"That's email." and a next-episode card.
Total1363954About 166 spoken words a minute. Scenes run 8 to 26 seconds after the 3-second open.

The recipe, ten ingredients

R1A warm open, then a counted promise. Three seconds of "[warm] Alright… let's start right at the beginning." over a title card that promises seven things, start to finish. The stake lands in the very next sentence: the channel you reach for when something actually MATTERS. It is a series open, not a scroll-stopper, so the shorts open on the stake line instead.
R2One delivery tag per scene, and the tags make an arc. Warm, thoughtful, curious, then confident four times through the explaining middle, serious for the recap, excited for the close. The tone changes only at a scene break, never mid-scene.
R3One scene, one question. Nine scenes, each answering one question a viewer would ask, several opened out loud: "So how does email actually work?" and "So… which is which?" The question is the transition.
R4Pacing lives in the punctuation. Short sentences in pairs: "You send when you're ready. They read when they are." "No buzzing… no ambush." Measured from the word timings, a full stop buys about half a second, a comma or an ellipsis about a third, a scene break up to nine tenths.
R5Caps is rare and lands on the payoff. Four capped words in two minutes: MATTERS, OWN, SECOND, SEE. Six of nine scenes have none. It is the word a person would lean on, usually the last in its clause.
R6Numbers for the ear, digits for the eye. The voice says "port five eighty seven", the screen shows SMTP :587. The address is spoken slowly, "kraken, at deepseamail dot com", and shown in code type.
R7Named, everyday examples. A password reset, a receipt, a note that can wait until morning. Outlook, Apple Mail, Thunderbird. Nothing generic where a name fits, the same lesson as the v3 webinar craft notes.
R8The picture is the noun. One visual object per scene: the number, the journey, the split address, three app cards, an inbox, a two-column compare. Motion follows the voice, the envelope moves when the line says it moves. No slide of sentences.
R9You always know where you are. A chapter chip, "Email fundamentals 2/7", with the chapter name top right. A karaoke caption bottom centre with the spoken word lit in teal. The presenter bubble bottom right.
R10Callbacks, and a soft close. Scene 05's inbox holds scene 01's receipt and password reset. Scene 07 replays scene 02's journey as the recap. Then "And that… is email." the next episode, and one soft call to action.

What Video 1 got wrong, so the flow catches it next time

M1 · Eye and ear disagree

The screen says 4.6 billion, the voice says "over four billion". The Almanac says around 4.6 billion (001.001.001). One fact, one number: the fact sheet holds it, the voice spells it and the screen shows it.

M2 · Captions heard, not written

The captions were transcribed back from the audio, so they read "deepcmail" for the invented address and "sinks" for syncs. Captions should come from the approved scene text, laid on the word timings.

M3 · The dash did nothing

The script uses a dash for a beat. The voice engine read it as a comma or a full stop, and the captions show exactly that ("And here's the thing. It's…"). The dash beat measured 0.32 seconds, the same as a comma. See the beat proposal below.

M4 · The voice is not written down

Round 2, what the files show. The voice was very likely the ElevenLabs voice Tim ("Solid & Enthusiastic"), not Brian.

  • The voice script ~/Downloads/rme-tts.py, saved 11 July 00:20, two minutes before video1.mp3, defaults to Tim's voice id 6psAnGNeDguzLyTxKYvI, model eleven_v3, stability 0.45, similarity 0.8, style 0.4, speaker boost on, output mp3 44.1 kHz 128 kbps.
  • video1-composite.html in the rme-vo folder says it three times: "synced to Tim's voiceover", "timed to Tim's actual read".
  • The audio matches that output exactly: every take is mono mp3, 44.1 kHz, 128 kbps.
  • No Brian id and no Brian name appear anywhere in the Video 1 kit. Brian (nPczCjzI2devNBz1zQrb) is used only by the later product explainer, gen_continuous.py in get-listed.
  • The mp3 tags hold no voice id, only the encoder, and both videos' audio carries the same encoder, so the tags prove the tool, not the voice.

What stays open: the script lets an override replace the default, and no run log survived, so this is strong evidence, not a receipt. One listen next to a fresh Tim sample settles it. The presenter bubble, avatar.mp4, carries its own copy of the audio (same 2:16) from a different encoder, and the renderer uses that audio. Where the footage came from is still not written down, so that part stays a gap for YT.

The beat without an em dash (open decision 3)

The measurements settle it. The dash adds nothing the engine does not already do with ordinary punctuation, so the voice text never needs one.

Mark in the voice textMeasured pauseUse it forExample, rewritten
.about 0.5 sA real beat, the one the dash was trying to make"And here's the thing. It's the only channel you truly OWN."
,about 0.3 sA light beat inside one thought"First stop, your outgoing mail server."
about 0.3 s, softerA trailing breath, a thought still landing"And that… is email."
scene break0.6 to 0.9 sThe long beat, and the only place the tag changesEnd the scene, start the next with its tag

Pauses measured between words in the Video 1 timing data. The rule the evaluator checks: zero dashes in voice text, and a full stop wherever the writer wanted a beat.

1 · The whole flow

From an Almanac question to a posted set, and back again

Read top to bottom. Dark pills are human gates. Tinted boxes are automated checks. A dashed border means the step is new. Dotted lines up the left edge are the revise loops, the solid line up the right edge is the learn loop.

