Home · Marketing & Content · brand-guidelines
brand-guidelines
Anthropic's official brand card: seven exact colours and two font stacks for styling artifacts with Anthropic's look-and-feel.
Score
“Seventy-three lines of exact, copy-paste brand data that measurably beat a strong baseline on fidelity — 7 of 7 colours exact, both fonts right. But it is a data card, not a workflow: no steps, no checks, a trigger that can fire on other companies' brands, and feature text describing a PowerPoint script that does not ship.”
The baseline knew the palette cold — 4 of its 9 colours were exact from memory. The skill still won on discipline: 7 of 7 colours exact, zero invented, and both fonts right where the baseline got both wrong.
Sub-scores & evidence
| Task completion | Invitation restyled end to end, every element on-palette. One point off: the date line set in Mid Gray #b0aea5 on #faf9f5 is barely readable — a defect produced by following the skill's own 'secondary elements' prescription. | 11/12 |
| Beats the no-skill baseline | With skill: 7/7 colours exact, 0 near misses, 0 invented, fonts 2/2. Baseline: 4/9 colours exact, 3 near misses, 2 invented shades, fonts 0/2. A clear fidelity win, though the baseline's page still looked the part. | 7/10 |
| Self-verification in output | SKILL.md defines no check step of any kind. The output was right because the values were copied, not because anything verified them. | 1/5 |
| Failure behaviour | No failure case injected this round. The one stated degradation path — Arial/Georgia when Poppins/Lora are unavailable — was encoded in the output's font stacks. | 2/3 |
There is no workflow. SKILL.md is a reference card — values plus feature claims, no ordered steps, no checks, no definition of done.
Sub-scores & evidence
| Clear steps | No procedure at all. Sections list colours, fonts, and features described in the present tense ('Applies Poppins font to headings'), but nothing tells the agent what to do first, next, or last. | 2/6 |
| Inputs & outputs defined | Input is any artifact, unstated. Output expectations exist only as python-pptx hints ('Applied via python-pptx's RGBColor class'); for HTML or CSS the agent is on its own. | 2/4 |
| Checks & verification | Nothing. No contrast check, no palette-compliance check, no review pass — in a skill whose prescribed mid gray fails on its prescribed light background. | 0/4 |
| Stop conditions | Nothing irreversible is in scope, but the skill describes itself as post-processing existing artifacts and says nothing about preserving originals or asking first. | 2/3 |
| Failure recovery | Font fallback to Arial/Georgia is specified — the most likely failure is covered. Every other failure mode is unaddressed. | 2/3 |
73 lines and roughly 560 tokens — the smallest package we have tested. One flat file is fine at this size; the cost is a Features section describing machinery that is not in the box.
Sub-scores & evidence
| Leanness | SKILL.md is 73 lines / 2,235 bytes — about 559 tokens in the always-loaded layer. | 5/5 |
| Progressive disclosure | No reference files; everything sits in one file. Acceptable only because the total is tiny. | 3/5 |
| Organisation & naming | Overview, Colors, Typography, Features, Technical Details — findable in seconds; the name field matches the folder name. | 3/3 |
| No padding | Roughly a third of the file (Features, Technical Details) documents a python-pptx post-processor that does not ship — no script, no code, nothing to run. | 1/2 |
Fires reliably for Anthropic styling — and just as easily for the wrong brand. The description claims 'company design standards' in general; the content covers exactly one company.
Sub-scores & evidence
| What it does | 'Applies Anthropic's official brand colors and typography to any sort of artifact' — the first sentence is unambiguous. | 4/4 |
| When to use | An explicit 'Use it when…' sentence exists, but its conditions are circular: fire when 'brand colors or style guidelines… apply' is nearly any design task. | 3/4 |
| User-word keywords | Ships a keyword list (branding, corporate identity, visual identity, styling, brand colors, typography) that mirrors how people actually phrase these requests. | 3/3 |
| When NOT to use | No boundary at all — and the generic 'company design standards' phrasing actively invites firing on other companies' brand work, where the skill would impose Anthropic's identity. | 0/3 |
| Naming & voice | Short functional name; description written as an instruction to the agent. The name is generic for single-company content, but the description names Anthropic. | 1/1 |
Two text files, pure ASCII. Nothing to execute, nothing to exfiltrate, no URLs anywhere in the package.
Sub-scores & evidence
| Least privilege | Requests no tools, files or network access. Byte scan of SKILL.md: zero non-ASCII characters, zero URLs, no commands or download instructions (the only 'install' mentions advise that fonts be pre-installed). | 4/4 |
| Declared tool scope | No allowed-tools field declared. | 2/3 |
| Sensitive-action confirmation | No sensitive operations exist in scope; nothing stated either way. | 2/3 |
The data is exact and copy-paste ready — seven hex values, two font stacks, one size rule. Around it: no worked example, and feature copy for a PowerPoint post-processor that does not ship.
