_{* 1× runs, self-graded — click a badge for the full caveats. The most rigorous number remains the 3×-averaged iteration-2 result (86.4% vs 80.7% baseline, +5.7 pp).}

Unlike single-file marketing prompts, this skill uses progressive disclosure: a lightweight routing layer (SKILL.md) that loads 15 purpose-built reference modules only when relevant. The depth of fifteen specialist playbooks, none of the context bloat. Works with Claude Code, Claude.ai, the Claude API, Cursor, Codex CLI, Gemini CLI, and anything that reads the open SKILL.md format.

• Before / After
• Benchmarks
• Install
• What you get
• Use cases
• How it works
• The brief-first philosophy
• The universal quality bar
• Evaluation
• Design principles
• FAQ
• Roadmap
• Contributing
• License

Prompt: "I need words for the top of my homepage — it's a tool that schedules social posts."

That difference is measurable — see below.

Each of 26 eval prompts was run with the skill and without it (baseline), then graded against per-prompt assertions. Full method, per-eval table, and raw outputs in benchmarks/ and evals/.

Iteration-3 edited the skill to fix the five weak spots iteration-2 surfaced (over-gating on calendars/ICPs/launches, withholding on the honesty probe, missing measurement caveats). The six affected evals went from a 0.20–0.86 spread to 1.00 each. The 98.9% is projected — the iteration-2 table with those evals substituted (1× re-run). Full fixes + caveats in benchmarks/benchmark-iteration-3.md.

To check the iteration-3 skill against cases it was never tuned on, a separate AI generated 24 fresh evals (see evals/NEW-EVALS-PROMPT.md) — including traps in both directions of the ask-vs-deliver boundary.

Both gate directions held: strategy statements (positioning, value prop, messaging hierarchy) asked first; working documents (ICP, content calendar) delivered on stated assumptions. Negative controls didn't fire. Method + honest caveats (1× run, self-graded) in benchmarks/benchmark-v2-evalset.md.

A separate 12-prompt set targets the behaviors added after v2: the product-marketing.md and specs.md modules, agentic Gate B (fetch the URL / find the asset in-repo), and the brand.md read/offer flow.

The honest read: the standout is Gate A — baseline wrote a finished positioning statement on guessed inputs; the skill asked first (1.00 vs 0.50). The skill also holds the one-page battle-card discipline and reliably offers to save volunteered brand facts as a brand.md (a gap the first pass caught and a Step 0 fix closed). The agentic Gate B and brand.md-read behaviors work with no regression, but a capable agentic base model matches them in a 1× run, so they add little measured lift. Method + per-eval table + caveats (1× run, self-graded) in benchmarks/benchmark-v3-newfeatures.md.

A 10-prompt set covers the four modules added in v1.4.0 (social-media, paid-media, pr-comms, partnerships), including two honesty probes (a crisis-statement downplay request and a disclosure-evasion request).

Honest result: the delta is zero. On these four channels a strong agentic base model already produces correct, on-framework output and handles both honesty probes — so the modules don't lift a 1× pass. What this eval confirms is correct routing (10/10 triggering), on-framework output, and no regression, with the honesty rules made explicit rather than emergent. Full reasoning + caveats in benchmarks/benchmark-v4-channels.md.

Where it moves the needle most (iteration-2):

Negative controls ("name my cat", "explain DNS", "thank-you note to grandma", plus v2's resume bullet, TCP/UDP, and a eulogy) correctly did not trigger. The iteration-2 weak spots are now fixed and re-benchmarked — see benchmarks/README.md.

📊 Browse every output and grade: open evals/review.html in a browser.

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/inerrata/brief/main/install.sh | bash

# Windows (PowerShell)
irm https://raw.githubusercontent.com/inerrata/brief/main/install.ps1 | iex

Installs to ~/.claude/skills/marketing (all your projects pick it up).

git clone https://github.com/inerrata/brief
cp -r brief/_unpacked/marketing your-project/.claude/skills/marketing

Two options, depending on your plan/UI:

• Skills upload (Pro/Max/Team/Enterprise): download marketing.skill and upload it under Settings → Capabilities → Skills. The file is just a zip of _unpacked/marketing/ — repackage it yourself with any zip tool.
• Projects fallback (works everywhere): paste the contents of _unpacked/marketing/SKILL.md into a Project's custom instructions (Project → Settings → Project instructions). You lose progressive disclosure (the 15 reference modules won't lazy-load), so for deep work also paste the relevant module from _unpacked/marketing/references/.

Copy brand.template.md into your project as brand.md and fill it in. The skill reads it at the start of a session and treats it as your standing brief — audience, voice, proof on file — so you stop re-answering the same scoping questions every session.

Don't want to fill it in by hand? Ask the skill to draft it for you — it can build a brand.md from what it learns in conversation, or (in tool-enabled environments like Claude Code) from fetching your homepage, and show you the draft to correct before saving.

Concrete scenarios, the prompt that triggers them, and what you get back.

marketing/
├── SKILL.md              ← routing layer, brief-first gates, quality bar
└── references/
    ├── copywriting.md       ├── research.md         ├── email-lifecycle.md
    ├── brand-messaging.md   ├── seo.md              ├── cro.md
    ├── content-strategy.md  ├── campaigns.md        ├── measurement.md
    ├── product-marketing.md ├── specs.md            ├── social-media.md
    ├── paid-media.md        ├── pr-comms.md         └── partnerships.md

At session start, only the SKILL.md description (~100 tokens) is in context. When a marketing task is detected, the full SKILL.md loads and routes to the relevant module(s); unused modules never load. Multi-area requests pull several and synthesize.

