A Claude Code /logo skill built from five wrong turns — SVG paths, regex bounds, a wrong kanji
Redesigning this blog's logo and code font with an AI agent: a broken SVG letterform, a regex bounding box, a wrong kanji, and the /logo skill it produced.

On this page
Introduction
I sat down to fix a logo I had not been happy with for a long time: a plain text mark this project generated for itself on day one, and that nobody had made a decision about since.
Designing a logo with an AI agent takes rounds: the agent generates candidates, I look, I cut and say why, and we go again. What I wanted to know was whether that loop was worth writing down as a Claude Code skill.
The agent was
confidently wrong five separate times across those rounds, and I count none of
it as wasted. Four of the five wrong turns named a rule, and the rules are now
a /logo skill and three scripts that the next mark starts from.
What shipped is a vermilion seal bearing 桜 next to “oharu” set in Fraunces, plus JetBrains Mono on article pages only.
The blog had no design, and I had been living with it
Here is what the favicon was:
rect and one <text> element asking for
Helvetica,Arial,sans-serif, at the three sizes a favicon is used at. The
letter is whatever the visitor’s operating system supplied.The header had no mark at all, just the site title as text.
The typography was the same story:
--font-sans: ui-sans-serif, system-ui, -apple-system, 'Segoe UI', Roboto, …;--font-mono: ui-monospace, 'SF Mono', SFMono-Regular, Menlo, Consolas, …;Pure system stacks: there was no font. Every visitor saw whatever fonts their operating system shipped. That is a legitimate trade (zero bytes, zero layout shift), but it was not a decision anyone had made.
Wrong turn 1: twenty-four variations of one idea
I asked for breadth first: twenty-four concepts across icon-only, wordmark, and combination marks, rendered into a gallery showing each at 128px, 32px and 16px at once.
It produced twenty-four of them. Every one was a logo built around the letter o.

That is the icon-only third of the sheet. Aperture, orbit, terminal, arcs, torii, squircle, prompt, split. Eight names for one drawing.
Every one is a monoline geometric stroke: circles, arcs, even weight, rounded caps. Icon or wordmark, they share a skeleton, and that skeleton is the common default for generated logos. I had asked for range and received repetition wearing eight hats.
The fix was to look at real work first. Six identities, screenshotted and actually examined:
| Reference | What it does | What round one had done |
|---|---|---|
| Linear, Astro | Solid fills, mark small and quiet | All strokes, which close up at 16px |
| Rauno Freiberg, Josh Comeau | No mark at all. Type plus one committed colour | Assumed a mark was mandatory |
| Maggie Appleton | Display serif. It is a writing site | Geometric sans, the template default |
| Anthony Fu | Personal monogram, hand character | Pure geometry cannot produce “personal” |
Round two came back genuinely different, because this time the constraints came from real sites:

An AI asked for N options returns N samples from one distribution.
Wrong turn 2: letterforms that were never rendered
Hand-drawn SVG letterforms carry bugs that only rendering shows, and round one
went into a gallery without being rendered once. The worst was the letter u:
M9,49 L9,99 A41,41 0 0 0 91,99 L91,140Read it as a pen. Left stem down from the x-height to y=99. Bowl arc across to
(91,99). Then a line down to the baseline. The arc ends at y=99, so the
right stem only spans 99 to 140: it is missing its entire top half. The
letter renders as a u with a stunted right side.
The fix draws the right stem as its own subpath, at full x-height:
M9,49 L9,99 A41,41 0 0 0 91,99M91,49 L91,140
A bug like this is obvious in a browser and invisible in the path data.
Wrong turn 3: a character the agent assumed
The strongest concept to come out of round two was a hanko, a personal seal, which is close to what a blog actually is. Vermilion, one character, knocked out in white.
The name is “oharu”, so the agent first built the seal on 春 (haru, spring). It looked good, but it was wrong. I corrected it: “oharu” is written 桜明, so the kanji for haru should have been 明. In the end I chose 桜 for the seal, because it has more character.
桜明 being two characters changed the design, not just a glyph: two characters cannot survive a 16px favicon. At that size a 10-stroke kanji turns to texture.