Sub-scores & evidence
| Examples & templates | No styled example artifact, no template, no before/after. The colour and font lists are the entire payload. | 1/3 |
| Scripts & instruction precision | Ships no scripts. Where instructions exist they are precise (exact hexes; headings are '24pt and larger'), but the claimed features — smart font application, accent cycling, python-pptx RGBColor — have no implementation to be precise with. | 2/3 |
| Dependencies stated | Fonts 'should be pre-installed in your environment' is stated twice. python-pptx is named as the application mechanism but never declared as a requirement — and no code uses it. | 1/2 |
| Concrete instructions | Every value is copy-paste ready with a role attached: '#141413 — Primary text and dark backgrounds', '#b0aea5 — Secondary elements'. Our with-skill run applied them verbatim. | 2/2 |
Best for / not for
Best for
- Artifacts that must carry Anthropic's identity — partner decks, community event pages, integration docs
- A template to fork for your own brand: the structure is 73 lines and the licence is Apache-2.0
- Styling work where exact hex values matter more than design direction
Not for
- Your own company's brand — despite the generic name and description, the only identity inside is Anthropic's
- Anyone expecting the described post-processing features to be automated — no code ships with this skill
- Accessibility-sensitive output — the prescribed mid gray on the prescribed light background fails contrast for small text, and the skill never mentions contrast
What it does
brand-guidelines is a data card. The whole package is two files — SKILL.md (73 lines, 2,235 bytes) and an Apache-2.0 LICENSE.txt — containing seven hex values with roles (Dark #141413 for text and dark backgrounds, Light #faf9f5, Mid Gray #b0aea5, Light Gray #e8e6dc, and orange/blue/green accents), two font stacks (Poppins for headings, Lora for body, Arial/Georgia fallbacks), and one size rule (headings at 24pt and larger get Poppins). There are no scripts, no assets and no reference files.
Around that core, roughly a third of the file describes features in the present tense — “Smart Font Application”, “Cycles through orange, blue, and green accents”, colour application “via python-pptx’s RGBColor class”. No code ships. The text reads like documentation for a PowerPoint post-processor that was never included, and an agent styling HTML, CSS or anything else has to translate the values itself. SKILL.md was last modified on 2025-12-01; the folder’s only later touch is a one-line LICENSE.txt copyright fix on 2026-04-20.
How the live test went
We restyled the same plain HTML invitation twice: once following the skill, once from a capable agent’s own knowledge of the brand. Then we stopped judging by eye — every hex value in both files was extracted and measured against the seven values printed in SKILL.md.
The baseline was better than we expected. From memory it produced #141413, #faf9f5, #e8e6dc and the signature clay orange #d97757 exactly — 4 of its 9 colours were official values, and two more were warm whites within 13 RGB units of brand Light. But it drifted: an invented brown-black for body text, an invented darker clay for hover states, a muted gray with no palette equivalent. And it got both fonts wrong, pairing a serif display face with Inter where the skill prescribes Poppins and Lora. (Its pairing tracks the current anthropic.com look more closely than the skill’s own spec does — the skill’s font choices may simply be older than the brand’s present site. We scored against the skill as written, because that is the artifact under review.)
The with-skill output used all seven official colours and nothing else: 7 of 7 exact, both fonts as specified, the three accents cycled in a bar across the top as the shape-colour rule describes. Its one visible defect was also the skill’s — the date line, set in Mid Gray because the skill assigns mid gray to secondary elements, sits at roughly 2:1 contrast on the Light background. The skill prescribes the pairing and never mentions contrast.
Where it falls short
The trigger boundary is the structural problem. The name is generic and the description claims “company design standards” in general, but the content covers exactly one company. Ask an agent to apply your company’s standards with this skill installed and there is a real chance your document comes back wearing Anthropic’s identity — the description never says when not to fire, and our desk test flags two of five ordinary prompts as misfire risks.
The second shortfall is the gap between the feature copy and the box. A skill that says it applies fonts via python-pptx and cycles accent colours across shapes is describing automation; what you download is a list. The list is accurate — our test applied it verbatim and it beat a strong baseline — but at 66 points this is a well-made reference card with no workflow, no verification step and no worked example, one brand wide.
Live A/B test — the evidence
One brief, run twice: restyle a plain, unbranded HTML event invitation (title, date line, two paragraphs, RSVP button; system fonts, greys) to match Anthropic's brand identity. Ground truth is the skill itself: the seven hex values and two font stacks printed in SKILL.md. Every colour in each output was extracted and measured against that palette — exact match, near miss (within 40 units of RGB distance), or invented.
