Skip to content
Back to search

agency-of-one

MarketplaceWorks with Claude CodeGitHub

Agency-grade, non-AI-looking web design for Claude Code.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today

Check it deeper on your machine

It has changed since its scan on 20 Nov 2026.

As a Claude Code plugin, install it and then ask Claude Code:

/plugin marketplace add keithmackay/mackayi/plugin install scanskillsafety@mackayiDo a deep scan of https://github.com/sam-pop/agency-of-one with scanskillsafety

Or from a terminal, clone it and run it:

git clone https://github.com/keithmackay/scanskillsafety-skill ~/.claude/skills/scanskillsafetynode ~/.claude/skills/scanskillsafety/cli.cjs --deep 'https://github.com/sam-pop/agency-of-one'

It checks the exact commit: every file, its packages against OSV (api.osv.dev), how it compares with this page's scan (findsafeskills.com receives the repository's address), and, if you agree, Claude reading every file with your own Claude subscription. It never runs the code.

Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/sam-pop--agency-of-one/badge.svg)](/listings/sam-pop--agency-of-one)

HTML

<a href="/listings/sam-pop--agency-of-one"><img src="/api/listings/sam-pop--agency-of-one/badge.svg" alt="findsafeskills safety rating" /></a>

Install

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

Tell your agent

Paste this into Claude Code:

Install the Agency of One design plugin for me:
1. Run `claude plugin marketplace add sam-pop/agency-of-one`
2. Run `claude plugin install agency-of-one@agency-of-one`
3. In this project, run `npm i -D playwright && npx playwright install chromium` so you can screenshot your design work
4. Run `claude plugin details agency-of-one` and show me the result
Then tell me to restart Claude Code so the plugin loads.

Or do it yourself

/plugin marketplace add sam-pop/agency-of-one
/plugin install agency-of-one@agency-of-one
npm i -D playwright && npx playwright install chromium   # per project, for screenshots

<details> <summary><b>No plugin system? Drop-in install</b> (copies into <code>.claude/</code>)</summary>

git clone https://github.com/sam-pop/agency-of-one
cd your-repo && ../agency-of-one/install.sh      # this repo only
../agency-of-one/install.sh --global              # every project (~/.claude)

This merges the lint hook into `.claude/settings.json` (never overwrites), adds the rules block to `CLAUDE.md` (idempotent), and gitignores `.design-review/`. Flags: `--target DIR`, `--no-hook`, `--no-claude-md`. </details>

Then just ask

Build the homepage for Northline Capital, a private lender making $1M–$25M bridge loans
on commercial real estate in the Mid-Atlantic. Term sheet in 48 hours. Get brokers to submit a deal.

The skill triggers on its own. **For the best result, hand Claude real photos, bios, client names and contact details up front.** It’s the single biggest lever (see [the evidence](#the-evidence)).

