Corpus evidence

38 Claude Code sessions, 2026-09-25 to 2026-11-02, 587 prompts after filtering harness noise. 31% are a pasted screenshot or contact sheet with no text; the 406 text prompts have a median length of 14 words. The author iterates in many short turns and uses images as the main feedback channel.

Share of the 406 text prompts: "make more / less" delta 35%; a direction word 22%; names an element inside the comp 21%; correction of a misread 17%; targets a numbered comp 14%; camera vocabulary 13%; removal or ban 13%; attaches a file 12%; easing vocabulary 10%; numbered list 9%; named style or era 9%; keep-clause 7%; multiplier 6%; timeline anchor 5%; asks to expose params 3%.

The 17% correction rate is the cost the skill removes. Nearly every correction maps to an empty slot in the checklist.

Opening briefs

The 29 openers (sessions that start with a "create" of 12+ words) are a different genre from the deltas, and they decide the result:

The idea lives in the reference (recreate this <mp4>, dancing like in <youtube>, in the texture of embroidery <img>) and the beats chain turns: a big wave splats across the screen, turning the entire screen into a new texture, the camera zooms out to reveal that the wave washes the new scene into a chinese shanshui painting, the landscape painting unravels into a single thread which eventually becomes a woven running silhouette, the camera zooms out, reveal all ~100 carousels in grids. An opener never picks the safe default look.

The first skill version (2026-11-02, before the rewrite) was extracted from the deltas and produced spec-shaped briefs: ms staggers, px rises, bans, pixelRatio checks, "linear style" and "flat vector" defaults, references found after the beats to match them. In the before/after test the results were indistinguishable from an unassisted agent, because a capable agent already picks sensible easing; what it does not do on its own is commit to a look, copy a striking move, use a real asset and add a turn.

Real openers, all the same skeleton (location, style, numbered beats with inline refs, globals after the list, params, realism demand):

• in stewdio file-b, in the style of a 16bit dungeon crawler game, create the following: + three numbered beats, each with a reference.
• in stewdio file-c, create the following: a text animation where each letter is an individual silver balloon with this type of texture <img>, and each of them animate in from by inflating from a deflated state, and animate out by reversing that process, the inflation and deflation should be very snappy, ease-in-out + expose params to edit the text and the texture and color of the balloon, and the inflation extent and timing + it should be very realistic.
• in stewdio's carousel project, create the following comps + five one-liners + for all of them use white bg and 3 place holder images <three files>.
• in the stewdio file-b, for the main comp, do the following at 7s: (sequence mode: the location line carries the time anchor).

Deltas that landed in one round

for 60, zoom in more by 1.5x / make the final pan 3x faster at top speed / in 17, make the squares just rotate 30 degree clock wise to straighten/align / at 2.5s have the camera zoom fast into the pupil, steep quadratic ease-in / after the orb forms make the pause 1.5s shorter before the drop / make the balloon 70 inflated popin and finish inflating after the entrance.

Deltas that needed follow-ups

Correction families (71 turns)

• Orientation / sense: upside down, other way, clockwise, mirror. Say axis + sense + what "up" means.
• Magnitude: faster, bigger, more zoomed. Multiplier + phase + reference point (the fastest we have had, the original size).
• Which target: wrong file, comp or element. Location prefix and comp number on every change.
• Medium: i meant rebuild the scene in 3d, make the eye an illustrator vector animation then actually revert back to the real eye. Photo / footage / 3D rebuild / vector / shader, up front.
• Lost state: you got rid of all of them (params), keep my new tweaks too. Keep-clause on every edit.
• Over-literal reads: they still need to be individual frames rather than a literal tunnel, it doesn't have to literally contain a square. Describe the composition of parts, not the metaphor alone.
• Pace: there is a slight pause after the dip, shouldn't be a pause when the uis appear. State holds, including "no hold".

Process guidance worth keeping verbatim

• treat each of them as a contraption and animate their special mechanisms (not a default draw-in)
• think of the whole thing as an in-process morph, rather than a complete deconstruction and reconstruction
• use some form of shape morph rather than the tracing play you are doing
• make the shape of the splash more dramatic and irregular / more random and natural and crystal-like (organic over orderly)
• it is still not fully one to one to the one in app, can you check the render properly
• take your time with everything before (do not compress the earlier beats)
• create a copy that tries another approach (branch, do not overwrite)
• no need to do backward directions, they are not a valid variation / mirror are not valid either

How the author used references

From 271 attachments across the transcripts and the prompts that cite them:

1. One reference, one job. Two qualities from outside means two files: a tunnel of this pattern <jpg> in the color way of the iridescent woman <png>. No single file was ever asked for both motion and look.
2. Moving media for motion, stills for look. Every mp4 and gif was cited for a mechanic or transition; every still for a subject, palette, texture or detail.
3. Motion references get the subject swapped. recreate this animation, but make it nyc subway; what is kept is named (only the spiral form factor and the scroll animation), the look is not inherited.
4. References get scoped to a part. the first two patterns in this picture, the outer portion of the first comp.
5. References come out of curated channels. Both of his video motion refs were blocks in his are.na mo-graph channel. Search by channel, never by block keyword.
6. Copies get loyal only after a correction. replicate these more loyally, it doesn't look like the glb i uploaded exactly, why?. Say the fidelity up front, and never re-interpret an uploaded asset.

Fence failures (2026-11-02)

Blake names niche aesthetics himself; other users name mainstream ones in plain words (clean and minimal, lots of white, like stripe's style, like claymation). The "remarkable, not matching" rule was written for the first case and, applied to the second, overrode what the user asked for (2026-11-02: minimal → a graphic score, stripe → the Sessions conference sizzle, claymation → clay.com's marble run; all three read as misses). So:

The second fix (rerun r3, same day) overshot the other way: the agents screenshotted the fence (the stripe docs page, five product-UI screenshots) and fenced the motion too (UI micro-interactions as m1). The builds came out as web pages at browser scale with a small thing moving on them. Both runs missed for the same reason: the fence was treated as a place to copy from, not as a query. So: