Opus 5.5 Motion Lab Make Claude Animate Like a Studio

Opus 5.5 Motion Lab: Make Claude Animate Like a Studio ![rw-book-cover](https://pbs.twimg.com/profile_images/2100596838915092480/4Mv5e9Qa.jpg) 🔗 **Source:** [Opus 5.5 Motion Lab: Make Claude Animate Like a Studio](https://twitter.com/flxrnc/status/2105311836190978392/?rw_tt_thread=True) > [!note]+ Document Notes > tags: [motion-design, claude-opus, prompt-engineering, three-js, ai-video, animation-principles] > author: flxrnc > source: X (Twitter) > reference: https://twitter.com/flxrnc/status/2105311836190978392 > date created: 2026-10-12 13:53 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ Title: Opus 5.5 Motion Lab: Make Claude Animate Like a Studio >  ## 🧭 Table of contents - [📝 Summary](#-summary) - [📖 Main content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning styles assessment](#-learning-styles-assessment) - [📊 Accessibility assessment](#-accessibility-assessment) - [🌍 DEIB assessment](#-deib-assessment) - [✅ Logical validity](#-logical-validity) - [🛡️ DARVO analysis](#️-darvo-analysis) - [🎯 Conclusion](#-conclusion) - [🔄 Refactor appendix](#-refactor-appendix--opus-55-motion-lab-make-claude-animate-like-a-studio) ## 📝 Summary This is a long X (Twitter) article by the handle @flxrnc (Reader author: Florence), posted at https://twitter.com/flxrnc/status/2105311836190978392. It is a tutorial and personal showcase: the author prints the prompts and follow-ups behind three animations made with Claude Opus 5.5, then gives a setup guide, a number cheat sheet, and a fix list. It is practitioner advocacy for one model and one workflow, not a controlled study. The author's closing call is to follow the account and bookmark the article (¶245). The article does not disclose a commercial tie to Anthropic, and the capture shows no sponsor line. Claude Opus 5.5 is made by Anthropic, which also made the assistant that wrote this analysis. This analysis weighs the claims on their evidence, not on that fact. - **What happened.** The author publishes a "lab notebook" for three loops: a 6-second liquid-chrome title (¶4-¶14), a vague-versus-directed product-shot comparison of earbuds (¶41-¶51), and a 3D mascot loop built on five animation principles (¶136-¶150). Each comes with timings, follow-up counts and a sound pass. Around them sit a render contract (¶69-¶72), a setup guide (¶73-¶98), a prompt template and ban list (¶99-¶118), a numbers cheat sheet (¶119-¶134), and an export and fix section (¶208-¶234). - **The claim.** Directed prompts full of numbers (easing, damping ratios, frame ranges, light strengths) beat adjectives, and a render-look-fix loop with frame-numbered notes finishes the piece: "Every adjective you replace with a number is a decision you take back from the model" (¶58). The author adds that honest receipts ("post every prompt with the work") convince better than "one prompt" claims (¶186). - **Why it matters.** The article gives a reproducible method for code-drawn motion: Opus writes an HTML or Remotion file, a headless browser freezes each frame, ffmpeg stitches it (¶60-¶61, ¶88-¶90). It also supplies concrete starting numbers, for example a damping ratio of 0.7 to 0.85 for premium feel (¶124) and a beat-grid formula (¶130). - **Key stakeholders.** Motion and design practitioners, developer advocates who make demo clips, documentation teams that explain generative workflows, trainers, marketers who post launch loops, and tool vendors named in the equipment list (Remotion, HeyGen HyperFrames, GSAP, Motion, three.js). - **Risks and tradeoffs.** The decisive artifacts are missing from the capture: 28 "Some content could not be imported from the original document" placeholders replace every prompt, follow-up list, code block, table and the render contract, and all 28 matching entries in the link list point back to the thread URL, not to the content. The captured text therefore can't confirm "the exact prompt." Times, follow-up counts, the 71 numeric checks and the 80% energy-under-60-Hz figure are the author's own, with no log or repository in the link list. Social view counts (1.8 million, 3.6 million, 1.75 million) are quoted from posts, not independently shown. - **Bottom line.** A specific, usable method with concrete numbers and an honest framing against "fake one-shots." What it lacks in the capture is the evidence behind its own receipts. The pattern is stronger than the proof: the numbers in the cheat sheet are starting points the author labels as such (¶128), and several attributions to Anthropic's guidance (¶65, ¶93, ¶94, ¶101) have no link in the captured markup. ## 📖 Main content **Opening and framing (¶1-¶3).** The author says every animation was made by Claude Opus 5.5 and comes with the exact prompt and every follow-up: "No fake one-shots." They say that since the September 22 launch, the feed fills with pieces captioned "one prompt," and that the middle part (second prompt, third, the frame-84 note) is rarely posted. The article promises three animations with picture and sound, the numbers that made them work, and a setup from an empty folder to an X-ready loop. **The opening shot (¶4-¶14).** A 6-second loop: a chrome blob splits into 7 drops, each pours into a letter, a glint sweeps the word, and the letters melt back (¶5). The liquid is a raymarched shader shape, the letters are three.js TextGeometry, and frames render on the GPU at 60 fps with subframe motion blur. One prompt line exists because of a three.js r160 trap: TextGeometry reads "height" for extrusion depth and overwrites "depth" with 50 (¶8). There were 10 follow-ups, 5 from stills and measurements and 4 from a frame-by-frame check of 12 transitions, plus one for a horizon line (¶9). Sound came from one prompt and 5 follow-ups, all measured on decoded audio, because the model can't hear (¶11). Total: about 2 hours of picture plus 45 minutes of sound (¶13). **Contents list (¶15-¶26).** A table of contents lists ten parts: Experiment 1, how it works, setup, prompt anatomy, the numbers, Experiment 2, the lab loop, equipment, the last mile, and fixes. The subhead says a browser and Claude are enough to start. **Experiment 1, vague versus directed (¶27-¶58).** The author first discusses the most copied launch-week prompt: a 26-word showreel request by @shneural (kirill sh), posted September 24 with Opus 5.5 beside GPT 6 Astra, from an account the author says has 885 followers and 1.8 million views (¶28). The author notes the showreel has no brief to miss, so it makes a good model demo and a poor test of your own launch (¶32-¶33). The most viewed music video is by donald (@donaldjewkes), whose post says he spoke for 5 minutes and Claude worked 12 hours, and the author reads that as direction doing the heavy lifting (¶34-¶40). The controlled test then uses one subject (wireless earbuds), one technical line, and two prompts: A (adjectives) and B (448 words of direction). A got 0 follow-ups and came back as dark navy space with a glow ring, bloom and particles; B took 15 follow-ups in 4 rounds, 12 sound follow-ups, and about 30 versus 90 minutes plus 90 minutes of sound (¶47-¶51). The five changes from A to B turn adjectives into materials, a studio, a camera, beats and springs, and a named ban list (¶53-¶57). **How Opus makes motion (¶59-¶72).