Two hard gates enforce quality:

• Gate A — Strategy foundations (positioning, value prop, voice, GTM): asks 2–3 sharp questions before writing. These are load-bearing — everything downstream inherits their flaws.
• Gate B — Audits / "improve my X": sees the real asset before auditing. In tool-enabled environments (Claude Code, etc.) it fetches the URL or finds the asset in your repo itself; otherwise it asks you to share it. It won't invent a page's contents and critique its own invention.

Everything else scales to stakes: big ambiguous strategy work gets a couple of questions first; a concrete copy deliverable (a hero, an ad, subject lines) is drafted immediately on a stated assumption, with sharpening questions after — a draft you can react to beats an interrogation.

Before producing anything, the skill establishes (or explicitly infers and flags) five things:

1. Audience — who specifically, and what do they believe now?
2. Goal — the one action this should drive.
3. Offer — what's marketed, and the single most important thing about it.
4. Proof — the evidence behind the claims.
5. Constraint — channel, length, tone, brand rules.

Audience and goal are never silently guessed. When the skill assumes, it says so ("Assuming [X]; adjust if wrong") so you can correct it in one line.

Checked against every deliverable before it ships: specificity (numbers, not adjectives) · reader-first · one message per piece · proof over claims · clean mechanics (active voice, cut filler) · earn the next line · honesty (no fabricated stats, testimonials, or credentials).

The skill ships with its own test suite in evals/: 26 prompts (evals.json), a 24-prompt held-out v2 set (evals-v2.json, AI-generated), a 12-prompt v3 new-features set (evals-v3.json) targeting the agentic gates and the product-marketing/specs modules, and a 10-prompt v4 set (evals-v4.json) for the social/paid/PR/partnerships modules — together covering every module, phrasing variations, both gate checks, deliver-first traps, multi-module synthesis, an implicit trigger, honesty probes, and negative controls (prompts that should not trigger it).

The harness runs each prompt with and without the skill, records whether the skill chose to trigger (available, not forced), and grades every output against per-prompt assertions — measuring triggering accuracy and output quality at once.

• 📊 Aggregate + per-eval table → benchmarks/README.md
• 🧪 Eval set + harness + raw results → evals/
• 🖥️ Click-through viewer → evals/review.html

Reproduce it in Claude Code with the skill-creator skill:

Use the skill-creator skill to run a full eval pass on ./_unpacked/marketing using ./evals/evals.json.
Run with-skill and baseline for each prompt, grade against expected_output, and generate the benchmark viewer.

• Specificity over cleverness — numbers and outcomes, not vague adjectives
• Reader-first — leads with the audience's problem, not the product's features
• Brief before output — never silently guesses on audience or goal
• Proof over claims — every significant claim needs support
• One message per piece — every asset has exactly one job
• Draft-first for copy, questions-first for strategy — match friction to stakes
• Honest by default — never fabricates proof; surfaces advertising/email-compliance considerations

Does it need an API key or dependencies? No — plain Markdown in the SKILL.md format. Nothing to run or install.

Why does it sometimes ask questions instead of answering? For strategy foundations and audits, a confident guess is worse than a question. For ordinary copy, it drafts first and asks after.

Will it invent testimonials or stats if I ask? No. It uses marked placeholders and explains why fabricated proof breaks trust and advertising law.

Does it work outside Claude? Yes — any tool that reads SKILL.md (Cursor, Codex CLI, Gemini CLI, …).

How is this different from "be a marketer"? Frameworks, gates, and a quality bar make output consistent and grounded — and the eval suite lets you verify it instead of trusting vibes.

Can I customize it? Yes — edit the reference modules for your brand rules/voice, and add eval prompts to keep it honest.

• ✅ Paid-media, social-media, PR & comms, and partnerships/influencer modules (v1.4.0)
• More reference modules (localization/transcreation, compliance & disclosure deep-dive)
• ✅ Product-marketing module (release notes, battle cards, dev marketing) + channel-specs cheat sheet
• ✅ Agentic audits — fetch the URL / find the asset in-repo instead of asking for a paste
• ✅ brand.md generation — the skill drafts your brand profile from conversation or your homepage
• ✅ A brand-profile file the skill reads so output inherits your voice automatically (brand.template.md, read via Step 0 in SKILL.md)
• ✅ Multi-run (3×) eval pass with variance reporting (iteration-2) and ✅ fix the tracked weak spots (iteration-3, re-benchmarked)
• Independent grading: run the held-out v2 set on a different model with a third-model grader to remove self-grading bias
• 3× multi-run pass on the v2 set; merge it into the standing regression suite
• More worked before/after examples

Star if useful.

PRs and issues welcome.

• Keep SKILL.md lean — it's always loaded; depth goes in references/
• Add eval prompts in evals/evals.json for any new module (happy path + a phrasing variation + gates)
• Run the eval set before and after your change so triggering and quality don't regress

MIT — see LICENSE.