ip-as-logo

There is not one line of code in this repository. Five files, zero executable code, zero tests, zero dependencies. The substance is a single 114-line instruction sheet telling an image-generation agent how to draw a mascot, which makes it the best specimen I have torn down for seeing the minimum unit of a skill. But one of the two pillars I had called its core invention lives only in README.md and appears nowhere in SKILL.md — and SKILL.md is the only file the agent loads, so its runtime effect is zero. The other pillar, a set of numeric budgets, was a value that flipped four times inside 45 hours. The verdict is rejection, and the ground is not taste: this machine has zero image-generation tools, so installing it leaves nothing to call.

What it is — the product is 114 lines of prose

One skill in the Agent Skills format. SKILL.md plus its frontmatter is the whole of it — no parser, no validator, no post-processor, no style reference images. So “installing” is copying a file, “fixing a bug” is editing a sentence, and the concept of a release does not exist.

Item	Measured 2026-08-22
Files	5 · 739,374 B checked out — 96% of that is one showcase image (713 KB, zero runtime role)
The product	SKILL.md, 114 lines / 16,976 B. A 13-step workflow plus an 8-line prompt skeleton
Code	0 lines. The GitHub languages API answers with {}
Dependencies	0. The one runtime dependency is “the host agent owns an image generator”
Scale	3,481 stars · 171 forks · 3 watchers · 1 contributor · 0 external PRs
Versioning	0 releases · 0 tags · 0 changelog — users tell versions apart by commit timestamp
Lifespan	29 commits packed into 45 hours, then a stop

The operating rules all sit in those 13 steps. With no subject given and the current workspace being a product repo, it read-only skims the README, the package metadata and the landing copy → if that still isn’t enough it fires background questions in exactly one round (a second questionnaire is forbidden) → it offers three one-line directions and generates nothing before agreement. On approval it makes two variants per direction, six images in total.

The prompt is pinned down too, as a fixed eight-line format: Background / Subject / Complexity / Color behavior / Composition / Style / Finish / Constraints. So what this repository actually is, is one hardened prompt plus a procedure for when and how to fire it.

The techniques — aesthetics written as numbers, not adjectives

Technique	What it says
Complexity budget	Outer silhouette = 4–7 large shapes, ≤1 species-identifying feature, ≤2 interior colour regions, 85–95% canvas coverage. It closes with a shrink test: “if it doesn’t read at 32×32, enlarge, merge or delete”
Three-colour semantic palette	Colours are counted as meanings, not pixels. Value and saturation variation is explicitly declared incidental drift inside one colour family, not a violation
Corner emergence + deterministic assignment	The character rises out of the lower-left or lower-right corner rather than the centre, and placement is fixed by label rather than by feel (A1·B1·C1 = left, A2·B2·C2 = right). An odd count must have its imbalance written into the report
Purpose-hiding prompts	Even though the repo is named ip-as-logo, the model is never told the output is a logo, brand mark or app icon. The reasoning is that naming the purpose drags the model toward cards, frames, masks and text
Constraint routing	Negative constraints travel by a different channel per model generation — newer models get one inline Constraints: line, older ones exposing negative_prompt get the dedicated parameter and the inline line is deleted (no double delivery)
One-pass unvetted hand-off	The most unusual part. Inspecting, grading, retrying or post-processing the output is entirely forbidden; whatever came back is labelled and passed through

The sentence behind the unvetted hand-off is this one — “Treat generation as a stochastic draw, not a conformance test.” It was not that way from the start: the author tore his own review loop out within two days.

What this section cost me to learn — writing a number and enforcing a number are different jobs. The 4–7 and the 85–95% are targets thrown at a probabilistic model in natural language, and the lines of code that count compliance number zero. The author himself renamed the section from “What it enforces” to “What it guides”, conceding that it does not enforce.

What broke — a core invention sat in a file the agent never reads

The rule that “95–100% of open briefs should be a familiar animal” is not in SKILL.md. Grepping, animal appears in two places: the frontmatter lists animals, robots, ghosts, plants and objects as equals, and line 18 of the body says the opposite — “don’t line up three animals with no reason.” 95, clock and vehicle appear zero times. That rule lives only in README.md, and as the README itself states, the agent loads exactly one file: SKILL.md. In the docs and absent from the artifact means zero runtime effect.

The “must read at 32×32” requirement doesn’t lack a judging device — it had one and it was deleted. An earlier revision carried an eleven-item “when to mark as not recommended” checklist whose first line was “it collapses at small sizes.” One commit removed that section and the retry loop wholesale and replaced them with the hand-off rules. Today, inside the same file, line 41 (merge or delete if it’s small) and line 113 (never block hand-off over detail or composition) cancel each other out.

The compatibility claim didn’t hold either. The README lists seven supported agents, but the canonical target list of the one install command the same README prescribes is missing four of them. Conversely, most of the installable targets lack the “built-in generator that returns images as assets” the README declares mandatory. The real intersection is about two, which matches the field reports in the open issues.

And there is one place where it shot its own foot. The author scoped his own “never name the purpose” rule with “this rule applies to generation prompts only” — and then, in the same commit, stripped logo, mascot and app-icon out of the frontmatter description. That sentence is the only text an agent reads when deciding whether to auto-invoke the skill. Not one keyword is left for “make me a logo” to catch on.