** The model outputs text, and the clip is a program. The author quotes Gideon Frost of OrcaRouter: "does not generate pixels. It writes code that paints pixels" (¶60), and Eric Kang's rule that every pixel is a pure function of t (¶63). Three consequences: reproducible frames, cheap edits, and self-checking. The self-checking claim leans on an Anthropic statement about dense-chart reading (¶65). A render contract pasted under any prompt closes failure modes such as clocks, unseeded randomness and a duplicated first frame (¶69-¶72). **Setup (¶73-¶98).** Path 1 is the Claude app with a screen recording, plus a frame counter and a space-bar pause (¶75-¶82). Paths 2 and 3 use Claude Code on Mac or Windows, with a render.mjs script: `node render.mjs anim.html 6 24` writes 144 PNGs (¶88-¶90). The 3D pieces use a real GPU (an RTX 4060 Ti) and blend subframes spread over half a frame, which the author says matches a 180-degree shutter (¶91). Effort stays at medium unless a piece misses twice (¶93). Fast mode costs $8 and $40 per million tokens against $4 and $20 (¶94). A Shikhar post, "Third iteration with Opus 5.5 medium," is cited at 1.75 million views (¶95-¶98). **Prompt anatomy (¶99-¶118).** Each line of a directed prompt removes one decision from the model, and the model's own decisions are the default look (¶100). The author cites Anthropic's prompting guide that "avoid a generic AI look" swaps one default for another, and says naming exact patterns works (¶101). The template starts with GOAL and TIME, then SHOTS, with a frame range, one hero move and a handoff each (¶105-¶106). The ban list and the rank rules for secondary and ambient motion follow (¶108-¶117), with a Rexan Wong quote (¶118). **Numbers cheat sheet (¶119-¶135).** Easing values come from Andrey Sitnik's Easings reference (¶121). Rules: enter with ease-out, exit with ease-in at 40 to 50% shorter, linear only for rotation and progress (¶123). Springs translate through the damping ratio: 1.0 never bounces, 0.7 to 0.85 premium, 0.4 to 0.6 playful (¶124), and Remotion's default works out to 0.5 (¶125). The spring function returns progress for any t, so frame 84 needs no frames 0 to 83 (¶127). Durations, stagger, the beat-grid formula, holds, camera speeds and WCAG contrast and flash limits follow (¶128-¶134). **Experiment 2, the mascot (¶136-¶151).** An 8-second loop with a voxel Claude mascot hops pad to pad along a floor that works as a motion timeline, with three studio lights and labels pinned to frames 69, 140, 200, 285 and 303 (¶138, ¶148). It took 4 follow-ups and a numeric audit of 480 frames with 71 checks; the first run failed 6 and all were fixed (¶141). Sound needed no follow-ups (¶143). Total: about 2 hours plus 70 minutes of sound (¶145). The prompt pins the mascot's proportions as a guardrail (¶149). **The lab loop (¶152-¶171).** Render stills, look, write frame-numbered notes, fix, re-render only what changed, two to four rounds (¶153). A URL parameter freezes any frame (¶154), a 12-frame contact sheet works as a storyboard (¶156), and onion skin and speed graphs sharpen notes (¶158). Feelings translate to numbers: "make it pop" becomes a 1.0 to 1.08 scale over 6 frames (¶163). A log table of receipts is in a placeholder (¶169). **Other people's receipts (¶172-¶186).** Miles reports 3 prompts for a launch video (590K views, ¶173-¶175). Vox reports a gravitational-lensing lab: one prompt, three overnight polish rounds of 3 agents each, 5 hours 28 minutes, about $90 at API prices (¶176-¶185). The author's moral: count your follow-ups and post every prompt (¶186). **Equipment (¶187-¶207).** Three.js (r186 shipped September 24), Remotion (free for individuals and companies of up to 3 employees), HeyGen HyperFrames (open-sourced April 17, 2026, Apache 2.0), GSAP (now 100% free per the author), Motion (MIT, no video renderer) and plain HTML and CSS (¶189-¶194). A rule of thumb picks one by job (¶196). Examples: Chubby's 3-minute film of about 7,400 lines of Remotion code (¶198-¶202) and Meng To's playable boat scene (¶203-¶207). **The last mile (¶208-¶228).** Sound: the feed plays muted, so picture works first; mix to -14 LUFS with peaks at -1 dBTP; sound lands 1 frame after its picture in the author's pieces (¶210-¶212). Loops: render frames 0 to N-1 and drive from phase = frame / N (¶214-¶215). Export: MP4 for posts, GIF for Articles, one palette, and a GIF delay trap (five frames at 40 ms, one at 50 ms) (¶217-¶222). Canvas sizes are in ¶223. Leon Lin's code-only launch video for a fake calendar app (one prompt, about 8 hours, 85K views) closes the section (¶224-¶228). **Fixes and Experiment 3 (¶229-¶245).** Failures return in every project; the fix table is a placeholder (¶230-¶231). If a fix doesn't hold, render the two frames on either side and describe the difference (¶232). Fixes get added to the render contract (¶234). Experiment 3 is the reader's: swap words for your product, run at medium, render stills at five frames, write two notes, fix, export, and post the prompt with every follow-up (¶236-¶241). Keep the numbers on the first run and change one at a time on the second (¶242). The article ends with a follow-and-bookmark ask (¶244-¶245). ## 🗒️ Notes #### 📄 ¶1-¶3 Premise: receipts, not fake one-shots - ¶1 states the standard: every animation ships with its exact prompt and every follow-up. - ¶2 names the problem: feed posts captioned "one prompt," with the second and third prompts missing. The launch date given is September 22. - ¶3 sets scope: three animations, picture and sound, with the numbers and the setup. #### 📄 ¶4-¶14 Opening shot: liquid chrome - ¶5: 6-second loop, 7 drops, raymarched shader liquid, three.js TextGeometry, 60 fps GPU render with subframe blur. - ¶8: the r160 TextGeometry trap ("height" versus "depth" overwritten with 50); fix is to pin the version and name the right parameter. - ¶9-¶11: 10 follow-ups (1 to 5 from stills, 6 to 9 from 12 transition checks, one for a horizon line); 5 sound follow-ups from decoded audio, none from listening. - ¶13-¶14: about 2 hours plus 45 minutes of sound; the wobble runs through frame 0 so the seam doesn't show. ¶7, ¶10 and ¶12 hold the prompt, follow-ups and sound prompt, which the capture replaces with import placeholders. #### 📄 ¶15-¶26 What's inside - ¶15-¶25 list the ten parts. The list gives the article its structure. - ¶26 says no After Effects is needed. #### 📄 ¶27-¶33 The showreel prompt as demo - ¶28: the 26-word showreel prompt by kirill sh, September 24, Opus 5.5 versus GPT 6 Astra; the author cites 885 followers and 1.8 million views. - ¶29-¶31: the embedded post text, which includes the poster's judgment that one model has "awful taste." That is the poster's opinion, not the article author's. - ¶32-¶33: the author's point is that a showreel has no brief to miss, so a one-liner says nothing about your own product or timing. #### 📄 ¶34-¶40 The "one prompt" music video - ¶34: credit split. The song "Claude-Pop" and its first video are by deckard (September 9, before Opus 5.5); donald made the Opus 5.5 version. - ¶35-¶39: embedded posts. donald: 5 minutes of speaking, 12 hours of Claude work. - ¶40: 3.6 million views; the author reads "one prompt" as a spoken brief plus autonomous work. #### 📄 ¶41-¶58 Controlled test: vague A versus directed B - ¶41-¶42: same model, same subject (earbuds), same technical line (one HTML file, three.js, 960x1080 per side, 