copywriting (coreyhaines31/marketingskills)

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). Each task gives the agent a short fact sheet (product.md, a fictional seaweed-snack subscription called Tidepool) and tells it to use only those facts. The skill's 'ask for context first' step was answered by the prompt.

Three-arm comparison

Measured with Codex CLI over 18 runs, scored by rule-based checks with no LLM judge. Total 24.0 / 30, verdict Situational.

Verdict: Situational (24.0 / 30). Source: coreyhaines31/marketingskills, skills/copywriting, pinned at commit 88c79f6. Install: npx skills@latest add coreyhaines31/marketingskills --skill copywriting. Licence: MIT (repo LICENSE).

What it is

The copywriting skill (v2.3.2) from coreyhaines31/marketingskills: a conversion-copywriting playbook. Before writing it looks for a product-marketing context file, then diagnoses the input (blank brief, existing voice, point edit); principles (clarity, benefits over features, specificity, one CTA), page-section structures, rules against AI tells ('it's not X, it's Y' reveals, 'no X, no Y, no Z' lists), supportable claims only. Three reference files (copy frameworks, AI tells, natural transitions) and an evals file. Markdown only. It triggers on 'write copy', 'rewrite this page', 'sounds like AI' and similar.

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).

Each task gives the agent a short fact sheet (product.md, a fictional seaweed-snack subscription called Tidepool) and tells it to use only those facts. The skill's 'ask for context first' step was answered by the prompt.

1. Landing page copy from a fact sheet — Write landing page copy for Tidepool (all facts are in product.md; use only those). Include: hero headline + subhead + CTA, three benefit blocks, three FAQ items. Save it to copy.md. Max 300 words.
2. Short-form ad variations — Write 3 Facebook ad variations for Tidepool using only facts in product.md. For each give 'Primary text:' (max 125 characters) and 'Headline:' (max 40 characters) lines. Save to ads.md.
3. Rewrite bland copy — Rewrite the paragraph in bland.txt as a sharper hero section (headline, subhead, CTA) for busy office workers. Use only facts from product.md. Save to hero.md, max 90 words.

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: "Write clear, specific, benefit-led copy with one primary call to action, use only facts you were given, avoid AI tell-tale patterns ("it's not X, it's Y", "no X, no Y, no Z" lists), and respect every length limit exactly."

Mean score: 5.0 without vs 5.0 with the skill (delta +0.0). Tokens per run: 63k vs 93k (1.48x); time 25s vs 29s. One-liner group: 5.0 (the skill's gain over the one-liner: +0.0; one-liner's own gain over no skill: +0.0); one-liner tokens 68k per run, so the skill costs 1.36x the one-liner.

What happened

• Every run in every group scored 5/5: all 18 runs were within the word and length limits, contained no invented numbers or claims, no banned buzzwords and none of the AI-tell patterns we check for ('it's not X, it's Y', 'no X, no Y, no Z'). With scores at the ceiling the test cannot show a quality gain; it shows the skill did not hurt.
• Reading the outputs, the three groups look alike: a headline built on the crunch/desk-drawer angle, three benefit blocks, a three-item FAQ copied from the fact sheet. We could not rank them by eye either.
• The skill was read in 6/6 treatment runs, so triggering works from the description alone (no prompt mentioned it).
• Cost: 93k tokens per run against 63k without the skill (1.48x) and 68k with the one-sentence version of its rules (1.36x); 7.7 commands against 5.7.

Dimension scores

Limits of this test

• A ceiling effect: with a fact sheet and an explicit 'use only these facts' instruction, plain Codex already writes clean, claim-free copy. The skill is designed for harder inputs (a blank brief, existing brand voice, competitive positioning, long pages); we did not test those.
• Taste, voice and conversion are not measured; the checks only catch banned words, AI-tell patterns, invented numbers and length limits.
• The skill's context file (.agents/product-marketing.md) and sibling skills (emails, popups, copy-editing) were not installed.
• 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

On fact-sheet copy tasks the skill changed nothing we can measure: the agent already avoids buzzwords, invented numbers and AI tells when told to use only given facts, and the skill adds about 1.5x tokens. It may help on blank-brief or voice-matching work, which this test did not cover, so it is a low-cost keep rather than a clear win.