---

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

.claude-plugin/marketplace.json · Manifest · 15 lines
  1. {
  2. "name": "agency-of-one",
  3. "owner": {
  4. "name": "Sam P.",
  5. "url": "https://github.com/sam-pop"
  6. },
  7. "plugins": [
  8. {
  9. "name": "agency-of-one",
  10. "source": "./",
  11. "description": "Make Claude Code's websites and UIs stop looking AI-generated."
  12. }
  13. ],
  14. "description": "Agency-grade, non-AI-looking web design for Claude Code."
  15. }
README.md · README · 192 lines
  1. <p align="center"><img src="docs/assets/hero.png" alt="Agency of One: thousand-dollar agency results from a Claude subscription. Blind reviewers who'd say 'AI made this' fell from 65% to 37%." width="100%"></p>
  2. <p align="center">
  3. <a href="#install"><b>Install</b></a>&nbsp;&nbsp;|&nbsp;&nbsp;<a href="#the-evidence"><b>Evidence</b></a>&nbsp;&nbsp;|&nbsp;&nbsp;<a href="#how-it-works"><b>How it works</b></a>&nbsp;&nbsp;|&nbsp;&nbsp;<a href="#whats-inside"><b>What’s inside</b></a>&nbsp;&nbsp;|&nbsp;&nbsp;<a href="#faq"><b>FAQ</b></a>
  4. <br><br>
  5. <img alt="Claude Code plugin" src="https://img.shields.io/badge/Claude%20Code-plugin-0f3b33">
  6. <img alt="License: MIT" src="https://img.shields.io/badge/license-MIT-0f3b33">
  7. <img alt="Node 18+" src="https://img.shields.io/badge/node-18%2B-0f3b33">
  8. <img alt="Always-on cost: ~366 tokens" src="https://img.shields.io/badge/always--on%20cost-~366%20tokens-0f3b33">
  9. </p>
  10. **Ask Claude Code for a website and you get the same website**: centred hero, gradient headline, three icon cards. Ask it to "avoid AI slop" and you get the second AI look: cream background, serif headline, clay accent, ALL-CAPS labels, `→` on every link. **Agency of One** fixes the inputs instead of adding rules. It gives Claude a studio’s process, takes the riskiest choices away from it, and won’t call the work done until a blind reviewer signs off.
  11. ## Install
  12. ### Tell your agent
  13. Paste this into Claude Code:
  14. ```text
  15. Install the Agency of One design plugin for me:
  16. 1. Run `claude plugin marketplace add sam-pop/agency-of-one`
  17. 2. Run `claude plugin install agency-of-one@agency-of-one`
  18. 3. In this project, run `npm i -D playwright && npx playwright install chromium` so you can screenshot your design work
  19. 4. Run `claude plugin details agency-of-one` and show me the result
  20. Then tell me to restart Claude Code so the plugin loads.
  21. ```
  22. ### Or do it yourself
  23. ```text
  24. /plugin marketplace add sam-pop/agency-of-one
  25. /plugin install agency-of-one@agency-of-one
  26. ```
  27. ```bash
  28. npm i -D playwright && npx playwright install chromium # per project, for screenshots
  29. ```
  30. <details>
  31. <summary><b>No plugin system? Drop-in install</b> (copies into <code>.claude/</code>)</summary>
  32. ```bash
  33. git clone https://github.com/sam-pop/agency-of-one
  34. cd your-repo && ../agency-of-one/install.sh # this repo only
  35. ../agency-of-one/install.sh --global # every project (~/.claude)
  36. ```
  37. This merges the lint hook into `.claude/settings.json` (never overwrites), adds the rules block to `CLAUDE.md` (idempotent), and gitignores `.design-review/`. Flags: `--target DIR`, `--no-hook`, `--no-claude-md`.
  38. </details>
  39. ### Then just ask
  40. ```text
  41. Build the homepage for Northline Capital, a private lender making $1M–$25M bridge loans
  42. on commercial real estate in the Mid-Atlantic. Term sheet in 48 hours. Get brokers to submit a deal.
  43. ```
  44. The skill triggers on its own. **For the best result, hand Claude real photos, bios, client names and contact details up front.** It’s the single biggest lever (see [the evidence](#the-evidence)).
  45. ---
  46. ## The evidence
  47. We ran the kit end to end on real briefs with fresh agents, then had independent reviewers judge the screenshots **blind**: they weren’t told which pages used the kit. Same model, same one-paragraph briefs.
  48. <p align="center"><img src="docs/assets/results.png" alt="Bar chart. Neutral reviewer: 'AI made this' 65% without the kit vs 37% with it; skeptic reviewer 88% vs 78%; 'a top agency made this' 25% vs 55%." width="100%"></p>
  49. | | “AI made this” (neutral reviewer) | “AI made this” (skeptic hunting for AI) | “A top agency made this” |
  50. |---|---:|---:|---:|
  51. | Without the kit (control) | 65% | 88% | 25% |
  52. | With Agency of One (mean of 3 sites) | **37%** (35–40) | **78%** (75–80) | **55%** (50–60) |
  53. <p align="center"><img src="docs/assets/craft.png" alt="Dumbbell chart of craft scores 0–10, control vs kit: concept 6→7.5, typography 6.5→7, layout 5.5→7.5, colour 5.5→7.5, imagery 5.5→7.5, motion 3→6, responsive 6.5→7, copy 8→8.5, technical 7→7.5, credibility 6.5→7.5." width="100%"></p>
  54. ### Same brief, before and after
  55. <table>
  56. <tr>
  57. <td width="50%" valign="top"><img src="examples/00-control-no-kit.png" alt="Control: Claude Code without the kit"><br><sub><b>Without the kit.</b> One reviewer: “a B2B consultancy template in the Webflow mould.” Another, blind: “the one a design-literate viewer calls AI fastest.”</sub></td>
  58. <td width="50%" valign="top"><img src="examples/clausewell-fold.png" alt="Clausewell with the kit"><br><sub><b>With the kit: Clausewell</b> (AI contract review). Ranked #1 by both blind panels. The hero is the product doing its job: a real, interactive redline.</sub></td>
  59. </tr>
  60. <tr>
  61. <td valign="top"><img src="examples/northline-fold.png" alt="Northline Capital with the kit"><br><sub><b>Northline Capital</b> (commercial lender). One true figure and a day-by-day funding clock. Reviewer: “a stranger gets the offer in five seconds.”</sub></td>
  62. <td valign="top"><img src="examples/keelson-fold.png" alt="Keelson Advisory with the kit"><br><sub><b>Keelson Advisory</b> (fractional CTO). A persistent column turns the brand idea into structure. Reviewer: “holds its own on voice, colour discipline and typesetting.”</sub></td>
  63. </tr>
  64. </table>
  65. All three were unattended runs from a one-paragraph brief, with no photos, no client proof and no human edits.
  66. ### What the reviewers still said, and what changed because of it
  67. | Finding | What the kit does now |
  68. |---|---|
  69. | “No people, no photography, no proof”: credibility and imagery stall around 6/10 | Asks for real photos, bios, clients and contact details **before** designing; slots them professionally when missing, never fakes them |
  70. | “These share a hand”: an honest-assistant voice, code-drawn imagery only, tinted off-white grounds | Catalogued as a 4th-generation tell (G4), with 6 new lint rules and revised voice and palette guidance |
  71. | “Good in-house or mid-tier studio work” | The honest ceiling for code-only output today. Real photography and proof are what lift it further |
  72. <details>
  73. <summary><b>Method, in full</b></summary>
  74. - **Builds:** fresh Claude agents with no conversation context ran the plugin unattended on three briefs (fractional CTO practice, AI contract-review SaaS, commercial real-estate lender) and did one fix round from the kit’s own reviewer. The control was the same model given the CTO brief with “make it look great” and no kit.
  75. - **Panels:** two independent reviewer agents saw five pages under shuffled letters (three kit builds, the control, and an earlier expressive-mode build). One was a neutral “Fortune-50 studio creative director”; the other was told to hunt for AI-generated sites. Neither knew a kit existed.
  76. - **Scores:** “AI made this” and “a top agency made this” are 0–100 likelihoods; craft axes are 0–10, where 8 means it could sit next to the best sites in its category.
  77. - **Caveats:** small sample (one control, three kit sites, two reviewers); reviewers are models, not people. Treat the numbers as directional evidence, not a benchmark. The full transcripts informed every rule in `references/tells.md`.
  78. </details>
  79. ---
  80. ## How it works
  81. <p align="center"><img src="docs/assets/pipeline.png" alt="Three stages. Decide: strategy and proof, random levers, five directions then an independent critic, DESIGN.md. Build: hero set piece first, page from tokens, craft finish pass. Prove: render and measure, plan vs render diff, blind two-axis review. A 57-rule linter runs on every edit; a design log blocks repeats across projects." width="100%"></p>
  82. Models collapse to the most likely design. Banning that look only moves them to the *next* most likely one, which is how the cream-and-serif look became the second AI tell. So the kit doesn’t ask Claude to be creative. It changes what Claude designs **from**:
  83. | Mechanism | What it does | Why it works |
  84. |---|---|---|
  85. | **Strategy first** | Positioning, the one belief, 5 competitor “category codes” marked keep or break, real proof inventory | Distinction comes from strategy, not decoration |
  86. | **Random levers** | A script deals a composition system, brand device, palette strategy, type strategy, quality reference, and a **brand hue sampled in OKLCH** outside the default zones (corporate blue, indigo, violet, clay). **~100 million combinations** | Randomness from outside the model breaks mode collapse |
  87. | **Two baselines** | Claude writes the *AI twin* and the *agency-cliché twin* of the brief, then must differ from both on every axis | Name the default to avoid it |
  88. | **Independent critic** | Five directions go to a separate agent that scores typicality and strength; a script picks at random among the atypical **and** strong ones | Verbalized sampling widens range 1.6–2.1×; a separate scorer can’t be gamed; the strength gate kills “unusual but drab” |
  89. | **Hero first, then finish** | The hero is built and polished in isolation, then the page, then a line-by-line craft pass | Where agency work earns its first impression |
  90. | **Lint on every edit** | 57 detectors for 4 generations of AI tells and agency clichés; findings go straight back to Claude | Executable gates beat adjectives |
  91. | **Measure the render** | Screenshots at 1440/390 plus states and scroll positions; which fonts *actually* loaded; colour and size distribution | Agents leave >25% of stated design decisions unimplemented |
  92. | **Blind two-axis review** | Distinctiveness plus 10 craft axes against named category benchmarks; SHIP needs every axis ≥ 8; max 3 rounds, keep the best | External critics beat self-critique; unbounded self-revision degrades |
  93. | **Memory** | A design log blocks reusing a display face, or repeating a past project on more than 3 of 6 axes | Stops Claude converging on its own house style |
  94. Two modes: **agency** (default; business-forward and polished) and **expressive** (`--mode expressive`; non-web reference media and constraints for campaigns, editorial and experimental work).
  95. ---
  96. ## What’s inside
  97. | Component | What it is | Token cost* |
  98. |---|---|---|
  99. | `agency-of-one` skill | The 4-phase process plus 5 reference files | ~130 always-on, ~2.3k when it runs |
  100. | `design-reviewer` agent | Blind ECD: two-axis scoring, fixed verdict rules | ~100 always-on, ~1.1k per review |
  101. | `direction-critic` agent | Scores directions for typicality and strength | ~100 always-on, ~490 per run |
  102. | `/agency-of-one:design-review` | Lint, screenshots, plan-diff and blind review in one command | ~30 always-on, ~370 per run |
  103. | SessionStart hook | Loads the core rules, **only in frontend repos** | ~680 in those repos, 0 elsewhere |
  104. | PostToolUse hook | Runs the linter on every edited UI file | none (harness-side) |
  105. <sub>*From `claude plugin details`: about **366 tokens always-on**. A full run (five directions, critic, build, up to three reviews) is token-heavy; budget for it like a real design engagement.</sub>
  106. **Scripts** (zero runtime dependencies besides Playwright; Node 18+; run from your repo root, `$S` = the skill folder):
  107. | Command | Does |
  108. |---|---|
  109. | `node $S/scripts/draw.mjs --project "Name" [--mode expressive]` | Deal the levers and the brand hue |
  110. | `node $S/scripts/draw.mjs --pick "A:0.3:8,B:…"` | Pick among atypical and strong directions (scores from the critic) |
  111. | `node $S/scripts/type-specimen.mjs --fonts "A, B, C" --text "…"` | Render font candidates with your real copy, then choose by looking |
  112. | `node $S/scripts/slop-lint.mjs [--json]` | Detect tells (exit 1 on errors) |