6-second loop). Prompt A is three adjectives; B is 448 words. - ¶47: A had 0 follow-ups and returned navy space, halo ring, bloom, particles and a made-up tagline; B took 15 follow-ups in 4 review rounds. - ¶49-¶51: 12 sound follow-ups in 4 rounds; first mix had 80% of energy under 60 Hz and hits only 0.5 to 1.6 dB over the bed. Time: about 30 versus 90 minutes plus 90 for sound. - ¶53-¶58: adjectives became numbers (roughness 0.38, clearcoat 0.35, chrome 0.05, 30-degree FOV, 120 BPM, damping 0.6 and 0.7). ¶58 is the thesis line. #### 📄 ¶59-¶72 How motion gets made - ¶60-¶61: the model writes code; a headless browser freezes frames; ffmpeg stitches them. - ¶63-¶65: three consequences (reproducible frames, cheap edits, self-checking). The Anthropic vision claim has no link in the capture. - ¶66-¶68: Kevin Ngo's 28-second code-drawn story, 640K views. - ¶69-¶72: the render contract (placeholder in ¶70) and why each line exists: clocks, unseeded randomness, and the duplicated first frame. #### 📄 ¶73-¶98 Setup from zero - ¶74-¶82: Path 1, the Claude app, with a frame counter and pause key; screen recording isn't frame-exact. - ¶83-¶90: Claude Code on Mac or Windows, a render.mjs script, two Windows notes (no && in PowerShell 5.1; use npx.cmd). The commands and script are placeholders in the capture. - ¶91-¶92: GPU rendering with subframe blending; Claude Code runs the loop from one instruction. - ¶93-¶98: medium effort as default, fast mode pricing, and Shikhar's 3D browser world at 1.75 million views. #### 📄 ¶99-¶118 Prompt anatomy - ¶100-¶101: each line removes a decision; naming exact patterns to avoid beats "avoid a generic AI look," per the author's reading of Anthropic's guide. - ¶103-¶107: the template (placeholder). GOAL first, TIME second, SHOTS with frame range, hero move and handoff, and a QA line. - ¶108-¶115: the ban list and the advice to extend it after every first render. - ¶116-¶118: one hero per shot; secondary moves at 50 to 70% travel, 100 to 200 ms later; ambient under 5%. #### 📄 ¶119-¶135 The numbers - ¶121-¶123: easing rules; the values table is a placeholder. - ¶124-¶127: damping ratio as the cross-library number; spring function returns progress for any t. - ¶128-¶133: durations, stagger, frames at 24 fps, beat grid (frames per beat = fps x 60 / BPM), holds, camera and parallax. - ¶134-¶135: WCAG-derived 4.5:1 text contrast and a 3-flashes-per-second ceiling. #### 📄 ¶136-¶151 Experiment 2: mascot and five principles - ¶137-¶138: Thomas and Johnston's 12 principles (The Illusion of Life, 1981); 8-second loop with labels on frames 69, 140, 200, 285 and 303. - ¶141-¶145: 4 follow-ups, a 480-frame numeric audit with 71 checks (6 failed first), no sound follow-ups, about 2 hours plus 70 minutes. - ¶148-¶150: labels double as a test; proportions pinned as a guardrail; playhead and labels for iteration, off for final. - ¶151: push key poses 10 to 30% further for phone size. #### 📄 ¶152-¶171 The lab loop - ¶153-¶154: render, look, note, fix, re-render; freeze a frame by URL (?t=2.5). - ¶156-¶158: contact sheet, onion skin, speed graph. - ¶159-¶167: seven paste-ready follow-ups (placeholder) and five feeling-to-number translations. - ¶168-¶171: receipts table (placeholder), creative versus technical follow-ups, and the sound follow-up counts 5, 12 and 0. #### 📄 ¶172-¶186 Other creators' receipts - ¶173-¶175: Miles, 3 prompts, 590K views. - ¶176-¶185: Vox, 5h 28m, about $90 at API prices, three rounds of two reviewers and one fixer, P0 to P2 ranking. - ¶186: the moral line on counting follow-ups. #### 📄 ¶187-¶207 Engine choice - ¶188-¶194: Three.js r186, Remotion licensing, HyperFrames (Apache 2.0, April 17, 2026), GSAP, Motion, plain HTML and CSS. - ¶195-¶197: a comparison table (placeholder) and a rule of thumb; put the engine in the OUTPUT line with numbers. - ¶198-¶207: Chubby's 3-minute Remotion film (about 7,400 lines, "about 1h and 7% weekly rates" per the embedded post, 196K views) and Meng To's boat scene (473K views). #### 📄 ¶208-¶228 Sound, loops, export - ¶210-¶212: mute-first feed, -14 LUFS, -1 dBTP, sound after picture; sync checked by decoding the MP4. - ¶214-¶216: loop on frames 0 to N-1, phase = frame / N, a finished frame 0. - ¶217-¶223: MP4 versus GIF, X limits (table placeholder), the 350-frame cap that gives about 14.5 seconds at 24 fps, the 40/50 ms delay trick, canvas sizes. - ¶224-¶228: Leon Lin's code-only launch video, about 8 hours, 85K views. #### 📄 ¶229-¶245 Fixes and your turn - ¶230-¶233: failures with fix prompts (placeholders); two-frame comparison when a fix doesn't hold. - ¶234: add each working fix to the render contract. - ¶235-¶243: Experiment 3 steps; change words first, then one number at a time. - ¶244-¶245: the author posts every piece with its prompt and follow-ups; ask to follow and bookmark. ## 📚 References Checked against the 74-entry link list. The article links its embedded social posts and one live demo. It does not link the technical sources it names, so claims about those have "no link in the captured markup," not a stronger label. 28 of the 74 entries are "View content" links that all point at the thread URL itself. **The article** - Opus 5.5 Motion Lab: Make Claude Animate Like a Studio, @flxrnc, X (Twitter): https://twitter.com/flxrnc/status/2105311836190978392. The 28 "View content" entries point to https://twitter.com/flxrnc/status/2105311836190978392/?rw_tt_thread=True and stand in for the prompts, follow-ups, code and tables (¶7, ¶10, ¶12, ¶44, ¶46, ¶48, ¶50, ¶70, ¶85-¶86, ¶89, ¶103, ¶122, ¶126, ¶140, ¶142, ¶144, ¶147, ¶155, ¶157, ¶160, ¶169, ¶195, ¶213, ¶218, ¶220, ¶231, ¶233). **Embedded posts the article cites (link list entries)** - kirill sh, @shneural, "Posted Sep 24, 2026 at 3:54PM": https://twitter.com/shneural/status/2103151003272962130 (¶28-¶31). - donald, @donaldjewkes, "Posted Sep 23, 2026 at 4:44PM": https://twitter.com/donaldjewkes/status/2102801274173587569 (¶35-¶37). - NotinReality, @other__reality, "Posted Sep 22, 2026 at 9:45PM": https://twitter.com/other__reality/status/2102514581684052169. This post's text isn't attributed by name in the article. The link order suggests it is the quoted post behind ¶38, but the capture doesn't say so. - deckard, @slimer48484, "Posted Sep 9, 2026 at 6:22PM": https://twitter.com/slimer48484/status/2097752569212756134 (¶34, ¶39). - Kevin Ngo, @kevin_t_ngo, "Posted Sep 22, 2026 at 4:40PM": https://twitter.com/kevin_t_ngo/status/2102437977435893771 (¶66-¶68). - Shikhar, @xikhar, "Posted Sep 27, 2026 at 12:14AM": https://twitter.com/xikhar/status/2104001664793600012 (¶95-¶98). - Miles, @uxmiles, "Posted Sep 28, 2026 at 5:04PM": https://twitter.com/uxmiles/status/2104618175803609305 (¶173-¶175). - Vox, @Voxyz_ai, "Posted Sep 24, 2026 at 1:40PM": https://twitter.com/Voxyz_ai/status/2103117246860345550, with a shortened link https://t.co/BE1xui93rF (¶176-¶185). - Chubby, @kimmonismus, "Posted Sep 23, 2026 at 7:36PM": https://twitter.com/kimmonismus/status/2102844654169575547 (¶198-¶202). - Meng To, @MengTo, "Posted Sep 23, 2026 at 2:03PM": https://twitter.com/MengTo/status/2102760783344189761, live site https://valley.mengto.here.now (¶203-¶206). - Leon Lin, @LexnLin, "Posted Sep 27, 2026 at 9:56AM": https://twitter.com/LexnLin/status/2104148233106723099 (¶224-¶227). The post says "repo in replies," but no repository link appears in the capture. **Named sources with no link in the captured markup** - Anthropic's Opus 5.5 prompting guide and frontend-design skill (¶101, ¶112), the dense-chart and effort-setting statements (¶65, ¶93), Claude Code docs and installer (¶84), and fast-mode pricing (¶94). - Gideon Frost of OrcaRouter (¶60), Eric Kang on the Hugging Face blog (¶63), Rexan Wong (¶118), and Andrey Sitnik's Easings reference (¶121). - Material 3 motion rules (¶123, ¶125), Nielsen Norman Group on 100 ms (¶128), WCAG (¶134), and X's media docs and Help Center (¶219). - Remotion, HeyGen HyperFrames, GSAP, Motion and three.js release and license claims (¶189-¶194), and Thomas and Johnston's The Illusion of Life (¶137). **Primary sources worth pulling when you verify (not in the link list)** - WCAG 2.2: https://www.w3.org/TR/WCAG22/ (success criteria 1.4.3 and 2.3.1 cover the 4.5:1 and flash numbers). - Easings reference: https://easings.net/. - Remotion: https://www.remotion.dev/. HyperFrames repository by HeyGen. GSAP: https://gsap.com/. Motion: https://motion.dev/. three.js: https://threejs.org/. **Images.** The imgs list shows 19 images: 1 with the article title as alt text, 11 with no alt text, and 7 with the generic alt "Image." ## 🗣️ Paraphrase The author says the "one prompt" posts that fill the feed since Opus 5.5 launched on September 22 skip the middle part, the follow-ups. This article prints three animations with every prompt and every follow-up, plus the sound pass. The core claim is that Opus 5.5 writes code, not pixels, so a render-look-fix loop with frame-numbered notes can finish a piece, and that replacing adjectives with numbers (roughness, field of view, BPM, damping ratio) takes decisions back from the model. The author proves the point with a side-by-side: the same earbuds scene from a 3-adjective prompt and from a 448-word directed prompt. The first comes back with a glow ring, bloom and particles. The second takes 15 follow-ups and looks designed. The article adds a setup guide for the Claude app and Claude Code, a prompt template and ban list, a cheat sheet of easing, spring, duration, beat-grid and loudness numbers, a mascot experiment built on five animation principles, a list of engines, export advice for X, and a closing assignment: run the opening prompt with your own words, post it with every follow-up, and change one number at a time. The author's stated rule is that receipts persuade and "one prompt" breeds suspicion. ## 📰 Newsletter **Posting an AI animation? Print the follow-ups.** @flxrnc published a long X article on Opus 5.5 motion work with an unusual rule: every clip ships with its exact prompt and every follow-up, "no fake one-shots." The author's three loops took about 2, 2 and 1.5 hours of picture work, plus 45 to 90 minutes of sound each, and 10, 15 and 4 creative follow-ups. The practical part is the numbers. Replace "premium" with a roughness value, "cool" with a three-light studio, "animated" with a beat grid and a damping ratio (0.7 to 0.85 reads premium, per the author). Write notes as frame numbers: "the punch at frame 84 peaks late, move it 2 frames earlier." Render frames 0 to N-1, never N, or the loop hiccups at the seam. Read it with a filter. The prompts, code, tables and render contract are missing from the captured text, and the times, counts and view totals are the author's own. Treat it as a method to try, not a result to cite. Original: [Opus 5.5 Motion Lab](https://twitter.com/flxrnc/status/2105311836190978392). ## ✅ Action items These items concern how your teams document, teach and demo generative-motion work. They don't assess whether the author's numbers are right. #### DevRel - Developer advocacy: before you publish an AI-made demo clip, include the prompt and the follow-up count, as ¶186 recommends, and say which model and effort setting you used. - Education: build a 30-minute workshop from the lab loop (¶152-¶167): render stills, write frame-numbered notes, fix, re-render. Run it on your own product loop and keep the receipts. - Developer experience: test the render.mjs approach (¶88-¶90) on macOS and Windows yourself. The script is missing from the capture, so write your own from the description and check that `?t=` freezing gives identical frames across runs. - Developer marketing: don't call a clip "one prompt" if a spoken brief or a long autonomous run came first (¶40). State the real count. - Community management: invite members to post their prompt and every follow-up in the first reply (¶241), and set a thread format for it. - Internal/external: if someone asks whether Opus can do motion design, answer with the method (code, frame-exact renders, self-checking from stills) and the unverified parts (times, counts, view totals), not with a verdict. #### OSPO - Policy and governance: check the license terms the article states before adopting an engine: Remotion's company-size rule (¶190), HyperFrames Apache 2.0 (¶191), GSAP's no-competing-builder clause (¶192), Motion MIT (¶193). Confirm each against the vendor's own page. - License compliance: add the chosen engine and any font or voxel asset to the dependency inventory with its license. - Security management: generated HTML run in a headless browser (¶88) executes arbitrary code. Render in an isolated environment and don't point it at secrets or internal URLs. - Software inventory: no direct action from this source beyond the engine entries above. - Developer support: prepare an answer for "why does my GIF run 4% fast" using the 40/50 ms delay pattern (¶222). - Community engagement: no direct action from this source. - Strategic alignment: no direct action from this source; nearest relevance is that code-drawn motion is reproducible and diffable, which fits open-source review habits. #### Community - Maintainers: if you maintain an animation library named in ¶189-¶194, check whether its Agent Skills, license text and docs say what this article says. - Contributors: write a "directed prompt template" snippet (¶102-¶107) for your project's contribution guide if you accept AI-made demos. - Sponsors: no direct action from this source. - Ecosystem: no direct action from this source; nearest relevance is that several named engines are open source with different license terms. #### WordPress - Core: no direct action from this source; nearest relevance is that Interactivity API and CSS animation code face the same easing and reduced-motion questions. - Core Performance: no direct action from this source. - Core AI: no direct action from this source; nearest relevance is the self-check pattern in ¶153 (render, look, note, fix) for agent-driven UI work. - Core Program Team: no direct action from this source. - Mobile: no direct action from this source. - Playground: try a Playground blueprint that loads a single-file HTML animation and a frame-freeze URL parameter, as a shareable test bed for the lab loop. - Community: no direct action from this source. - Training: add a motion-numbers module (easing, damping ratio, stagger, holds, ¶119-¶134) to a Learn WordPress workshop on block and theme animation, after verifying each number. - Marketing: if you post release or event loops, follow the sizes and limits in ¶219-¶223 only after checking X's current media docs. - Support: no direct action from this source. - Design: adopt the ban list (¶108-¶115) as a review