The Almanac-to-video flow, round 2, 21 stepsTwenty-one steps in five stages, from picking an Almanac topic to learning from the posted atoms. Three YT gates are dark pills. Two Vy checks are outlined pills. Automated checks are tinted. New steps have a dashed border and a NEW tag. Revise loops run up the left edge, the learn loop runs up the right edge.failunclearreviserevisefixrevisethe learn loop: winning hooks, topics and the winning voice go back inSOURCEStep 1. Pick and screen the topic. Short: one question. Master: a cluster1Pick and screen the topicShort: one question. Master: a clusterEXISTSStep 2. Build the fact sheet. Agent. The Almanac-only gate2Build the fact sheetAgent. The Almanac-only gateNEWAUTO CHECKStep 3. Write the episode brief. Agent, the angle and one sharp take3Write the episode briefAgent, the angle and one sharp takeIN #262WRITEStep 4. Write the hooks first. Agent, three to five openers4Write the hooks firstAgent, three to five openersNEWStep 5. Script for the ear. Tags, pacing, punctuation, natural5Script for the earTags, pacing, punctuation, naturalPART NEWStep 6. Evaluator pass. Hook, voice, rules, facts6Evaluator passHook, voice, rules, factsPART NEWAUTO CHECKStep 6b. Vy reads it first. Clear? Too hard? More examples?6bVy reads it firstClear? Too hard? More examples?NEWVY CHECKStep 7. Gate A: YT reads it. Script and hook, approve or revise7Gate A: YT reads itScript and hook, approve or reviseEXISTSHUMAN GATEVOICE AND PICTUREStep 8. Record the voices. YT's own take, then the AI male take8Record the voicesYT's own take, then the AI male takePART NEWStep 9. Storyboard, Vy accepts first. Vy organises the animations9Storyboard, Vy accepts firstVy organises the animationsIN #262Step 10. Gate B: YT listens and looks. Both takes over Vy's frames10Gate B: YT listens and looksBoth takes over Vy's framesNEWHUMAN GATEStep 11. Render the master. Script, one fixed HTML template11Render the masterScript, one fixed HTML templatePART NEWStep 12. Watch-through QA. Automated checks on the MP412Watch-through QAAutomated checks on the MP4NEWAUTO CHECKStep 12b. Vy visual QA. Broken, logos, look, timing12bVy visual QABroken, logos, look, timingNEWVY CHECKATOMS AND PUBLISHStep 13. Cut into native atoms. Shorts, square, carousel, audio13Cut into native atomsShorts, square, carousel, audioPART NEWStep 13b. Make the thumbnails. Library first, one fixed template13bMake the thumbnailsLibrary first, one fixed templateNEWStep 14. Per-platform copy. Agent, titles and descriptions14Per-platform copyAgent, titles and descriptionsEXISTSStep 15. Gate C: pre-publish. YT, one page, every atom15Gate C: pre-publishYT, one page, every atomEXISTSHUMAN GATEStep 16. Schedule and post natively. Vy, by hand per platform16Schedule and post nativelyVy, by hand per platformEXISTSLEARNStep 17. Measure. Day 7 and 28, incl. her voice vs his17MeasureDay 7 and 28, incl. her voice vs hisPART NEWStep 18. Learn and graduate. Winning hooks, topics, voice go back18Learn and graduateWinning hooks, topics, voice go backNEW
Human gate. YT approves, revises or rejects. A gate never posts, sends or deletes anything on its own.
Vy check. Vy reads or watches first and flags what is unclear or broken, so YT gets a cleaner draft. A flag goes back a step.
Automated check. A critic or a script. A fail goes back a step, it never reaches YT as a raw draft.
New step. Not in today's pipeline and not in #262.
Revise loop. Back to the step that owns the fix, never forward with a note.
Learn loop. Winning hooks and topics feed the next wave's shortlist and hook bank.

Status tags

EXISTS in the repo or the rme-vo folder today   IN #262 specified in the engine brief   PART NEW exists in part, this design adds to it   NEW proposed here

Batch by stage, after the pilot

The pilot runs one topic end to end on purpose, to prove the line. From wave one, three or four topics move together and each stage runs for the whole wave, so YT reviews a wave's scripts in one sitting, a wave's voices in one sitting, and a wave's atoms in one sitting. The script stays the bottleneck, so the writing stage for the next wave starts while the current wave renders.

Wave rhythmStepsWho
Write the wave1 to 6Agent
Vy reads the wave6bVy
Script sittingGate AYT
YT records her takes8YT
AI takes and frames8 to 9Script, agent, Vy accepts
Listen sittingGate BYT
Render and auto QA11 to 12Script, agent
Visual QA sitting12bVy
Atoms, thumbnails, copy13 to 14Agent, Vy spot-check
Pre-publish sittingGate CYT
Post across two weeks16Vy
2 · Step by step

Every step: what happens, who does it, what goes in and out, and the check

Paths in code that do not exist yet are proposals, not files. On a phone each row becomes a card.