  113. | `node $S/scripts/screenshot.mjs --url URL [--widths 1024,768] [--state name=click:#sel]` | Render, capture, measure |
  114. | `node $S/scripts/plan-diff.mjs` | DESIGN.md tokens vs what actually rendered |
  115. | `node $S/scripts/designlog.mjs check\|add\|list` | Cross-project memory |
  116. <details>
  117. <summary><b>What the linter catches</b> (57 rules, 4 generations)</summary>
  118. | Generation | Examples |
  119. |---|---|
  120. | **G1**, the original AI look | Inter/Roboto/system fonts, purple and indigo, gradient text, glows and glass, blurred blobs, three icon cards, “Trusted by” logo strips, emoji headings, ✨, buzzword copy, invented stats, generic CTAs | <!-- slop-ok -->
  121. | **G2**, “tasteful AI” | Fraunces/Instrument Serif/Space Grotesk/Geist and friends, cream `#F4F1EA` + clay `#D97757`, near-black + acid accent, ALL-CAPS eyebrows, `A · B · C` meta, `→` on links, one italic word in a headline, 01/02/03, untouched shadcn theme | <!-- slop-ok -->
  122. | **G3**, the anti-slop house style | The page built as a paper artefact, hairline cells, lettered keys, mono annotations, neobrutalist offset shadows, marquee strips | <!-- slop-ok -->
  123. | **G4**, the honest-assistant fingerprint | “Nobody else.” / “What X is not” voice, narrated demos, “Question N of 5”, the same CTA five times, straight apostrophes |
  124. Full catalog with signatures and sources: [`references/tells.md`](skills/agency-of-one/references/tells.md).
  125. </details>
  126. ### Configure
  127. - **Your brand wins.** If the brief or a brand guide pins a look, Claude follows it and records the exceptions.
  128. - **Existing `DESIGN.md`** at the repo root is followed exactly, so commit it.
  129. - **Deliberate exceptions** go in `DESIGN.md §9` and `slop-lint.config.json`:
  130. ```json
  131. { "allowFonts": ["Fraunces"], "allowRules": ["eyebrow"], "ignore": ["legacy/**"] }
  132. ```
  133. Or add an inline `slop-ok: rule-id` comment.
  134. ---
  135. ## FAQ
  136. **Does it work on a $20 Pro plan?** It installs and runs, but a full run (five directions, critic, build, up to three blind reviews) is token-heavy, so expect to hit usage limits on Pro. It’s comfortable on Max. The always-on cost is ~366 tokens.
  137. **Will every site I build look the same?** That’s what the design log is for. It blocks reusing a display face from your last six projects or repeating a past project’s fingerprint, and the levers and brand hue are dealt with real randomness.
  138. **Next.js / Tailwind / shadcn?** Yes. Tokens go in CSS variables or Tailwind v4 `@theme`, and shadcn/Radix are used for behaviour only, with every primitive restyled.
  139. **Can I use it on an existing site?** Yes. Run `/agency-of-one:design-review` for a blind review and lint report on what you have, or ask Claude to write a `DESIGN.md` for it first.
  140. **Why does it ask for photos and client names?** Because the blind reviewers scored credibility and imagery around 6/10 without them, whatever the design quality. Code can’t produce proof.
  141. **Is a lint pass enough?** No. Lint is the floor. The critic, the craft pass and the blind review are the real gates.
  142. ---
  143. ## Contributing
  144. Spotted a new default creeping into AI-built sites? Add it to `references/tells.md` with its signature and generation, add a detector to `scripts/slop-lint.mjs`, add a case to `tests/fixtures/slop`, and check that `tests/fixtures/clean` stays clean. `npm test` runs the self-tests, and CI runs them on every push. **Never add a recommended-fonts list:** whatever gets recommended becomes the next tell.
  145. ## Research basis
  146. Anthropic’s frontend-design skill (June 2026 rewrite); the Hallmark, Impeccable, pixelslop, taste-skill and Refero anti-slop projects; Verbalized Sampling (arXiv 2510.01171); Vibe Design Agents (Google, 2026); ReLook (arXiv 2510.11498); Design Creativity Bench (arXiv 2610.11539); “Design Theater” (arXiv 2607.22928); the DesignArena 64k-page study. Sources in [`references/tells.md`](skills/agency-of-one/references/tells.md).
  147. ## License
  148. MIT © Sam P.

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.