checklist for AI-assisted motion, and extend it after each first render, as the article says. - Documentation: document any in-product animation with numbers: duration, easing, and the reduced-motion behavior. Use the cheat sheet (¶123, ¶128) as a starting reference, not a spec. - Photos: none. - TV: none; nearest relevance is that a screen-recorded demo is not frame-exact (¶82), so use a rendered export for recorded tutorials. - Polyglots: no direct action from this source. - Accessibility: apply the article's own WCAG numbers (4.5:1 text contrast, no more than 3 flashes per second, ¶134) and add prefers-reduced-motion to any code-drawn loop. The article doesn't mention reduced motion. - Meta: no direct action from this source. - CLI: none. - Hosting: no direct action from this source. - Tide: none. - Openverse: none. - Themes: when a theme ships hero motion, define durations, easing and the reduced-motion fallback as design tokens (¶127 says "use only these tokens"). - Plugins: no direct action from this source. #### The WPOCC - Programs: no direct action from this source; nearest relevance is a "show your receipts" norm for AI-assisted community demos. - Projects: no direct action from this source. - Funding experiments: none. ### FAIR - Technical independence: no direct action from this source. - Policies and covenants: no direct action from this source. - Productized governance artifacts: no direct action from this source; nearest relevance is that a printed prompt-plus-follow-up log is a provenance artifact for AI-assisted output. ## 🧪 Hypotheses 1. Replacing three vague adjectives with a numeric brief (materials, light strengths, FOV, BPM, damping) cuts the creative follow-ups needed to reach an acceptable result, compared with the same prompt without numbers. 2. Frame-numbered notes ("frame 84 peaks late") resolve in fewer rounds than feeling-based notes ("make it pop") on the same piece. 3. A named ban list reduces the share of default-look elements (glow, bloom, particles, dark backgrounds) in the first render, and extending it after each render lowers that share further. 4. A numeric audit of live scene state across all frames (the author's 71 checks) catches timing errors that a human review of stills misses. 5. Sound written against frame numbers and verified by decoded audio syncs to within 1 frame without anyone listening. 6. Viewers rate a loop that follows the article's loudness, contrast and flash limits as equally appealing to one that doesn't, while accessibility improves. ## 📊 Methodology - **H1.** Pick 10 subjects. For each, write prompt A (3 adjectives) and prompt B (the template in ¶102, filled). Use the same model, effort and engine. Count creative follow-ups to a fixed rubric score, with two blind raters. - **H2.** Give the same 8 first renders two follow-up styles, chosen at random: frame-numbered and feeling-based. Count rounds to the same rubric score and note which note types needed clarification. - **H3.** Run 20 first renders per condition (no ban list, article's ban list). Count default-look elements by checklist: glow, bloom, particles, dark background, fade-and-slide-up. - **H4.** On 10 loops, run a script that reads scene state per frame against the prompt's numbers, and separately have two reviewers check contact sheets. Compare the defects each finds. - **H5.** Decode each finished MP4's audio, find each hit's peak, and compare it with the picture frame. The author's claim is 1 frame after; measure the real offset. - **H6.** Show 100 viewers randomized pairs of loops, with and without the 4.5:1 and 3-flash limits. Collect appeal ratings and an accessibility check against WCAG 2.2 criteria 1.4.3 and 2.3.1. ## ❓ Questions - Where are the exact prompts, follow-ups and render contract the article promises? The capture holds 28 placeholders and no repository link. - How was "about 2 hours" measured: wall-clock time, human attention, or agent runtime? The article also cites 12-hour and 8-hour runs for other creators (¶36, ¶225). - What is the basis for the damping-ratio bands (0.7 to 0.85 premium, 0.4 to 0.6 playful, ¶124)? They are presented as rules without a source. - Does the "better at seeing what changed between two versions" claim (¶65) come from an Anthropic document, and does it hold for a 12-frame contact sheet at phone-size detail? - How did the 71 numeric checks (¶141) get written: by the author or by the model? If the model wrote the checks, they test the model against itself. - What does the vague prompt A look like with a banned-list line only, with no numbers? The test changes two things at once (numbers and ban list). - How does the loop hold up with effort set to low or high, given that the author used medium throughout (¶93)? - Does the 5-minute spoken brief behind a "one prompt" post (¶36) count as direction the viewer can't see, and how should captions describe it? ## 🔮 Future research - **Controlled prompt study.** Run the author's A versus B test across several models and subjects with blind ratings, so the claim doesn't rest on one pair. - **Follow-up economics.** Log time, tokens and dollars per creative follow-up, since the article gives time for only some pieces and a cost for just one (Vox, about $90). - **Self-check reliability.** Test whether a model reading contact sheets and speed graphs catches planted timing errors at known rates. - **Audio without hearing.** Compare frame-number-scored sound against human-mixed sound on sync and loudness, and measure phone-speaker response given the author's 60 Hz finding (¶49). - **Accessibility of code-drawn motion.** Study reduced-motion handling, flash limits and contrast in AI-made loops, using WCAG 2.2. - **Provenance norms.** Test whether printing every prompt and follow-up changes viewer trust, as ¶186 asserts. ## 💡 Implications - **Practice.** The method treats the prompt as a specification and the model as a renderer: numbers in, frames out, notes by frame. Teams can copy that structure for any generated UI motion. The article's own cheat-sheet numbers are starting points, per ¶128, and need verification against the cited standards. - **Policy.** If a team publishes AI-made clips, a rule such as "post the prompt and every follow-up" (¶186, ¶241) gives an audit trail. It also exposes cost and time, which marketing may not want public. - **Security.** Running model-written HTML in a headless browser (¶88) and installing a CLI from a one-line installer (¶84) both need isolation and a check of the source. The article gives no security guidance. - **Education.** The lab loop (¶152-¶167) is a teachable procedure with a clear stop condition (run out of notes), and the feeling-to-number table (¶162-¶167) trains a vocabulary designers can reuse without AI. - **Funding.** Time claims (about 2 hours, 8 hours, 12 hours) and the Vox cost ($90 at API prices) give a rough budget range, but they come from self-reports. Fast mode at $8 and $40 per million tokens (¶94) is a cost lever for iteration sprints. ## 🧠 Learning styles assessment - **Reading level estimate.