brand-guidelines (Anthropic)

Review

Codex CLI 0.166 (gpt-6-astra, low reasoning) on a Mac test machine, a fresh temporary project and temporary HOME/TMPDIR/caches per run, skill installed (treatment) or absent (control) or absent with the skill's core requirement squeezed into one sentence appended to the prompt (one-liner group), 2 runs per group per task, network off, /tmp writes blocked. The prompts never mention the skill (except the wording noted below). Three tasks that name Anthropic's official brand without giving any value: restyle deck.html in place, write brand.css for internal documents, and build a small self-contained status card (status.html). Scored by rules: at least 5 of the 7 brand hex codes, Poppins headings, Lora body, Arial/Georgia fallbacks.

Three-arm comparison

Measured with Codex CLI over 18 runs, scored by rule-based checks with no LLM judge. Total 26.6 / 30, verdict Situational.

Verdict: Situational (26.6 / 30). Source: anthropics/skills, skills/brand-guidelines, pinned at commit ca45fc4. Install: npx skills@latest add anthropics/skills --skill brand-guidelines. Licence: Apache-2.0 / Anthropic terms (see LICENSE.txt in skill).

What it is

The brand-guidelines skill from anthropics/skills: Anthropic's official brand colours (dark #141413, light #faf9f5, mid gray #b0aea5, light gray #e8e6dc, accents orange #d97757, blue #6a9bcc, green #788c5d) and typography (Poppins for headings with Arial fallback, Lora for body with Georgia fallback), plus how to apply them. One SKILL.md of about 330 words; no scripts.

How we tested

Codex CLI 0.166 (gpt-6-astra, low reasoning) on a Mac test machine, a fresh temporary project and temporary HOME/TMPDIR/caches per run, skill installed (treatment) or absent (control) or absent with the skill's core requirement squeezed into one sentence appended to the prompt (one-liner group), 2 runs per group per task, network off, /tmp writes blocked. The prompts never mention the skill (except the wording noted below).

Three tasks that name Anthropic's official brand without giving any value: restyle deck.html in place, write brand.css for internal documents, and build a small self-contained status card (status.html). Scored by rules: at least 5 of the 7 brand hex codes, Poppins headings, Lora body, Arial/Georgia fallbacks.

1. Restyle an HTML deck — Style deck.html with Anthropic's official brand colors and typography. Edit the file in place.
2. Write a branded stylesheet — Write brand.css for our internal documents following Anthropic's official brand guidelines (colors and typography). Save it as brand.css.
3. Branded status card — Create status.html: a small self-contained status card (title, three status rows) styled in Anthropic's official brand look and feel.

Scoring is programmatic (hidden tests / planted-issue recall / constraint checks / render+OCR where relevant), 0–5 per run. No LLM judge decided any score.

Results

The one-liner given to the third group: "Use Anthropic's palette (dark #141413, light #faf9f5, mid gray #b0aea5, light gray #e8e6dc; accents orange #d97757, blue #6a9bcc, green #788c5d) with Poppins for headings (Arial fallback) and Lora for body text (Georgia fallback)."

Mean score: 4.42 without vs 5.0 with the skill (delta +0.58). Tokens per run: 104k vs 56k (0.54x); time 64s vs 39s. One-liner group: 5.0 (the skill's gain over the one-liner: +0.0; one-liner's own gain over no skill: +0.58); one-liner tokens 52k per run, so the skill costs 1.08x the one-liner.

What happened

• Mean score 4.42 without the skill, 5.0 with the skill, 5.0 with a one-sentence version that lists the colours and fonts. The control group's misses came from the status-card task: in one of its two runs (1.9/5) the agent did not use the brand palette and fonts, in the other it did (4.64); on the deck and the stylesheet task control scored 5.0 in both runs.
• So plain Codex often knows the palette from training, but not reliably; with the skill every one of 6 runs hit the full palette and fonts.
• The skill was read in 6/6 runs and was cheaper than the control: 56k tokens per run against 104k (0.54x), and about the same as the one-liner (52k). On the deck task control used 155-166k tokens against 64k with the skill; we did not investigate why control spent more (it made 8 commands against 6 in the first run of each).
• The one-sentence version, which contains the same hex codes and fonts, scored the same as the skill (5.0) at the same cost, so the skill's value is the stored facts, not an elaborate method.

Dimension scores

Limits of this test

• A rule check of hex codes and font names, not a visual judgement of how brand-like the page looks.
• Three small tasks, two runs each; the control miss is a single run.
• The skill only has facts for one brand (Anthropic's); the test shows nothing about how well a skill like this works for another company's guidelines.
• n = 2 per group per task: differences under about 1 point are noise.
• Rule-based scoring measures what the checks can see; it cannot judge taste or whether a result is the best possible.

Security

See security.md. Scan: 0 CRITICAL / 0 HIGH / 0 MEDIUM. Verdict: safe to run (no-network sandbox). A clean scan does not prove a skill safe.

Bottom line

The skill reliably supplies the exact palette and fonts and in these runs it cost half the tokens of a bare request. A one-sentence prompt with the same values matches it, so the skill is worth using only if you want the facts stored once instead of repeated in each prompt.