桜明 stacked vertically is the correct hanko for the name and the hardest to read in a browser tab. The outlined version keeps its frame, but 桜明 shrinks until it nearly disappears. So I made the seal round, with the single character 桜.
Wrong turn 4: a bounding box computed with a regex
For the shipped assets, the hand-drawn letters gave way to outlines baked from Fraunces (wordmark) and Noto Serif JP (桜), using fontTools and HarfBuzz. Correct kerning and contrast, frozen into path data, so the wordmark ships as geometry with no webfont request and no flash of fallback type.
Placing a glyph needs its inked bounds. SVG path data is a list of pen moves: a
letter for the move, then the coordinates to go to. L720,50 draws a line to
(720,50) and carries both x and y. outline.py pulled every number out of the
path string and treated them as alternating x, y pairs.
But letters are full of perfectly horizontal and vertical edges, and
SVGPathPen writes those as H and V shorthand. H700 draws straight across
to x=700 with y unchanged, so it carries a single coordinate. One
horizontal line flips the parity of every number after it:
H700 carries only an x, so every pair the regex reads after it is shifted by one number.It survived galleries looking merely slightly off. It only became undeniable when a lockup that should have been roughly 3200 units wide came out at 1802.
The fix is to stop parsing text and let the library measure:
from fontTools.pens.boundsPen import BoundsPen
bounds = BoundsPen(glyph_set)for info, pos in zip(buf.glyph_infos, buf.glyph_positions): xform = (scale, 0, 0, -scale, (x + pos.x_offset) * scale, 0) glyph_set[name].draw(TransformPen(pen, xform)) glyph_set[name].draw(TransformPen(bounds, xform)) # same transformMeasured by fontTools, the wordmark’s real bounds are 2456.4 × 758.6. The
regex had been returning different numbers all along, and it took a person
checking them by eye to notice.
Wrong turn 5: a preload on every page
Code blocks also got their own typeface. They are the surface a tech blog is actually read on, and mine were rendering in Menlo, the same “no decision was made” state as the body text.
JetBrains Mono Variable, Latin subset only, 40KB, self-hosted at a fixed path so it can be preloaded by URL. No CJK font is fetched for code: code is Latin, so the CJK stacks stay untouched.
The preload went into the shared layout. The browser said why that was wrong:
The resource http://localhost:4321/fonts/jetbrains-mono-latin.woff2 waspreloaded using link preload but not used within a few seconds from thewindow's load event. Please make sure it has an appropriate `as` value andit is preloaded intentionally.The home page, the tag listings and both CJK index pages have no code on them.
Each one downloaded 40KB and used none of it. Gating the preload behind a
preloadMono prop that only ArticleLayout sets:
| Page | rel="preload" links |
|---|---|
dist/index.html | 0 |
dist/ja/index.html | 0 |
dist/tags/index.html | 0 |
dist/blog/<article>/index.html | 1 |
What the rounds became: a /logo skill
The skill’s principles:
- Study real sites before generating anything.
- Render every concept and check it at 16px before showing it to anyone.
- Never assume a fact about the person, such as which kanji their name uses; ask.
- Set type from real typeface outlines, never hand-drawn letterforms, and let fontTools measure them.
- Decide colour by putting options side by side, not by theory.
- The user picks at every gate.
The skill is told to present, recommend with reasons, and stop: an opinion is useful, a decision is not yours. Every round in this article ended at a gate where I chose.
Three scripts do the mechanical half:
gallery.mjsbuilds the contact sheet and serves it on a local port.outline.pybakes typeface outlines into path data through fontTools and HarfBuzz.build-brand-lockup.mjsis the worked example that emits the favicon and the header component.
Concepts and sheets are written to a scratch directory, never into the repo: a design round is mostly rejects, and a repository is not where rejects go.
What shipped
A vermilion round seal bearing 桜, next to “oharu” set in Fraunces. I shipped it as two SVGs: the full lockup, seal plus wordmark, and the seal alone for narrow screens. The narrow version cannot be cropped out of the full one: the seal is deliberately shorter than the lettering, so a square crop takes in the letters beside it.
Below 50rem the seal appears alone, which also solved an old layout problem: on a phone, the text wordmark competed with the search control for space.
One colour decision came from putting the options side by side. The site accent
is #1d4ed8; the seal is #d8452a. Those sit almost opposite each other, and a
blue wordmark beside a red seal visibly fights. The wordmark takes the theme’s
ink and blue stays on links, so the two never sit adjacent.
| Wordmark path data | 3569 bytes (integer-rounded from 5145) |
| Seal path data | 1677 bytes |
favicon.svg | 1878 bytes |
| JetBrains Mono, Latin subset | 40KB, article pages only |
Webfont requests on a /ja/ article | 1 (JetBrains Mono), no CJK font |
Summary
Five wrong turns. Not one of them was caught by reasoning about the code.
Good design came only from comparing against real sites. The broken letterforms needed a browser. The wrong kanji needed me to correct it. The bounding box bug needed a number absurd enough to notice. The wasted preload needed a console.
These five are the argument for running this as a loop rather than as a prompt. Producing twenty-four polished concepts costs minutes now, which makes it easy to mistake volume for range and to accept a confident artefact built on a guess.
What a loop adds over a prompt is rendering, measuring, and asking when the design depends on a fact only I have.
None of the five felt like waste at the time, and none reads like waste now. The next mark starts at six rules and three scripts instead of at a blank prompt.



