Most marketers prompt Claude like a chatbot. Anthropic's own engineers prompt it like a structured collaborator — with verification, planning, specific context, and reusable skills. The gap between those two modes is the gap between an average output and a publishable one.
In May 2026 Anthropic published their official Claude Code best-practices page and the Claude channel released a walkthrough video on the Explore → Plan → Code → Commit workflow. Austin Marchese's distillation of those talks pulled 334K views in six days. The signal was loud: the formal playbook is interesting because most people are not running it.
Below are the four rules from those sources, with the source citation per rule, and a concrete marketing example for each. Skip the developer-specific bits — the rules are general.
The 4 Rules
Give Claude a way to verify its work
Pair every task with a tangible check — tests, screenshots, expected outputs, or a written rubric Claude can compare against.
Why it works (cited)
Anthropic's official Claude Code best practices page (code.claude.com/docs/en/best-practices) calls verification "the single highest-leverage thing you can do." Without a checkable reference, Claude produces output that looks right but may not actually match what you wanted — and you become the only feedback loop, catching every mistake by hand.
Marketer example: brand voice check before publishing
What most marketers do
"Write me a LinkedIn post about our new pricing page."
What the rule says to do instead
"Write a LinkedIn post about our new pricing page. Compare the draft against @brand-voice-guide.md and @top-3-past-posts.md. After drafting, flag any line that breaks a brand-voice rule and rewrite it. Return the final post plus a one-line diff for each flagged line."
Why it helps
Claude now has something concrete to check against. You stop reading every output line by line and start spot-checking only the flagged diffs.
Explore first, then plan, then code, then commit
Four phases, not one giant prompt. Read the context, write a plan, execute the plan, then ship.
Why it works (cited)
The Anthropic Claude channel walks through this exact loop in their official video, "The Explore → Plan → Code → Commit workflow in Claude Code." The docs put it bluntly: "Letting Claude jump straight to coding can produce code that solves the wrong problem." Same problem in marketing — let the model draft before you research the angle and you ship the wrong angle in a beautiful voice. Anthropic also notes when to skip planning: "If you could describe the diff in one sentence, skip the plan." For marketers that means a typo fix or a single-word swap goes straight to draft; a new sales-page hero earns the full loop.
Marketer example: launch announcement post
What most marketers do
"Write a LinkedIn launch post for our new free tool."
What the rule says to do instead
Phase 1 Explore: "Read the new tool page at @tool-page.md and our three best launch posts at @past-launches.md. Summarize the audience, the offer, and the hook pattern that worked." Phase 2 Plan: "Outline three angle options for the launch post: feature-led, problem-led, contrarian. For each, sketch the hook, the body promise, and the CTA. Recommend one." Phase 3 Draft: "Write the recommended angle as a LinkedIn post, max 250 words, ending on the CTA you outlined." Phase 4 Commit: " Suggest two A/B variations on the hook line, then save the final to @drafts/2026-05-22-launch.md."
Why it helps
Each phase has one job. The model does not skip from headline to closing CTA without ever validating the angle. You get a paper trail you can review before publishing — and the next launch reuses the same loop.
Provide specific context — files, patterns, symptoms
Reference exact files, point to existing patterns Claude should follow, describe symptoms not solutions.
Why it works (cited)
Anthropic's docs give a direct contrast: "fix the login bug" versus "users report that login fails after session timeout. check the auth flow in src/auth/, especially token refresh. write a failing test that reproduces the issue, then fix it." The first prompt sends Claude on a guessing tour. The second names the symptom, the suspected location, the example pattern (failing test first), and the success condition. Same delta in marketing: the more precise your instruction, the fewer corrections you write.
Marketer example: low-converting landing page
What most marketers do
"Improve this landing page copy."
What the rule says to do instead
"Our pricing page at @pricing.md is converting at 1.2% — half of the @past-launches.md average of 2.5%. Read @pricing.md plus our three top-converting pages in @top-pages/. Compare hook structure, social-proof placement, and CTA pattern. Identify the three biggest differences and rewrite the pricing page hero and first CTA to match the top-page pattern. Do not change the pricing data."
Why it helps
You named the symptom (1.2% vs 2.5%), the location (pricing.md), the pattern to copy (top-pages/), the change scope (hero + first CTA only), and the constraint (do not touch pricing numbers). Claude has one job, not five.
Use skills instead of bloating CLAUDE.md
Procedures and reference material go in SKILL.md files. CLAUDE.md is for short, always-on facts only.
Why it works (cited)
Anthropic's best-practices page is explicit: "CLAUDE.md is loaded every session, so only include things that apply broadly. For domain knowledge or workflows that are only relevant sometimes, use skills instead. Claude loads them on demand without bloating every conversation." The rule of thumb in the same doc: "Would removing this cause Claude to make mistakes? If not, cut it. Bloated CLAUDE.md files cause Claude to ignore your actual instructions." Same logic for marketing. Your brand voice guide, ICP profile, and content audit checklist do not need to load on every chat about every topic. They load when the Skill description matches the task.