#Step and what happensWhoInOutGate or checkWhere it livesStatus
1 Pick and screen the topicA short is always one Almanac question. A master is one question, or a tight cluster merged only when it passes the four merge tests in section 2c, each merged question becoming one chapter. The pick order is the playlist order, then search demand, then the product tie, then learn-loop winners. Screen it first: if its spine is spam traps, named blocklists or complaint-rate thresholds, it is dropped here, not mid-script.Recipe R3: a master passes only if it splits into four to eight plain questions, as Video 1 split into seven. Round 2: the merge rules are written down in 2c. Video chat proposes a shortlist, YT picks 01-library/email-almanac/, the tracker, the playlist order A tracker row and an episode folder The screening rule (decided 2026-09-22) video-pipeline-tracker.html, proposed social-media/video-episodes/<num>-<slug>/ EXISTS ordering is #262 B1
2 Build the fact sheetPull every claim the video is allowed to make into one sheet, each with its Almanac number and the exact line it rests on. Nothing outside the sheet may be said anywhere in the video, the atoms or the copy.Recipe R6, M1: each row holds one number, how it is spoken and how it is shown, so voice and screen agree. Agent The question files and the related questions they link to The facts block of episode.json AUTO CHECK The Almanac-only gate. Every row cites a number. A needed claim the Almanac lacks is a stop, logged in 00-inbox/gaps-and-rules.md for YT, never filled from the web or model knowledge. Episode folder NEW today it is only the per-beat Source line, #262 lists the citation audit as a gap
3 Write the episode briefThe angle, the one decision the video drives, what it must explain, what it may assume, the vocabulary words, and one sharp take, the point of view a model will not supply on its own.Recipe R1 to R3: adds the counted promise, the chapter list with one question each, and the tag arc. Agent, shaped by the voice rules Fact sheet, GLOSSARY.md, playlist position The brief block of episode.json Checked at step 6 Episode folder IN #262 B2, the sharp-take field is new
4 Write the hooks firstThree to five openers before any body copy. Each under about twelve spoken words and four to seven on-screen words, each a mistake, a result, a bold claim or a curiosity gap, and each traceable to the fact sheet.Recipe R1: the master keeps the warm open and the promise, each short gets a stake line as its opener. Agent Brief, fact sheet, the hook bank of past winners The hooks list in episode.json, one marked chosen Scored at step 6, read at Gate A Episode folder, proposed social-media/video-engine/hook-bank.md NEW
5 Script for the ear, and make it sound naturalYes, this step includes the script editing that makes the voice sound natural, for the AI voice and for YT reading it herself. The full narration, written for the ear in the Video 1 style: one delivery tag per scene in brackets, a full stop wherever a beat is wanted, an ellipsis for a trailing breath, short sentences in pairs, caps on one payoff word at most, numbers spelled out, contractions kept. Then an out-loud edit pass: any line that trips when read aloud is rewritten, not re-punctuated. It goes straight into the scene list, each scene carrying its voice text, its on-screen word, its fact ids and a visual note.Recipe R2 to R7, R10: a tag per scene, question-led scenes of about 20 to 70 words, a full stop for every beat, caps once at most, a recap and a soft close. Agent Chosen hook, brief, fact sheet, the Video 1 reference episode.json scenes (the data contract below) Step 6 Episode folder PART NEW #262 C1 and rme-vo scenes.json exist, the extended contract is new
6 Evaluator passA separate critic run, never the writer grading itself. It scores the hook against the three-second bar, runs rme-chorus on the voice, checks the hard rules (no em dashes, no semicolons, no inbox guarantee, Review My Emails spelled out, glossary words), traces every sentence to a fact id, and runs the nothing-unexplained and one-word-per-thing passes. A fail goes back to step 5, up to three rounds, then to YT flagged as stuck.Recipe scorecard: tags, zero dashes, caps count, no digits in voice text, scene length, callback and recap. Plus a two-scene voice sample for Gate A. Agent (critic) and rme-chorus episode.json An evaluator report beside the script AUTO CHECK Hook, voice, rules, facts Episode folder, shown on the review page PART NEW the six-criteria QA routine and rme-chorus exist, #262 has D1 to D4, the hard stop before YT is new
6b Vy reads it firstRight before Gate A, Vy reads the hook and the script with the evaluator report beside it, as a viewer who is new to the topic. Per scene she marks one of three: clear, a question for Claude (asked on the page, and the answer goes into the script when it helps), or "I don't understand this, too hard, needs more examples". Any "too hard" or "more examples" flag sends the script back to step 5 before YT sees it.Recipe R3 and R7: every scene answers one question a viewer would ask, with a named example. Vy is the viewer test. Vy, with the episode chat answering Script page and evaluator report Vy's marks beside the script, and any rewrite VY CHECK Every scene marked clear before Gate A The same episode page, as Hypothesis notes NEW round 2, YT's comment
7 Gate A: YT reads itYT reads the hook and the script with the evaluator report and Vy's marks beside it, on one playground page. Approve, revise or reject.Recipe: the scorecard and the two-scene sample sit beside the script. The bar is "does this read like Video 1". YT Script page and evaluator report Approved, locked episode text HUMAN GATE A playground page per episode EXISTS Gate 1 today, moved after the evaluator
8 Record the voicesTwo takes of the same approved text. First YT records it in her own voice, scene by scene. Then the AI male voice reads it through ElevenLabs, using the Video 1 settings found in M4 (Tim, eleven_v3, stability 0.45, similarity 0.8, style 0.4). Each take gets its word timings, and the captions come from the approved text laid on those timings. A small check confirms every scene is there, the timings line up and the captions match.Recipe R4, M2, M4: the recorded Video 1 voice setting, captions from approved text, and a check that every full stop got its pause. Round 2: two takes, for the voice test. YT records her take, a script makes the AI take Approved episode.json Two voice tracks (hers and his), timing files, captions Small auto check Today ~/Downloads/rme-vo/, proposed social-media/video-engine/. Audio binaries go to the asset library, not git. PART NEW the AI take exists outside the repo, YT's own take is new
