Two free skills for your coding agent: one finds a name and checks whether it's taken, the other builds three brand directions and a kit for the one you pick, and both tell you what they couldn't check.
npx skills add juztripper/already-looked
It named and branded itself, in headless runs where the prompt made the final selection automatic. That shows the skills run end to end. It doesn't show the name is good, or clear of conflicts.
Same model, same briefs. Claude Sonnet told to "suggest 6 names", then the naming skill.
| Fictional brief | Asked plainly | With the skill |
|---|
| Attack-surface scanner | Plainsight, Surfacewalk, Ledgerline, Tallyho | Tipoff: attackers don't hack, they notice. This is the quiet word you get before it matters.Orphan: engineers already say orphaned accounts for assets whose owner has left.Tenuki: in Go, ignoring the local fight to play elsewhere, as attackers do. |
| Eczema skincare in Germany, the Netherlands and the UK | calma, nulla, tolera | lau: the first thing every patient is told is lukewarm water, and lau is German for lukewarm.still: means quiet in all three markets.fine: the answer to "how's your skin?" |
| Small-firm legal software | Lantern, Harbor, Firmly, Ledgerline | Tenth: lawyers bill in tenths of an hour.Line One: "Mrs. Alvarez on line one."Shingle: a lawyer hangs out a shingle to open a practice. |
| Repairable cycling kit made in Portugal | Mend, Estrada, Darn | Trinta: riders wash kit at 30 degrees, and trinta is thirty in Portuguese.Devagar: Portuguese for "slowly", from "devagar se vai ao longe".Bartack: the stitch that reinforces where bib straps fail. |
| Board-game publisher | Marginalia Games, Lanternworks, Open Table Press | Colophon: the credit page of a book.Fair Copy: the clean final version made from a messy draft. |
Plain output came with one-line reasons too; they're in the results. It also reached Colophon once, as Colophon Games. On Codex with gpt-6.1-terra the skill did worse. Numbers below.
Real skill output, unedited boards, cropped, on fictional briefs. These are drafts and they aren't cleared. Names are illustrations, not recommendations, and may coincide with real companies. People, figures and test results on the boards are invented. Codex boards remain weaker overall. More boards, with the risk the skill wrote down, are on the landing page.
A strong prompt still beats the branding skill
Our first gallery was a stock font plus a small decoration on a plain tile. The owner looked at it and said it didn't look good. They were right, so we rebuilt the skill to make a whole visual world. Then we tested that against a strong prompt asked for the same world, on the same top-tier budget, on three briefs. This is the fourth round on them.
- Strong prompt on Claude Opus: 4.67, about $3 a run. On Codex: 3.40.
- The skill's recommended direction on Claude: 4.27, about $14 a run. On Codex: 2.50.
- Best of the skill's three directions beat the prompt on 1 of 3 briefs on Claude, +0.23 on average, and 0 of 3 on Codex, -0.90.
The prompt still wins on the recommended direction, on both agents. The skill is close on Claude, and clearly weaker on Codex. Claude's +0.23 is carried by one brief: the same strong-prompt lau board scored 4.50 in an earlier panel and 3.00 in this one, because panel scores are relative to what's beside them. Without lau, Claude's mean is -0.55.
On Codex: use the naming skill, with codex -m gpt-6.1-sol. For one deep identity, use the one prompt: it scored 3.40 against 2.50, and no direction beat it. Use the branding skill there only if you want the three directions, the checks and the kit.
Re-judging moves scores. The Claude boards are the identical files in the last two rounds, and best of three on Claude went from +0.47 to +0.23. The same Upstroke prompt board went from 4.50 to 4.90. Treat differences under about 0.4 as noise.
If you want one deep identity, use this prompt.
Why we think it still wins: a top model left free plays its best idea. Direction 1, the free one, was the panel's best in 4 of 6 runs. The prompt's free runs also converge: 3 of 6 picked red-orange, against 1 of 6 of the skill's recommended directions. This may partly be judge bias: LLM judges share the models' taste, so this panel can't separate quality from typicality. No human designer or client has judged any of this.
What we changed: direction 1 is free, directions 2 and 3 are pushed away from defaults, each at full depth. The recommendation chose direction 1 in 2 of 6 runs and matched the panel's best in 0 of 6, so we added one sentence telling the skill not to pass over it. The Claude boards predate that sentence. The Codex rerun read it, on one run per brief, so nothing here shows it works.
We also found our harness told runs not to open "any viewer", and every Codex direction said it hadn't looked at its board. We fixed the line and reran both Codex arms. None says that now, though the tool call isn't logged, so we can't prove each one looked. The gap barely moved, -1.04 to -0.90, which is within re-judging noise. So that wasn't the main cause, and we don't know the cause. Our read of the boards, not a measurement: Codex's are tidy, modular and low on conviction.
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
