# A skill that writes my thumbnail prompts

> Every key note here gets a claymation thumbnail with the same characters and palette, so I built a skill, [garden-thumbnail-prompt](https://github.com/weijie-ng/garden-thumbnail-prompt), to write the prompt instead of hand-tuning each one. The fix that stuck was a reusable principle, not patching the one broken case a model reaches for first.

I point a [skill](https://agentskills.io/home) at a note. It reads the note, picks the one idea, and stages that idea as a single clay scene with the two reference characters. The style block and the character descriptions are fixed, so the whole set stays in one family. The skill supplies the words; the two characters come from reference sheets I made.

Before any of the automation, I had to lock the cast. Words alone let the characters drift. Ask for a gardener and a clay Shiba twice and you get two different gardeners and two different Shibas. So I drew a reference sheet for each, a gardener I call Ben and a small Shiba Inu robot, pinning down their turnarounds, palette, and expressions. I attach the matching sheet in my image tool and the character lands the same every time.

![Character reference sheet for Ben the gardener: a claymation figure in a beige and blue cable-knit sweater and tan trousers, shown from several angles with facial expression studies and a colour palette.](/garden-assets/tinkering/a-skill-that-writes-my-thumbnail-prompts/hero.png?v=8c4e2da0f1) ![Character reference sheet for the Shiba Inu robot: a rounded tan and cream clay shiba with visible joints, a glowing blue chest core, and a coiled red-tipped antenna tail, shown from several angles with detail callouts.](/garden-assets/tinkering/a-skill-that-writes-my-thumbnail-prompts/shiba-inu.png?v=d00d54054a)

The cast, standardised before any automation: Ben the gardener and the Shiba Inu robot.

Getting there took a few rounds. The loop-engineering note is where it broke. I asked for "a loop" and got a racetrack.

![Claymation scene: a gardener kneels inside concentric oval clay tracks laid on soil while the robot shiba sits outside them. The rings read as a running track, not a cycle.](/garden-assets/tinkering/a-skill-that-writes-my-thumbnail-prompts/fail_image.png?v=8f5257586f) Before

What "a loop" produced: a running track.

The prompt had asked for the wrong thing in plain sight:

> A round clay *loop-track* laid into a garden bed, with the Shiba Inu robot *running happily around the ring* ... Ben kneels just outside the ring and watches it run on its own.

"Loop-track" and "around the ring" are object words. The generator built the object, and a ring on the ground is a racetrack.

When I pointed that out, the model went for the narrow fix. It wanted to write a loop recipe straight into the skill, a special case for one word. That would have patched this note and left the same trap for the next abstract word, so I steered it the other way: state the principle, not a case. The skill got a rule it could reuse:

SKILL.md

\- Self-running loop or cycle: draw curved vine-arrows or a waterwheel,

\- never a flat ring with a figure running laps.

\+ Sculpt the idea, not its name. The generator renders your nouns

\+ literally, so name the concept's distinctive shape, not a word

\+ that also happens to be an object.

With that rule in the skill, the next run came back neater. Instead of a ring on the ground, it named the shape that actually reads as a loop, the endless-loop ribbon that folds back on itself:

> A single continuous clay ribbon wrapped in green vine, twisting back into itself as one unbroken *endless loop* that stands upright, clearly not a flat circle or ring. A plant grows up through it and blooms where the loop comes back around.

![Claymation scene: a terracotta clay ribbon wrapped in green vine forms a figure-eight endless loop standing upright in a garden. The robot shiba sits in one curve, a sprout rises from the top, and Ben the gardener stands to the side holding a blueprint scroll.](/garden-assets/tinkering/a-skill-that-writes-my-thumbnail-prompts/pass_image.png?v=33db528e65) After

The final thumbnail: one continuous ribbon, no track in sight.

The specific prompt only fixes one note. The rule fixes every note like it. Left to itself the model reaches for the specific fix first, so the part that scales, turning a fix into a rule, is on me.