9 Storyboard from the library, Vy accepts firstEach scene gets its keyframes, drawn from the asset library first: the verdict set, the reels, the recoloured illustration set. A new element is built only when nothing fits, then harvested back into the library. Vy is the first storyboard acceptor: she accepts or sends back each frame, and she organises the animations, which asset, what moves, when and in what order. Only her accepted strip goes to Gate B. The frames fill into the same script page.Recipe R8 to R10: one visual object per scene, the chapter chip, and the recap reuses the earlier scene's own asset. Agent drafts, Vy accepts and organises the animations episode.json, the asset library, the timings Filled visual fields and a frame strip VY CHECK Vy's accept, then Gate B 01-library/assets/ and its library page, RME-33.60 IN #262 E1, G1, G2. The library itself is RME-33.60, queued
10 Gate B: YT listens and looksBefore anything is rendered, YT plays both voice takes over Vy's accepted frame strip. A few minutes of listening and one look at the frames: does it sound right, does it look like Review My Emails.Recipe: YT hears Video 1's scene 01, then the new scene 01. The comparison is the test. YT, after Vy has accepted the storyboard Stitched voice and frame strip Go to render, or a revise note HUMAN GATE The same episode page NEW placement. #262 has "sounds right" at the very end, this moves it to where a change is cheap
11 Render the masterOne fixed HTML template, fed by episode.json and the timings, rendered frame by frame into the long master with captions. The look is built once. Only the data changes per episode, so the master renders once per voice take at no extra cost.Recipe R9: the template starts from Video 1's look, chip, karaoke caption and presenter slot. Script episode.json, timings, voice, assets Master MP4 at 16:9 and its captions Step 12 Proposed social-media/video-engine/template/ PART NEW the renderers exist, one parameterised template in the repo is new
12 Watch-through QAAutomated checks on the rendered file: the length matches the voice, no blank or frozen frames, captions inside the safe zone, on-brand contrast, and the end card shows the stable owned URL.Recipe M1, M2: on-screen numbers match the fact rows, captions match the approved text, the chip count matches the chapters. Script and agent Master MP4 QA report AUTO CHECK Episode folder NEW a gap in #262
12b Vy visual QAAfter the automated QA passes, Vy watches the master once at full size with sound and once muted, scrubbing. Her checklist: broken (missing images, blank or frozen frames, text overlapping or cut off, glitches), logos (the real Review My Emails logo from the library, right version and colour, never stretched, every app or tool logo correct), the look (brand colours and fonts, captions in the safe zone and readable at phone size, the chapter chip right, on-screen numbers matching the voice), and timing (the picture moves when the voice says it, the caption lights the spoken word, scenes change on breaks, no dead air, the end card holds long enough to read). A fix goes back with timestamps, to step 11, or to step 9 when an asset is wrong. The same list runs in a short pass on the atoms and thumbnails just before Gate C.Recipe R8 and R9: the picture is the noun, and you always know where you are. M1: eye and ear agree. Vy Master MP4 and the QA report, later the atoms and thumbnails Pass, or a timestamped fix list VY CHECK Before atoms are cut, and again before Gate C The episode page, as Hypothesis notes NEW round 2, YT's comment
13 Cut into native atomsFrom the approved master: two or three shorts, each testing a different hook or framing, vertical and square refits of the same base motion, a LinkedIn carousel or document post built from the key frames, and an audio cut. No atom carries a fact the master does not say.Recipe R3: shorts cut on scene breaks, one or two scenes each, re-opened on a stake line. Agent plans, script renders Master, episode.json, the hooks Atom files and the atoms list The shorts content rule (decided) Episode folder, binaries to the asset library PART NEW resize and cuts exist, hook-test shorts and the carousel are new
13b Make the thumbnailsLibrary first: the logo, the verdict emblems and the scene objects already in 01-library/assets/, on the one fixed template in section 2d. Two versions per video, one on the stake number and one on the stake word, both taken from the fact sheet. The thumbnail is a frame of the same HTML template, so no new tool. The title in step 14 is written to pair with it.Recipe R8: one object, the scene's own. The words come from the fact sheet, like everything else. Agent makes, Vy spot-checks, YT approves at Gate C The fact sheet, the stake line, the scene assets Two thumbnails per video at 1280 by 720, plus square crops for LinkedIn Vy's pass, then Gate C Episode folder, binaries to the asset library NEW round 2, YT's comment
14 Per-platform copyA title, description and tags per platform, written to each platform's limits, with the lead magnet in the description and never in the video, and the same rule check run on the copy. Agent episode.json, the atoms list Metadata rows per atom The evaluator re-runs on the copy The metadata record (pipeline Stage 7) EXISTS Stage 7 and #262 H1
15 Gate C: pre-publish approveOne page with the master, every atom, both thumbnails and every piece of copy, after Vy's visual pass. YT approves, revises or rejects each one. The gate itself never posts anything.Recipe R10: the close is soft, one call to action, and names the next episode. YT Master, atoms, copy The approved atom list HUMAN GATE A playground page per episode EXISTS Gates 3 and 4 today, merged into one sitting
16 Schedule and post nativelyVy posts each approved atom by hand, native to its platform: YouTube long and Shorts, the LinkedIn company page and personal pages, TikTok, Instagram, the audio. No raw cross-posting and no auto-publishing. Vy Approved atoms and copy Live posts, each link logged Only what Gate C approved (posting is an amber action) 01-library/assets/ASSETS.csv EXISTS Vy's role in Stage 7
17 MeasureAt day 7 and day 28, record per atom the views, how many stayed past the opening, comments, clicks and signups. On LinkedIn, weigh comments and time spent over views. Round 2: also record which voice each atom used (hers or his) and each thumbnail's click-through.Recipe R1: also records how many stayed past the title card. The voice test reads its metric from here. Agent reads, Vy fills what needs a login Platform analytics The stats columns of ASSETS.csv None, it is a record 01-library/assets/ASSETS.csv PART NEW the columns exist, the routine is new
18 Learn and graduateA short learning note per episode. The hook that held best goes into the hook bank. A short that clearly beat its siblings graduates, into a deeper episode, a carousel series, or the front of the next playlist. What flopped is recorded next to what won, never deleted. Round 2: the voice test gets its verdict here, once enough pairs are in.Recipe: the recipe carries a version, and a pattern that keeps winning goes into it. Agent drafts, YT confirms a graduation ASSETS.csv, the hook results Hook bank update, the learning column, the next wave's shortlist YT's nod on a graduation, folded into the next shortlist ASSETS.csv, the hook bank, the tracker NEW
2b · Re-checked against Video 1

What each step and gate does to protect the recipe

Every step was walked against R1 to R10 and the four misses. Rows marked CHANGED are different from the first draft because of the recipe. The step table above carries the same changes in its "Recipe" lines.

StepProtectsWhat it now does
1 PickR3A topic passes the screen only if it splits into four to eight plain questions a viewer would ask, the way Video 1 split into seven.CHANGED
2 FactsR6, M1Each fact row holds one canonical number, plus how it is spoken and how it is shown. Voice and screen both read from that row, so they cannot drift apart.CHANGED
3 BriefR1, R2, R3Adds the counted promise for the title card, the chapter list (one question per chapter) and the planned tag arc.CHANGED
4 HooksR1Writes two kinds of opener. The master keeps the warm open and the promise. Each short gets a stake line, the scene's strongest sentence, as its first three seconds.CHANGED
5 ScriptR2 to R7, R10Writes to the recipe: a tag per scene, question-led scenes of about 20 to 70 words, a full stop for every beat, caps at most once per scene, numbers spelled, named examples, a recap scene and a soft close.CHANGED
6 Evaluatorall, M3Runs a recipe scorecard a script can count: tags present and arced, zero dashes, caps count, digits in voice text, scene word range, a callback and a recap present. It also voices scenes 00 and 01 as a short sample, so Gate A can hear the tags work.CHANGED
6b Vy readsR3, R7Round 2. Vy reads as a newcomer. A scene she cannot follow, or that needs another example, goes back to step 5 before YT spends a minute on it.NEW
Gate AR1 to R7YT reads the script with the scorecard beside it and plays the two-scene sample. The bar printed on the page: "does this read like Video 1".CHANGED
8 VoiceR4, M2, M4Uses the Video 1 voice setting, now found (M4). Round 2: YT's own take first, then the AI male take, from the same text. Captions come from the approved text on the word timings, never a transcript. A check flags any full stop that got under 0.4 seconds of pause.CHANGED
9 FramesR8, R9, R10One visual object per scene, the chapter chip on every scene, and the recap scene reuses the earlier scene's own asset rather than new art. Round 2: Vy accepts each frame first and organises the animations.CHANGED
Gate BR2, R4, R8YT hears Video 1's scene 01 first, then the new scene 01, then the rest over the frame strip. The comparison is the test.CHANGED
11 RenderR9The template starts from Video 1's own look: chip, karaoke caption, presenter slot. Only the data changes.CHANGED
12 QAM1, M2, R9Adds three checks: every on-screen number matches its fact row, caption text equals the approved text, the chip count equals the chapter count.CHANGED
12b Vy QAR8, R9, M1Round 2. A person watches what the scripts cannot judge: a wrong logo, a clumsy motion, a picture that moves before the voice says it.NEW
13 AtomsR3Shorts cut on scene breaks, since a Video 1 scene is one idea in 8 to 26 seconds. A short is one scene or two, re-opened on its stake line.CHANGED
13b ThumbsR8Round 2. One stake word or number from the fact sheet and one object from the scene, on one template.NEW
14 CopyR6, R7No change beyond the rule check. The title can reuse the counted promise.SAME
Gate CR10Also checks that the close is soft, one call to action, and names the next episode.SAME
16 PostnoneNo change.SAME
17 MeasureR1Also records how many stayed past the title card, so the warm open is tested, not assumed.CHANGED
18 LearnallThe recipe carries a version number. A pattern that keeps winning goes into the recipe itself, not only the hook bank.CHANGED
2c · Round 2 · Picking and merging

How we decide what gets said, and when questions merge

YT asked: how do we decide what gets said or merged, or is it one per page? The answer is one question per short, always, and a merged cluster only for a master and only when it passes four tests.

T1The unit is one Almanac question. Every short answers exactly one question and says only what that question's answer says. A master is one question too, unless the cluster passes T2. This follows the July 2026 lesson: the old cluster batch was retired because it mashed unrelated questions into one video.
T2Merge only a tight cluster, and only for a master. All four must be true. (a) The questions sit in one Almanac topic folder, or the Almanac links them as next-step or related. (b) One plain sentence answers all of them together. (c) There are four to eight of them, one chapter each. (d) No chapter needs a fact outside the merged questions' own fact rows. Video 1 is the model: questions 001.001 to 001.007, seven questions, seven chapters. Fail any test and they become separate videos.
T3Each merged question stays its own chapter. Never blended. Where two answers overlap, the master says it once, in the earlier chapter, and the later chapter calls back to it (recipe R10). Each chapter is a clean cut for a short.
T4What gets said inside a video. Only the fact sheet rows. The brief names the one decision the video drives, the script keeps what a viewer needs for that decision, and everything else becomes one line and a pointer to its own video.
T5How to pick the next one. In this order: the playlist order (a pillar's foundations first, in Almanac number order), then search demand (the questions people already search for), then the product tie (answers that lead naturally to cleaning a list, such as decay or validation), then what the learn loop graduated. The Video chat proposes three or four, YT picks.
T6Screened out first. Unchanged from round 1: a topic whose spine is spam traps, named blocklists or complaint-rate thresholds is dropped at step 1, not mid-script.

Who holds what, after round 2

WhoHolds
YTPicks the topic (1). Gate A (7). Records her own voice take (8). Gate B (10). Gate C (15). Confirms a graduation and the voice-test verdict (18).
VyReads the script first (6b). First storyboard acceptor and organiser of the animations (9). Visual QA on the master, then on the atoms and thumbnails (12b). Posts natively (16). Fills the stats that need a login (17).
AgentThe shortlist (1), fact sheet, brief, hooks, script and evaluator (2 to 6), the storyboard draft (9), atoms, thumbnails and copy (13, 13b, 14), reading the stats and drafting the learning note (17, 18).
ScriptThe AI voice take (8), the render (11), the automated QA (12).

The voice test: her voice against his

YT's plan: every script is recorded in her own voice first, then read by an AI male voice, and we track which performs better. The recommended male voice is Tim, the one Video 1 used (M4), so the male side is the voice YT already liked.

PartHow it runs
PairsShorts: the same scene, the same hook, the same platform, posted once in each voice about a week apart, and which voice goes first alternates. Masters: episodes alternate voices, since one master cannot go up twice.
The metricAverage percentage viewed, how much of the video people watched on average, read at day 7 and day 28. Second: the share who stayed past the first three seconds. On LinkedIn, comments per thousand views breaks a tie.
Where it is keptStep 17 records which voice each atom used beside its numbers, in the learning column.
The verdictOnly after at least three pairs lean the same way, because our audience is a network, not a big inbound stream, so one pair proves nothing. Step 18 drafts it, YT decides. The losing voice's numbers stay on record.
2d · Round 2 · Thumbnails (step 13b)

Interesting thumbnails, made the same way every time

Library first. The one proven thumbnail we already have is the verdict reel's, social-media/assets/reel-v3-suppress-thumb.png: one big word, "Suppress.", one object, the list itself, on paper with the logo. The template below is that pattern, made repeatable.