** Flesch-Kincaid Grade Level of about 8 and Flesch Reading Ease of about 62, from a rough sentence and syllable count over the roughly 6,500-word capture. Sentences are short and direct, but numbers, units and tool names (raymarched, damping ratio, dBTP, LUFS) raise the load for readers outside motion design. This is an estimate, not a tool run. #### 👀 Visual learners - Present: 19 images (title image, screenshots or clips) and a promised pairing of every loop with its video. The captured text can't show what they depict, and 11 of 19 have no alt text. - Improve: add before-and-after stills for A versus B and a labeled contact-sheet example, each with alt text. #### 🔊 Aural learners - Present: the article covers sound in depth (¶11, ¶49, ¶143, ¶210-¶212) and embeds videos with sound, per the text. No transcript or narration appears in the capture. - Improve: provide a short narrated walkthrough of one follow-up round, and captions for the three loops' sound design. #### 🗣️ Verbal learners - Present: short paragraphs, labeled parts, and lists with one idea per bullet. The feeling-to-number translations (¶163-¶167) give a plain vocabulary. - Improve: define LUFS, dBTP, raymarched, damping ratio and subframe blur in one clause each on first use. #### 🧩 Kinesthetic learners - Present: the strongest feature. Setup paths for the app and for Claude Code (¶73-¶92), a render command (¶90), a URL to freeze a frame (¶154), and an Experiment 3 assignment with steps (¶236-¶242). - Improve: put the code and prompts in the post as text or a repository, since the capture holds placeholders, and so readers can copy and run without hunting. #### 🧠 Logical learners - Present: formulas (frames per beat = fps x 60 / BPM, ¶130), a drift example (2 frames off after 32 beats), and a controlled A versus B design with one variable changed. - Improve: show the arithmetic for 14.06 versus 14 frames per beat, state each rule's source, and report variance across more than one run. #### 🧑‍🤝‍🧑 Social learners - Present: many embedded creator posts, a "post your prompt in the first reply" norm (¶241), and a follow ask (¶245). - Improve: offer a shared place or hashtag where readers post receipts, and a template for the first reply. #### 🧘‍♀️ Solitary learners - Present: the article is self-contained and the assignment can be done alone, at medium effort, in the Claude app. - Improve: add a self-check rubric so a solo learner can tell when "run out of notes" (¶240) has actually happened. #### 🌿 Naturalistic learners - Present: little; the content is software and screen graphics. Real-world physics appears in analogies (light beats sound, ¶211) and animation principles drawn from physical motion. - Improve: tie spring and arc examples to observable real-world motion (a bouncing ball, a pendulum) with a side-by-side. ## 📊 Accessibility assessment Judged from the captured text, link list and image list. Rendered contrast, heading levels in the real markup, and the behavior of embedded video are unknowable from this capture. #### Screen reader semantics - Present: the capture shows "##" markers on some lines (for example ¶26, ¶58, ¶72, ¶98, ¶135, ¶151, ¶186), which suggests heading or callout styling, but several of them are full sentences (¶58, ¶72, ¶186), not topic headings. If they render as headings, a screen reader user hears long sentences in the heading list. - Improve: use short topic headings and keep conclusions in body text (1.3.1 Info and Relationships, 2.4.6 Headings and Labels). Of 19 images, 11 have no alt text and 7 use the alt "Image," so 18 of 19 images have no useful text alternative (1.1.1 Non-text Content). The one with alt text repeats the article title. - 28 prompts, code blocks and tables are replaced in the capture by "View content" placeholders. If the published page does the same, that content may be inaccessible to anyone who can't open the linked thread (1.1.1, 1.3.1). #### Keyboard navigation - Present: the capture shows ordinary links. The article's own advice includes a space bar that pauses and a frame counter (¶81), which suggests keyboard control in the demos. - Improve: verify embedded video players and the "View content" links are keyboard operable with visible focus (2.1.1 Keyboard, 2.4.7 Focus Visible). A 74-link article with many identical "View content ↗" anchors makes tab navigation hard to scan. #### Contrast and readability - Present: unknowable from text. The article cites a 4.5:1 text target for video (¶134), which matches WCAG 1.4.3 for normal text. - Improve: check the page's body text, link color and any code-block styling against 1.4.3 and 1.4.11 Non-text Contrast. The article's own loops use dark backgrounds in some cases (¶57 bans them by default), so check captions over video. #### Text resizing and reflow - Present: text content that should reflow. Some lines are long strings of numbers and units. - Improve: check 1.4.4 Resize Text and 1.4.10 Reflow at 320 CSS px and 200% zoom, especially for the code and table content missing from this capture, which is where reflow usually breaks. #### Media and captions - The article embeds video loops with sound (¶11, ¶212) and references X media. The captured text doesn't mention captions or transcripts. The animations are non-speech, but sound carries meaning in the sync design (1.2.1 Audio-only and Video-only, 1.2.2 Captions, 1.2.5 Audio Description). Offer text alternatives that describe the motion and the sound cues. - The article recommends keeping flashes under 3 per second (¶134), which maps to 2.3.1 Three Flashes or Below Threshold. It doesn't mention prefers-reduced-motion or a pause control for looping clips (2.2.2 Pause, Stop, Hide). #### Forms and interactions - Present: no form appears in the capture. Interactions are links to posts and a live site (valley.mengto.here.now). - Improve: the interactive demos the article cites (a playable boat scene, an interactive black hole lab) need their own keyboard and pointer-alternative review (2.1.1, 2.5.1 Pointer Gestures, 2.5.8 Target Size Minimum). The article doesn't address this. #### Standards reference - WCAG 2.2 AA items to verify: 1.1.1 for 18 of 19 images, 1.2.2 and 1.2.5 for the video loops, 1.3.1 and 2.4.6 for heading use, 2.2.2 and 2.3.1 for looping and flashing motion, 2.4.4 Link Purpose for the repeated "View content ↗" anchors that all go to the same thread URL, and 1.4.3 for text over video. The article cites two WCAG-derived numbers in ¶134 without naming success criteria, and neither refers to prefers-reduced-motion. ## 🌍 DEIB assessment #### Cultural breadth - Present: the examples span English-language X creators and tools from several countries; the article is English only. Loudness, canvas and X-specific limits assume a social-media delivery context. - Improve: note what changes for other platforms, regions, and for audiences on low-bandwidth connections where a 6-second MP4 is still large. #### Inclusive language - Present: plain, direct language with idioms such as "cooked," "receipts," and "No fake one-shots," plus a quoted post that calls a model's output "honestly terrible" and says a company has "awful taste" (¶31). That line is the quoted poster's, embedded in the article. - Improve: gloss "cooked" and "receipts" for readers of English as a second language, and mark quoted trash-talk as a quote, not the author's view. #### Neurodiversity support - Present: numbered steps, short bullets, a ready-made assignment (¶236-¶241), and a "one or two per round" limit (¶161) that reduces overload. - Improve: the article is long (about 6,500 words in the capture), and the numbers come in dense runs (¶128). A one-page