With the skill
#141413
#faf9f5
#b0aea5
#e8e6dc
#d97757
#6a9bcc
#788c5d
Fonts: Poppins (headings) + Lora (body), with the skill's Arial/Georgia fallbacks
- 7 of 7 colours used are exact brand values — zero near misses, zero invented
- Both fonts exactly as specified: Poppins headings, Lora body
- Accent bar cycles all three accents (orange, blue, green), as the skill's shape-colour rule describes
- One defect came from the skill itself: the date in Mid Gray #b0aea5 on #faf9f5 is barely readable, and the skill offers no contrast check
Without the skill (baseline)
#141413
#faf9f5
#d97757
#e8e6dc
#f5f4ed
#ffffff
#6b6963
#3d3929
#c15f3c
Fonts: Source Serif 4 (headings) + Inter (body), loaded from Google Fonts
- 4 of 9 colours exact from memory alone — including the signature orange #d97757 and ink #141413
- 3 near misses: two warm whites close to brand Light, one muted gray with no palette equivalent
- 2 invented shades, including a darker clay #c15f3c for hover states the brief never asked for
- Both fonts wrong against the spec — a serif display face and Inter, where the skill prescribes Poppins and Lora
| Objective count | With skill | Baseline |
|---|---|---|
| Exact-match brand colours used | 7 of 7 | 4 of 9 |
| Near-miss colours | 0 | 3 |
| Invented colours | 0 | 2 |
| Heading font matches spec (Poppins) | Yes | No — Source Serif 4 |
| Body font matches spec (Lora) | Yes | No — Inter |
The baseline's memory of the palette was genuinely good — it named the signature orange to the exact hex. What it could not do was stay inside the palette or recall the fonts. The skill's entire value is that discipline, and on this task it was measurable: 7 of 7 exact against 4 of 9, and 2 of 2 fonts against 0 of 2.
Full outputs: With-skill invitation (full output) · Without-skill invitation (baseline)
Trigger tests
| Prompt | Expected | Result |
|---|---|---|
| “Restyle this slide deck in Anthropic's brand colours and fonts” | trigger | Triggers — Anthropic brand styling is the description's first sentence. |
| “Make this landing page match Anthropic's look and feel” | trigger | Triggers — 'look-and-feel' is the description's own wording. |
| “Apply our company design standards to this report” | skip | Likely triggers anyway — the description claims 'company design standards' generically while the content covers exactly one company, so your report gets Anthropic's palette. |
| “Choose a colour palette and fonts for my bakery's new brand” | skip | Borderline — 'brand colors' and 'typography' are the skill's stated territory; if it fires, the bakery gets Anthropic's identity. |
| “Summarise this PDF into three bullet points” | skip | Skips — no styling, formatting or brand element in the request. |
Desk-simulated against the skill's description matching rules; noted as simulated wherever the score depends on it.
Security screen
✓ Pass
Credential theft / exfiltration
✓ Pass
Remote code download & execute
✓ Pass
Data sent to unrelated third parties
✓ Pass
Hidden instructions (incl. invisible Unicode)
✓ Pass
Obfuscated payloads
✓ Pass
Privilege escalation requests
Screening only — not a security certification. The package is two text files (SKILL.md and an Apache-2.0 LICENSE.txt): we read both in full and ran a pattern and byte scan — no scripts, no assets, no URLs, no non-ASCII characters. Industry context: Snyk's ToxicSkills scan of 3,984 public skills found security flaws in 36.82% of them.
Install
Claude Code (plugin)
/plugin marketplace add anthropics/skills /plugin install example-skills@anthropic-agent-skills
Any agent (skills CLI)
npx skills add anthropics/skills --skill brand-guidelines
Cursor / Codex (manual)
# copy the brand-guidelines folder into: .cursor/skills/brand-guidelines/ # Cursor project .agents/skills/brand-guidelines/ # shared agents folder
Alternatives
frontend-design (Anthropic)
The opposite job. frontend-design invents a distinctive look when no brand book exists; brand-guidelines constrains output to one fixed identity. Use the first to create a direction, this one to stay inside somebody else's.
Fork it with your own values
For any brand other than Anthropic's, the realistic alternative is this file itself: 73 lines, Apache-2.0, seven colour slots and two font stacks to replace. That is also an admission about the original — the template is more reusable than the content.
Frequently asked questions
Is the brand-guidelines skill safe?
In our screening on 2026-10-23, brand-guidelines passed all six security red lines (credential theft, remote code execution, data exfiltration, hidden instructions, obfuscated payloads, privilege abuse) and scored 8/10 on the safety dimension. Screening is not a certification — re-screen after major updates.
What is the brand-guidelines skill used for?
Anthropic's official brand card: seven exact colours and two font stacks for styling artifacts with Anthropic's look-and-feel.
How do I install the brand-guidelines skill?
For Claude Code, add the marketplace and install the plugin set that contains it. With the skills CLI: npx skills add anthropics/skills --skill brand-guidelines. For Cursor or Codex, copy the brand-guidelines folder into .cursor/skills/ or .agents/skills/. Full commands are in the Install section of this page.
How good is the brand-guidelines skill?
brand-guidelines scored 66/100 (grade B) in our live test on 2026-10-23: Seventy-three lines of exact, copy-paste brand data that measurably beat a strong baseline on fidelity — 7 of 7 colours exact, both fonts right. But it is a data card, not a workflow: no steps, no checks, a trigger that can fire on other companies' brands, and feature text describing a PowerPoint script that does not ship.
Tested 2026-10-23 in an agent environment following the skill's own instructions — not inside the Cursor desktop app. Scores follow the Skills List methodology v1: red lines veto regardless of score; install counts and stars never count toward quality. More in Marketing & Content.