Marketer example: brand voice as a skill
What most marketers do
Pasting your 2-page brand voice guide into every chat where you ask Claude to write something.
What the rule says to do instead
Create a folder at ~/.claude/skills/brand-voice-auditor/ with a SKILL.md file. Frontmatter: name: brand-voice-auditor and description: Audits a draft against our brand voice rules. Use when a teammate or AI tool produces copy and we need to verify it matches our voice before publishing. Body: the four-step audit process. Add references/brand-voice-guide.md and references/top-3-past-posts.md alongside. Now Claude loads the full guide only when the situation matches the description — every other chat stays fast.
Why it helps
You stop pasting. The Skill becomes the single source of truth your whole team uses. Updating the brand voice means editing one file, not finding every chat where you pasted the old version.
A Note on the Rule Count
Marchese's widely-shared video extracted four rules from Anthropic's AI Code Summit talks. Anthropic's official best-practices page lists more headed sections than that — context management, course-correcting, subagents, hooks, permissions, MCP, scaling. The four rules above are the ones every other section composes onto: verification gates the output, the explore-plan-code-commit workflow is the loop, specific context is the prompt-quality layer, and skills are how knowledge persists across conversations.
Everything else in Anthropic's page (manage context with /clear, use subagents for investigation, set up hooks for deterministic actions, configure permissions) is operational tuning on top of these four. For marketers running Claude through the apps, the four rules are the complete playbook.
The Marketing Loop: Research → Plan → Draft → Publish
Rule 2 (explore-plan-code-commit) becomes the spine of every marketing task. The table below shows how the four phases translate.
| Engineer phase | Marketer phase | What Claude does |
|---|---|---|
| Explore | Research | Reads the trend signals, past best posts, competitor angles. No drafts yet. |
| Plan | Outline / Angle | Sketches 2-3 hook options with body promise and CTA. You pick. |
| Code | Draft | Writes the picked angle inside the platform's format rules. |
| Commit | Publish / Schedule | Adds two A/B variants on the hook, saves to your queue. |
Most marketers try to compress all four into one prompt and wonder why the output feels generic. Each phase has its own job. Anthropic's docs also note: if you could describe the change in one sentence, skip the plan. A typo fix or a single-CTA swap goes straight to draft. A new launch deserves the full loop.
Common Mistakes Marketers Make (and What the Rules Say to Do)
Mistake: One giant prompt with everything in it
Marketers paste the brand voice guide, the past posts, the angle, the platform constraints, and the CTA all into one message. Claude tries to honor every constraint and ends up hitting none of them well. Rule 2 says split it: research first, plan second, draft third.
Mistake: Vague hook brief ("make it punchier")
"Punchier" means nothing to the model. Rule 3 says reference the pattern. Pull the three highest-performing past hooks and ask Claude to match their structure (question-first, stat-first, story-first). The model has something to copy, not a vibe to guess at.
Mistake: Skipping the verification step because the output "looks fine"
Anthropic calls verification the highest-leverage thing you can do. The marketing equivalent: every draft gets compared against your brand voice guide before you publish. Rule 1 turns this from a manual chore into a Claude task — the model flags its own violations and rewrites them.
Mistake: Re-pasting your brand voice guide into every chat
Rule 4 fixes this directly. Encode the brand voice as a Skill. The model loads it when the situation matches the description (drafting copy that needs voice compliance) and ignores it when it doesn't (asking about ad spend benchmarks). Zero re-pasting, single source of truth.
Mistake: Treating Claude like a one-shot copywriter
The rules together describe a structured collaborator — one that researches, plans, drafts, and verifies. Marketers who use Claude as a copy generator ship faster but also ship more generic output. Marketers who run the full loop ship slightly slower per piece but ship at a quality bar that compounds. Google's Antigravity 2.0 covers similar workflow territory from the IDE side — the rules below remain the same regardless of which agent surface you run.
How Brand Brain Bakes These Rules In
If you want the same loop without managing folders and frontmatter, Brand Brain runs the four rules inside an app.
- Verification (Rule 1): the Brand Voice Analyzer learns your voice from your existing posts and audits every new draft against it automatically. Free, no signup.
- Research → Plan → Draft → Publish (Rule 2): Brand Brain's agent loop runs the four phases on a schedule. You set the topic, the agent researches trend signals, drafts angles, picks one, generates the post in your voice, and queues it for review.
- Specific context (Rule 3): every draft references your past best-performing posts plus the platform format rules. No vague briefs.
- Skills (Rule 4): Brand Brain ships 30+ marketing-specific skills in production (brand voice analyzer, content repurposer, post generators, audit checklists). You get the Skill recipes without writing SKILL.md files. See the 12 marketing Skills guide for the recipe list.
If you want to write your own first SKILL.md instead, the free Claude Skill Builder tool generates valid frontmatter and flags vague descriptions before you commit it to your folder.