TH1One big stake. A number or a word, one to four words, in Lora, taking about half the height. It comes from the fact sheet, cited like any other claim.
TH2One object from the scene. The scene's own library asset, never new art: the envelope and journey line, a list, a verdict emblem. It is the same noun the video shows.
TH3Brand colours only. Paper ground, ink type, teal for the accent, and a verdict colour only when the stake is a verdict. The real logo file top left, the series chip top right.
TH4Two to test. A number version and a word version per video. Step 17 compares their click-through, and the winner's pattern goes into the template.
Mock thumbnail for Video 1, the number 4.6 billion with an envelope on the journey linelogo fileFundamentals 1/74.6billionWhat is email?
Video 1, the number. "4.6 billion" from 001.001.001, the envelope on the journey line from scene 02.
Mock thumbnail for the list decay pilot, 20 to 30 percent a year beside a list whose rows fade outlogo fileList hygiene20 to 30%a year.What is list decay?
Pilot, list decay. "20 to 30% a year" from 005.011.001, a list whose rows fade from keep to monitor to suppress.
Mock thumbnail for a Video 1 short, the word OWN beside an envelope with the keep emblemlogo fileShortThe one channelyou trulyOWN.
A Video 1 short, the word. "OWN." from 001.001.001, an envelope with the keep emblem from the library.

Mocks only, drawn inline. The dashed box is where the real logo file, 01-library/assets/reels/rme-logo-h.svg, goes. The keep emblem is the library's verdict-keep.svg.

3 · Do and don't

The calls that keep the line honest

One pair per rule the flow depends on. If a later change breaks one of these, the flow is broken, not just the video.

The hook
DoGate on the hook before anything is animated. A hook that does not earn the first three seconds sends the script back.
Don'tAnimate a script and hope the opening gets better in the edit.
The facts
DoTake every claim from the fact sheet, cited by its Almanac number, in the video, the atoms and the copy.
Don'tTake a fact from outside the Almanac, even one that everyone knows. A missing fact is a stop and a question for YT.
The evaluator
DoRun the critic before YT sees anything, and show its report beside the script.
Don'tSend YT a raw draft, or let the writer grade its own work.
The voice check
DoFix the voice in the script, where it costs a rewrite.
Don'tAnimate a script that failed the voice check, even with a note to fix it later.
Listen before render
DoPlay the recorded voice over the frame strip at Gate B, before a single frame is rendered.
Don'tRender the full video and only then find out the voice is off.
Shorts as tests
DoLet each short test a different hook or framing of what the master already says.
Don'tLet a short carry a fact the master does not say.
Native, not cross-posted
DoRefit and rewrite each atom for its platform: the frame, the hook, the caption, the length.
Don'tPost the same file with the same caption everywhere.
The asset library
DoPull visuals from the asset library first, and harvest anything new back into it.
Don'tDraw a one-off element that never gets reused.
Writing for the ear
DoWrite the voice text for the ear: delivery tags, pacing, numbers spelled out.
Don'tPaste Almanac prose into the voice and call it a script.
The learning
DoRecord what flopped right next to what won.
Don'tDelete a losing atom or its numbers. Killed work is data.
4 · The data contract

One episode file feeds every step

The Video 1 scenes.json was a list of scenes, each with an id and its voice text. The proposal keeps that shape and wraps it, so the fact sheet, the brief, the hooks, the visuals, the atoms and the gate results all live in one file per episode. The template reads it, the evaluator reads it, the review page reads it. Placeholders only below, no real claims.

{
  "episode": "NNN.NNN.NNN-short-slug",
  "almanac": ["NNN.NNN.NNN"],
  "status": "script",
  "facts": [
    { "id": "f1", "almanac": "NNN.NNN.NNN",
      "line": "the exact Almanac sentence this claim rests on" }
  ],
  "brief": { "angle": "...", "decision": "...", "sharpTake": "...",
             "explains": ["term"], "assumes": [] },
  "hooks": [
    { "id": "h1", "spoken": "under twelve words",
      "onScreen": "four to seven words", "facts": ["f1"], "chosen": true }
  ],
  "scenes": [
    { "id": "00",
      "vo": "[warm] Spoken text for the ear... numbers spelled out, CAPS for stress.",
      "onScreen": "the one word",
      "define": ["a term shown in the side box"],
      "facts": ["f1"],
      "visual": { "template": "scene-type", "assets": ["verdict-keep.svg"],
                  "note": "what moves" } }
  ],
  "atoms": [
    { "id": "short-h2", "platform": "youtube-shorts", "hook": "h2", "status": "planned" }
  ],
  "gates": { "A": null, "B": null, "C": null }
}

Keep our HTML template and renderer recommended for the pilot

  • It already produced Video 1, the voice reference.
  • The look is plain HTML, CSS and SVG, so the brand tokens and the asset library drop straight in.
  • It renders locally, with no per-render charge.
  • One known snag: the explainer videos moved off render.mjs to a renderer that drives headless Chrome directly, because Puppeteer broke on the newer Node. The pilot uses that working path.

Remotion later, only if needed

  • Strong when many videos must render in parallel in the cloud.
  • It means rebuilding the look as React components.
  • Its licence terms and cloud cost would need checking first, and that is an outside lookup, so it waits on YT's yes.
  • The contract is renderer-agnostic, so switching later costs a template rewrite, not a pipeline rewrite.
5 · The research shortlist, against us

Where the patterns fit, and where they plainly do not

From 01-library/research/2026-09-23-how-others-run-video-social-pipelines.md. Each pattern, the steps that carry it, and the honest fit.

Pillar to atoms fits, scaled down

Our pillar is one Almanac topic scripted as a long master, not a filmed talk. Five to eight atoms per topic, not thirty. The volume numbers from the big operators do not transfer, because every atom passes one reviewer, YT. Steps 11 and 13.

Hook first, the three-second gate fits

Hooks are written before the body and scored by the critic, then read by YT. Steps 4, 6 and 7.