summary of the starting numbers would cut working-memory load, and a note on flashing and motion intensity would help readers with vestibular or seizure sensitivities. #### Representation - Present: named creators include kirill sh, donald, deckard, Kevin Ngo, Shikhar, Miles, Vox, Chubby, Meng To and Leon Lin. The source text uses no pronouns for any of them, and this analysis uses none either. - Improve: the roster is a set of X creators and one studio founder. Ask whether motion designers working without a model, or with different tools, appear in the evidence, since the article's case rests on viral posts. #### Psychological safety - Present: the "no fake one-shots" stance invites honesty about iteration, and the author asks readers to post failures and counts (¶186). The article tells people to expect 2 to 4 rounds as normal (¶153). - Improve: the quoted trash-talk about a competitor's taste (¶31) and the viral-count framing may pressure beginners to compare their first loop with 1.8 million-view posts. Add a line that the first run is a draft and that view counts reflect distribution as much as quality. ## ✅ Logical validity #### Strong reasoning - **Controlled comparison, ¶41-¶42, ¶47.** One subject, one model and one technical line, with only the direction changing. The left side got 0 follow-ups by design and ran in a separate context, so it didn't see prompt B. That design isolates direction better than the viral showreel comparison (¶28, ¶32). - **Honest attribution of the "one prompt" claim, ¶40.** The author reads donald's own post (5 minutes of speaking, 12 hours of Claude work) against its caption and says direction did the work. - **Mechanism for the code-over-pixels claim, ¶60-¶64.** Reproducible frames, cheap edits and self-checking each follow from the file being a program. "Pure function of t" gives a testable property: render frame 84 twice and compare. - **Measured fixes, ¶11, ¶49, ¶141.** Sound fixes cite measured offsets (8 to 11 ms late, 16 ms codec padding, 80% of energy under 60 Hz), and a numeric audit logs 71 checks with 6 initial failures. Reporting failures makes the process credible. - **Arithmetic stated, ¶130, ¶221-¶222.** The beat-grid drift (2 frames off after 32 beats) and the GIF delay fix (five frames at 40 ms plus one at 50 ms for a 6.00-second loop) can be checked by hand. #### Fallacies No instructions addressed to an AI, an assistant or "the reader" appear in the captured text. The article gives instructions to human readers (paste this prompt, install this tool, run this command, follow the author), and the pasted prompts addressed to Claude are among the 28 blocks missing from the capture, so this analysis can't say what they contain. What I checked: whether each causal claim follows from the evidence given, whether numbers carry sources, and whether the article treats self-reports as proof. These are gaps and weak points, not formal fallacies in every case: - **Single-case generalization, ¶47, ¶58.** "Adjectives are where the default look lives" rests on one A versus B pair with one model and one subject. *Fix:* run the pair on several subjects, or say "in this test." - **Confounded comparison, ¶53-¶57.** Prompt B adds numbers, a ban list, camera and beat direction, and 15 follow-ups, while prompt A gets 0. The article credits "direction" in general, but follow-up rounds and the ban list are separate variables. *Fix:* add a prompt-B-without-follow-ups arm. - **Popularity as evidence, ¶28, ¶40, ¶98, ¶175.** View counts (1.8 million, 3.6 million, 1.75 million, 590K) appear as signals of quality. Views measure reach. *Fix:* label them as reach. - **Appeal to authority without a link, ¶65, ¶93, ¶94, ¶101.** "Anthropic says" claims (dense-chart accuracy at low effort, medium matching Opus 5 at high, fast-mode prices, the prompting guide) have no link in the captured markup. The author doesn't claim a commercial tie, and Opus 5.5 is Anthropic's model, so cite the source and let readers check. *Fix:* link each. - **Unverified numbers, throughout.** Run times, follow-up counts, the 71 checks, the GPU model, 885 followers and view counts are unsourced in the capture. The prompts and logs that would back them are the 28 missing blocks. *Fix:* link a repository with the logs. - **Survivorship in examples, ¶172-¶207.** The cited creators are successful posts. The article says "Other people's numbers look the same once they are honest about them" (¶172), but it doesn't show failed attempts from others. *Fix:* include a failure case. - **Fast-mode and cost framing, ¶94.** Prices are stated as fact without a link and may change. *Fix:* link the pricing page and date the number. - **Internal tension on "no fake one-shots," ¶2 versus ¶36.** The author criticizes loose "one prompt" claims, then reports on a 12-hour autonomous run as direction. This is consistent, but the article doesn't say whether the author's own "10 follow-ups" count includes autonomous agent loops. *Fix:* define "follow-up." #### Tone of voice - Confident, practical and tutorial-like, with imperatives ("Screenshot this one," "Fill every line before you press enter"). The article is practitioner advocacy for a workflow built on one model, with a self-promotional close (¶244-¶245). The author discloses no sponsorship in the capture, and the commercial interest visible in the text is audience growth. Opus 5.5 is Anthropic's model; the article's praise of it should be read as one practitioner's report, and the neutral facts are the method and the arithmetic. ## 🛡️ DARVO analysis Not applicable to this source. DARVO (Deny, Attack, Reverse Victim and Offender) describes a response pattern to an accusation. This is a tutorial and showcase with no accusation against the author. Each prong is still checked. - **Deny.** None. The article doesn't deny any fault or any claim made against it. - **Attack.** Mild and indirect. It quotes a post that says a competing model has "awful taste" (¶31) and the author says "Credit where it's due" (¶34) in the credit split. The author's own critique targets a practice, "one prompt" captions that hide follow-ups (¶2, ¶186), not a named person. - **Reverse victim and offender.** None. No party is cast as a victim or offender. - **Reporter's voice.** The author's voice is instructive and promotional, not defensive. ## 🎯 Conclusion The article argues that Opus 5.5 produces finished motion work when you replace adjectives with numbers, drive time from the frame number, and run a render-look-fix loop with frame-numbered notes. It supports the point with three self-made loops, a vague-versus-directed pair of earbuds clips, a cheat sheet of motion numbers, and a set of other creators' posts. It also contributes a useful ethic: print every prompt and every follow-up. What the capture does not establish: the prompts, code and tables are missing (28 placeholders, all linked back to the thread), so "the exact prompt" can't be checked. Times, counts, the 71-check audit, the audio measurements and view totals are the author's own and have no linked log. Statements attributed to Anthropic, Material 3, Nielsen Norman Group, X's docs and the engine vendors have no link in the captured markup. The A versus B comparison varies several things at once. The method is plausible and internally consistent, and the arithmetic (beat grids, GIF delays, loop phases) can be checked by hand. **Next steps for readers:** run Experiment 3 (¶236-¶242) on your own product with the numbers fixed, log your own follow-up count and time, verify the cheat-sheet numbers against the cited standards before teaching them, and add prefers-reduced-motion and a flash check to any loop you ship, because the article doesn't cover them. ## 🔄 Refactor Appendix — Opus 5.5 Motion Lab: Make Claude Animate Like a Studio > Density: 2–5 Literature atoms (H3 pad), 0–3 Permanent candidates (H4 pad). > Emitted: 5 Literature, 2 Permanent. Resonance filter applied: view counts, engine licenses > and export limits were dropped as short-lived details. > > Konik claim-header rules: present tense, declarative, ≤10 words, searchable, unique. > > Atomicity: one declarative sentence per atom. "and"/"also" → split. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Split Pad #### [H1] Follow-ups are the part nobody posts > [!quote] [¶2] > Almost nobody posts the middle part: the second prompt, the third **Context & use**: - ¶2, the article's problem statement about "one prompt" posts since the September 22 launch. - A claim about posting practice the reader can check on any AI-demo feed. **Tags**: #xref #articleseed  #### [H2] "One prompt" can hide hours of direction > [!quote] [¶40] > The one prompt was a spoken brief plus 12 hours of autonomous work: direction did the heavy lifting. **Context & use**: - ¶40, the author's reading of donald's post (5 minutes of speaking, 12 hours of Claude work). - Use it as the counter-example when someone cites a "one prompt" clip as proof of the model alone. **Tags**: #xref #share  #### [H3] Numbers replace adjectives as decisions > [!quote] [¶58] > Adjectives are where the default look lives. Every adjective you replace with a number is a decision you take back from the model. **Context & use**: - ¶58, the thesis line of Experiment 1. The A versus B pair (¶41-¶57) is the only evidence in the article. - Test it with a second subject before relying on it. **Tags**: #xref #aphorisms  #### [H4] Receipts persuade, "one prompt" breeds suspicion > [!quote] [¶186] > Count your follow-ups and post every prompt with the work. **Context & use**: - ¶186, the article's provenance norm, repeated at ¶241 and ¶244. - Applies to any AI-assisted demo, including documentation and training clips. **Tags**: #xref #share  #### [H5] The model is shared; the numbers are yours > [!quote] [¶243] > The model is the same for everyone. The numbers you give it are yours. **Context & use**: - ¶243, the closing claim. It frames direction as the author's differentiator, not the tool. - Pair with H3, and with the caveat that the numbers' sources are mostly unlinked. **Tags**: #xref #aphorisms  ### 🧠 Atomic Split Pad #### [P1] Publish the follow-up count with AI demos **Claim** (one sentence, Courtney's voice): An AI-made demo should ship with its prompt and its follow-up count, because "one prompt" hides the direction and rounds that did the work. **Evidence pointer**: [[#H1]] · [[#H2]] · [[#H4]] **Scope & conditions**: Holds when the demo is meant to show what a person can reproduce. Weakens when the clip is clearly a showcase of the model with no brief, as the showreel prompt is (¶32). The article's own counts are unverified because the logs sit in the missing blocks. **Links**: - [[Provenance for AI-assisted output]] — because extends - [[Developer advocacy demo standards]] — because applies-in launch clips and workshop recordings - [[One-shot AI demo claims]] — because contrasts on hidden iteration  #### [P2] Specify numbers, then verify the source **Claim** (one sentence, Courtney's voice): Replacing adjectives with numbers gives the model less room to default, but each number needs a source before I teach it as a rule. **Evidence pointer**: [[#H3]] · [[#H5]] **Scope & conditions**: Holds for motion and visual specs where a number maps to a result (damping ratio, FOV, BPM). Weakens when the number is a taste threshold with no standard behind it, such as "0.7 to 0.85 feels premium" (¶124), which has no link in the capture. **Links**: - [[Prompt specificity and default styles]] — because extends - [[Motion design numbers and easing]] — because applies-in documentation and training material - [[Unsourced numbers in tutorials]] — because contrasts on verification  ### ✏️ Fleeting Queue - [F1] Re-run the earbuds A versus B test on a second subject and add a "prompt B without follow-ups" arm #fleeting #inbox - [F2] Check prefers-reduced-motion and the 3-flash limit on any code-drawn loop we ship #fleeting #inbox - [F3] Open question: where are the prompts, code and logs that the 28 placeholders replace? #fleeting #inbox - [F4] Article seed: "What a demo receipt should include," using this article's prompt, follow-up and timing log as the model #fleeting #inbox #articleseed - [F5] Verify the damping-ratio bands and easing rules against Easings and Material 3 before teaching them #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Run Experiment 3 on a Pantheon demo loop and log prompt, follow-up count and time — before the next DevRel demo #things - [ ] Draft a "show your prompts and follow-ups" rule for AI-made demo clips — DevRel and Documentation review #things  📄 **Full text:** [[Full Document Contents/Tweets/Opus 5.5 Motion Lab Make Claude Animate Like a Studio.md|Opus 5.5 Motion Lab: Make Claude Animate Like a Studio by flxrnc]] ## Highlights - Since Opus 5.5 launched on September 22, the feed has filled with motion pieces captioned "one prompt". Some really were one line. The most viewed music video was five minutes of talking and a 12-hour run. Almost nobody posts the middle part: the second prompt, the third, the note that says "frame 84 peaks late". ([View Highlight](https://read.readwise.io/read/01m4rm633da2rk85d2nfvan07j)) - Note: [H1] Follow-ups are the part nobody posts - Tags: atom/literature/h1 - It crossed 3.6 million views. Donald calls it one prompt, and in the same post he says he spoke to his computer for five minutes and Claude worked for 12 hours. The one prompt was a spoken brief plus 12 hours of autonomous work: direction did the heavy lifting. ([View Highlight](https://read.readwise.io/read/01m4rm661k9g7ezqz6yff0rftb)) - Note: [H2] "One prompt" can hide hours of direction - Tags: atom/literature/h2 - Adjectives are where the default look lives. **Every adjective you replace with a number is a decision you take back from the model.** -------------------------------------------------------------------------------------------------------------------------------------- ([View Highlight](https://read.readwise.io/read/01m4rm78m26e4g47ykxatg4w4z)) - Note: [H3] heading-level claim - Tags: atom/literature/h3 - **Count your follow-ups and post every prompt with the work. Receipts convince people; "one prompt" makes them suspicious.** ---------------------------------------------------------------------------------------------------------------------------- ([View Highlight](https://read.readwise.io/read/01m4rm79dnbc8ybm6zkp9dcq93)) - Note: [H4] heading-level claim - Tags: atom/literature/h4 - The model is the same for everyone. The numbers you give it are yours. ([View Highlight](https://read.readwise.io/read/01m4rm68xycwt3c534pz2mwwhe)) - Note: [H5] The model is shared; the numbers are yours - Tags: atom/literature/h3