What held up is worth recording too. 3,481 stars · 171 forks · 3 watchers matched my measurement, and so did the 739 KB checkout. My first-pass line that “all three external issues went unanswered” I retract — the maintainer replied to all three, but engaged only with the collaboration offer and selectively skipped the two technical defects. That is not neglect, it is selective avoidance, and the distinction is what sharpens the verdict.

Held against my own setup — installing it leaves nothing to call

This skill’s single runtime dependency does not hold on my machine. I swept the full text of all 99 of my SKILL.md files for image-generation vocabulary and got eight hits — every one a false positive, dall matching inside ReadAll and findAll. Zero local diffusion stack, zero generation tools in the session tool roster. Install it and there is no call site.

Held against the remaining intellectual assets, my side was already stronger.

Their technique	My counterpart	Result
Complexity budget (numeric ceilings)	This site’s scripts/check-weight.mjs — 10 numeric ceilings, the derivation of each number re-derived in a comment, and exit 1 blocking the deploy on breach	strict superset
Output review gate (which they deleted)	The mindmap generator’s check.py — three deterministic checks over a probabilistic artifact; without RESULT: PASS it cannot enter a report	opposite direction
One-round question ceiling	My comprehension skill’s SKILL.md — at most one calibration question, none if the answer is obvious. And it is explicit-invocation only, so the ordinary question count is zero	already stronger
Absorbing repo context first	15 global rule documents injected in full at session start, plus the project CLAUDE.md importing DESIGN.md	wiring, not instruction

They wrote numbers down; I enforce numbers. Where the difference comes from is visible in the measurements too — this repository deleted its most quantitative element, the colour band (OKLCH ranges, contrast-ratio targets), 12.5 hours after introducing it, with no stated reason anywhere. Whether it was dropped for having no effect or because the model ignored it is unknowable, which means someone repeats the same investigation six months from now.

There is also no slot for the output. This site’s design contract is locked with “no photographs on cards” and “every body figure is inline SVG”, and the real contents of public/ are a 327 B favicon and two share cards. There is nowhere to hang a 1536×1536 raster mascot. One side risk as well: this skill’s description runs 415 characters and contains none of logo, mascot, brand or icon. It would cost standing tokens and never fire.

And then the last takeaway vanished — my own log had already answered it

In the first pass I kept one. I wrote that I would add “when fanning out candidates, move one variable and pre-assign the slots by label” as a section of my own hook playbook. The stated benefit was “today only the chosen sentence survives, not what I changed to win, so after thirty-odd issues I still can’t count performance by hook type.”

That statement was false. Opening the publishing log, every issue already records its hook type as a label, marks first-time types with an (unused) tag 25 times over, runs an explicit single-variable discipline in lines like “format and length changed together — not a single variable”, and already aggregates performance by sentence form in a top-ten table. There is even a recorded rejected hypothesis: seeing reach fall as star counts rose, I had written “the strategy of using stars as social proof is rejected.” Worse, the experimental axis I was proposing was an axis already judged low-yield — two deliberately matched replications, aligned item by item on hook, format, length and topic, came out 18,000 against 423, a factor of 43.

The most expensive fact in this piece is not about the target but about my own assets — what I was trying to learn from someone else’s repository, I had already been doing for thirty-odd issues, and what that accumulation concluded was “changing the hook doesn’t help.” The same disease shows on both sides. They deleted their own numbers without a line of reasoning; I kept my records and then failed to read them, so I went shopping for someone else’s. Keeping a record and reading that record are different jobs.

Verdict

Item	Call
Install the skill	Rejected — its one runtime dependency (an image generator) is zero on this machine
Complexity budget	Rejected — check-weight.mjs is a strict superset (enforcing code plus stated derivation)
One-round question ceiling · absorbing context first	Rejected — already wired more strongly
One-pass unvetted hand-off	Cautionary example — every gate of mine is fail-closed
Purpose-hiding prompts	On hold — not a duplicate, but there is no surface to apply it to. My card pipeline is HTML rendering, so the prompt surface itself is zero
Candidate fan-out labelling	Adopted → rejected — my publishing log was already doing it, and that axis is already low-yield

Zero action items. What remains is only the reopen conditions. ① When an image-generation path exists on this machine. What I would measure then is the difference between “paste the prompt skeleton only” and “install the skill.” ② When my game work actually needs mascots or icons. Games sit outside this site’s figure policy, which releases one of the rejection grounds. ③ When upstream acquires a concept of versions. Conversely, a rising star count is not a reopen signal — stars ticked from 3,480 to 3,487 over this teardown while watchers stayed at 3, contributors at 1, and external PRs at 0.

Here is what I did not do. I never generated a single image. So this repository’s central claim — whether 4–7 shapes, three colours and 85–95% canvas coverage actually hold — stands at a sample size of zero. They didn’t measure it and neither did I; what this page established is “is the rule in the executable artifact”, not “does the rule land in the picture.” The deployed site renders client-side, so I could not machine-fetch the text of its terms of service, and the site source is split into a private repository and cannot be audited. The rights to a generated image are decided not by the repository’s MIT licence but by the terms of whichever image model you attached — and that fact is disclosed nowhere in the repository.