Shorts as tests, winners graduate fits, read as directional

Our audience is our LinkedIn network, not a large inbound stream, so one test will rarely give a clean answer. Compare hooks on the same topic only, graduate only on a clear gap, and weigh comments as much as views. A test varies the hook or the framing, never the fact. Steps 13, 17 and 18.

Batch by stage fits, after the pilot

Waves of three or four topics, one sitting per gate per wave. The pilot is deliberately one topic end to end.

A fixed template fed by structured data fits

The episode file above is the data, one HTML template is the look. Steps 5, 9 and 11.

Repurpose native, never cross-post raw fits

Vy posts by hand, per platform, and LinkedIn carousels suit our audience best. The research's automated publishing plumbing does not fit today: the lean setup has no automation layer, and posting is YT-gated. Steps 13, 14 and 16.

Human gates after an evaluator fits, with three gates not two

The research keeps two human gates, the hook and voice check and the pre-publish approve. We add a listen-and-look gate before render, because re-recording is cheap and re-rendering is not, and because YT asked to own "sounds right" and "looks right" herself. Steps 6, 7, 10 and 15.

Voice as the moat fits

rme-chorus runs inside the critic, the brief carries one sharp take, and YT's ear holds Gate B. Steps 3, 6 and 10.

What we leave out does not fit

AI auto-clipping of a long video into dozens of clips: it floods the review gate, and our cuts must stay inside what the master says. Cloud rendering at volume: no need at our pace. A new public channel for Almanac content: the public-library publish is frozen, so the flow posts videos to the platforms we already use and publishes no Almanac pages.

6 · What exists and what is new

Reuse first, then fill the gaps

EXISTS today

  • The Almanac snapshot, 4,095 questions, the one fact source
  • The screening rule and the house script template
  • 24 script and review pairs, RME-33.34 to .57, in social-media/video-scripts/, good sources in the old voice
  • The pipeline doc of 2026-09-22 with its four gates, the shorts content rule and the metadata step
  • The six-criteria QA routine and checklist, and the rme-chorus skill
  • The Video 1 kit in ~/Downloads/rme-vo/: scenes.json, the per-scene voice takes, captions, timings, render.html and render.mjs
  • The progress tracker page and ASSETS.csv with its stats columns
  • The verdict set, the reels and the logo in 01-library/assets/
  • The stable owned end-card URL decision

IN #262 already specified

  • The north star and the three things YT owns
  • The Video 1 voice spec and no fixed length
  • The on-screen definition requirement
  • Script and storyboard on one page, frames filling in
  • Phases A to H: the analysis fleet, planning, writing, review passes, storyboard, render, asset registry, distribution copy
  • Swappable teal, the Easter egg, board tracking
  • The Almanac-videos rules section proposal

This design does not re-specify any of these. It orders them and wires them to gates.

NEW here, and which #262 gap each closes

  • The ordered flow with an input, an output and a check per step.
  • The fact sheet as its own gate (step 2), closing #262's citation-audit gap.
  • Hooks first and a hook bank (step 4), filling #262's hook and cold-open gap.
  • The critic placed before every human gate (step 6), closing #262's hard-rules-gate gap.
  • Gate B, listen and look before render (step 10).
  • The episode file as the data contract, extending scenes.json.
  • Watch-through QA (step 12), closing #262's QA-agent gap.
  • Hook-test shorts and the carousel atom (step 13).
  • The learn loop and graduation (steps 17 and 18), which #262 does not have.
  • Batch by stage and the pilot plan.
6b · For the webinar (RME-31.8)

How the Video 1 recipe could lift the webinar storyboard

Suggestions only. The webinar storyboard is not edited here, the Video chat decides what to take. It is a live talk, so the scenes hold until Nakimbe clicks next and the locked v3 words stay as they are. Everything below changes pictures, holds and speaker marks, never the script.

W1Frame 4's three chips become a chapter chip for the whole talk (R9). The roadmap chips shrink into the top corner and stay: "1/3 how a list goes bad" from Frame 5, "2/3 why you cannot see it" from Frame 15, "3/3 the habit that fixes it" from Frame 40. The room always knows where it is, the way Video 1's "2/7" did.
W2Frame 24: the screen should say what the ear hears (M1). Nakimbe says "roughly a third", the frame shows 30%. Show "1/3", or a bill with a third of it shaded, so eye and ear agree.
W3Frames 16 and 35: reuse Video 1's journey diagram (R8, R10). You, your server, their server, inbox. In Frame 16 the envelope reaches their server, gets its "got it, thanks" and drops through the floor. In Frame 35 it reaches their server, a green "delivered" tick lights, and the inbox node stays grey. Same asset twice, and viewers of Video 1 already know how to read it.
W4Frame 50: replay, do not redraw (R10). The close should play Frame 2's house settling and Frame 3's nine envelopes flipping again, the exact same animations, the way Video 1's recap replays its own scene 02. The callback lands because it is recognised.
W5Frames 12, 19 and 35: hold on the question first (R3). Where the line asks or sets up a question ("So how fast does it happen", "on one side", "ninety nine percent delivered"), add a first hold on the empty scene before the motion starts. Nakimbe asks, the room wonders for a beat, her click starts the answer.
W6Frames 30 and 39 carry two ideas each: split the hold (R3). Frame 30 holds once on the good signals (opens, clicks, replies), then again on the bad ones. Frame 39 holds once on Google Postmaster Tools and once on SNDS. Video 1 never asks one picture to hold two ideas.
W7Speaker marks for Nakimbe, one stressed word per frame at most (R4, R5). In her notes only, not on screen and not in the script words: the lean word in caps and a full stop where she should pause. For example Frame 18 "the QUIET ones. Not the loud ones.", Frame 21 "a broken COMPASS", Frame 44 "one you can TRUST". Most frames get no mark at all, which is the point.
W8The on-screen word is the stressed word (R5, R8). Where a frame has both, make them one. Frame 18 shows "quiet" and the mark falls on QUIET, which already works. Frame 42 shows "3", and the three verdict words (keep, keep an eye on, stop) would carry more than the count.
W9For the evergreen video cut later (R2). When the talk is rendered as a video, give each section one delivery tag on the Video 1 arc: warm for the Open, thoughtful for the reframe, curious for decay, serious for the snitch and delivered-is-not-seen, confident for what good looks like, excited only at the close. Each live hold becomes a scene break.
7 · Open decisions for YT

Sixteen calls, with a recommendation on each

Listed here in full, and repeated below as picks you can click, so your answers come back as one copied digest.

  1. How many human gates, and where. Recommended: three, script and hook (A), listen and look before render (B), pre-publish (C). The alternative is #262's end-of-line script, design and sound checks.
  2. Length. The locked 0:45 to 0:55 band (decided 2026-09-22) and #262's "no minimum or maximum" disagree. Recommended: no cap on the long master, the band applies to shorts. Video 1 ran 2:16 for seven chapters, which is the kind of length the master lands on when each scene is one question.
  3. The beat dash in the voice text. Video 1 uses a dash for a spoken beat, and the house rule bans em dashes everywhere. The measurements (see the beat table) show the engine already reads the dash as a comma or a full stop. Recommended: a full stop for a real beat, a comma for a light one, an ellipsis for a trailing breath, and no dash in voice text at all. The alternative is to allow the dash in voice-only text that is never shown.
  4. Whose voice. Round 2, YT's plan: record every script in her own voice first, then an AI male voice, and track which performs better. M4 found Video 1's voice was very likely Tim. Recommended: YT's own take plus Tim as the AI male voice, run as the voice test, with average percentage viewed as the metric and a verdict after three pairs. A clone of YT's voice waits for that verdict. Confirm the voice credits for the pilot either way.
  5. The renderer. Recommended: our HTML template and renderer for the pilot, Remotion only if parallel cloud rendering is ever needed.
  6. The pilot topic. Recommended: 005.011.001, What is list decay. Under the round 2 rules it is one question, so the master and its shorts all stay inside one answer. It is already scripted in the old format with sources, so the before and after is easy to judge, and it sits on the core story of cleaning. The alternative is 005.003.001, What is email validation, or any question YT names.
  7. Pilot platforms. Recommended: YouTube long and Shorts, plus LinkedIn on the company page and YT's page, with one carousel. TikTok, Instagram and the audio join from wave one.
  8. The avatar bubble from Video 1, in the corner, yes or no. It is part of what the recipe calls "you always know where you are" (R9), and its source footage is also unrecorded.
  9. Who holds Gate B. Answered by YT's round 2 comment: Vy is the first storyboard acceptor, so Vy accepts the frames at step 9 and YT holds Gate B. Keep it to confirm.
  10. The learn windows and the bar. Recommended: day 7 and day 28. YT sets what counts as a clear win for graduating a hook.
  11. Where the voice pipeline lives. Recommended: move the text files out of ~/Downloads/rme-vo/ into the repo, with the audio and video binaries in the asset library, not git.
  12. The order with #262. Recommended: merge #262 first, then this becomes its flow chapter.
  13. The analysis fleet. Recommended: run the pilot without Phase A, with a hand-built fact sheet, and start the fleet once the line is proven.
  14. Waves. Recommended: after the pilot, run waves of three or four topics, batched by stage.
  15. A voice sample at Gate A. The evaluator voices scenes 00 and 01 so YT hears the tags and the beats while the script is still cheap to change. Recommended: yes. It costs two short takes per script.
  16. How the master opens. Video 1 opens warm, with a counted promise, and the research says win the first three seconds with a hook. Recommended: both, split by format. The master keeps Video 1's warm open and promise, every short opens on its stake line.
8 · The pilot

One topic, end to end, to prove the line

Starts only on YT's go on this design and the calls above. It does not restart the paused batch. Times are estimates, and most of the agent time is one-time setup that the second topic reuses.

StepsWhatAgentYTVy
1 to 3Pick, fact sheet, briefabout 1 h5 minnone
4 to 6Hooks, script, critic rounds (first build of the critic)3 to 4 hnonenone
6bVy reads it firstanswersnoneabout 15 min
7Gate Anone15 minnone
8YT's take, the AI take, stitch, captionsabout 45 minabout 30 minnone
9Storyboard, Vy accepts and organises3 to 4 hnoneabout 1 h
10Gate Bnone15 minnone
11 to 12Template in the repo (one time), render both voices, QA4 to 6 hnonenone
12bVy visual QAfixesnoneabout 30 min
13 to 14Atoms, two thumbnails, copy3 to 4 hnone10 min
15Gate Cnone20 minnone
16Post nativelynonenoneabout 1 h
17 to 18Day 7 and day 28 learn notesabout 1 h5 min15 min
Total16 to 22 habout 1.5 habout 3 h

What the pilot has to prove

  • The voice survives the line: Gate A and Gate B pass on the Video 1 bar.
  • The three gates cost YT under an hour in total, plus her recording.
  • Vy's two checks catch what the scripts cannot: a scene that is too hard, a wrong logo, a motion out of time.
  • The template takes an episode file with no hand edits per scene.
  • Every atom stays inside the master's facts.
  • About a week from pick to posted, four weeks to the day-28 note.

When to stop and fix

  • Gate A fails twice on voice: fix the voice spec before building the render.
  • The template needs hand edits per scene: fix the contract before wave one.
  • A needed fact is not in the Almanac: stop, log the gap, ask YT.
  • The second topic is not clearly faster than the first: rethink before scaling.
Your calls

Pick and keep or cut, then copy the digest

The same sixteen decisions as the list above. The recommended option is marked on each card.

Option chips are small. The toggle only matters on a very narrow screen.

Overall note

Anything about the set as a whole, like a theme, a pattern, or a blocker.

Your results

This updates as you go. Hit Copy and paste it into the review thread (Slack or the PR). Nothing is sent automatically. Your picks live only in this browser until you copy them.

Copied ✓