This handbook is written for senior operators who want to run a marketing organisation through Claude -- not as a one-shot copy generator, but as the daily operating substrate that drafts briefs, executes campaigns, audits SEO, and reports outcomes. It is built for founders, CMOs, VPs of Marketing, and senior strategists who already have a working theory of growth and need a faster way to execute it.

Every chapter is grounded in Anthropic's first-party documentation at docs.claude.com. Where we paraphrase, we attribute. Where we quote, we quote in full and link the page. Where we make a claim that goes beyond the docs -- a workflow, a cost discipline, an operating cadence -- the claim is grounded in Dcrayons' own engagement history, not a third-party blog. That is the only way a handbook on AI operations ages well.

Dcrayons has been AI-first since 2024. The plays that follow are the ones that have shipped, broken, been re-shipped, and now sit in production across our client portfolio. Skip the chapters that do not apply to your stage; the next-step prompts inside every chapter are written so you can pick a single one and start tomorrow.

The case for picking a single model family for marketing operations

Marketing teams who treat AI as "I will paste a prompt into whichever chatbot has a tab open today" never get past surface-level use. The discipline that produces real leverage is the opposite: pick one model family, learn its strengths, structure your prompts and projects around its features, and only reach for a second model when a task it cannot do well shows up. This handbook makes the case that Claude, built by Anthropic, is the right default choice for marketing operations.

What Anthropic positions Claude as

Per docs.claude.com, Claude is "a highly performant, trustworthy, and intelligent AI platform built by Anthropic. Claude excels at tasks involving language, reasoning, analysis, coding, and more." The platform offers two distinct surfaces for building with Claude: the Messages API, which gives direct model-prompting access for custom agent loops and fine-grained control, and Claude Managed Agents, a pre-built agent harness for long-running tasks and asynchronous work. Marketing teams that want to embed Claude into bespoke campaign workflows will use the Messages API; teams that want an out-of-the-box agent for things like "watch this Slack channel for brand mentions and summarise nightly" will use the Managed Agents.

Three traits that matter for marketing operations

1. A model family that ranges from cheap to top-tier

Per the models overview at docs.claude.com, the current generation includes three models: Claude Opus 4.8 ("Anthropic's most capable model for complex reasoning and agentic coding"), Claude Sonnet 4.6 ("the best combination of speed and intelligence"), and Claude Haiku 4.5 ("the fastest model with near-frontier intelligence"). Pricing per million tokens is $5/$25 (Opus), $3/$15 (Sonnet), and $1/$5 (Haiku) for input/output respectively. A marketing team can confidently route bulk classification work to Haiku, day-to-day drafting and analysis to Sonnet, and the weekly strategy synthesis to Opus -- without leaving the Claude family.

2. Tools built for long-running work

Claude's features overview at docs.claude.com lists capabilities that map directly to marketing operations: extended thinking for complex analysis, prompt caching (both 5-minute and 1-hour windows) to keep recurring context cheap, batch processing at "50% less than standard API calls" for bulk content tasks, citations for grounded research output, structured outputs for predictable JSON when building dashboards, and web search and web fetch for grounded competitive research. These are not afterthoughts; they are first-class features.

3. Context windows large enough to hold real brand context

Per the models overview, both Claude Opus 4.8 and Sonnet 4.6 ship with 1M-token context windows, and Claude Haiku 4.5 ships with 200K. Anthropic notes that 1M tokens corresponds to roughly 555K-750K words. That means an entire brand book, a year of campaign briefs, and the last two quarters of performance dashboards can all live inside a single prompt without summarisation. For a marketing team, this is the feature that makes the model usable across the lifecycle of a project, not just an ad-hoc task.

When Claude is the wrong default

We will return to this in Chapter 8: Where Claude breaks (and what to use instead), but the honest version is: for live-data SERP analysis and Google Search Console integration, Gemini wins because Google ships that integration; for image generation, OpenAI's DALL-E, Midjourney, or Google's Imagen are the right calls; for highest-throughput translation at the cheapest price point, dedicated translation APIs still beat any general-purpose model. Default to Claude; reach across when the task is one of these.

What to do this week

Pick the single marketing surface that currently absorbs the most senior time. That is the surface to test Claude on first. Read the Intro to Claude page at docs.claude.com (linked from this handbook's appendix), make your first API call, and write down the cost-per-task numbers you observe. The rest of this handbook is structured so you can keep adding surfaces in priority order.

Source: Intro to Claude (docs.claude.com/en/docs/intro), Models overview (docs.claude.com/en/about-claude/models/overview), Features overview (docs.claude.com/en/build-with-claude/overview). Last reviewed 2026-06.

From hobby use to production use

The fastest way for a marketing team to underuse Claude is to keep it stuck inside the "open a tab, ask a question, copy the answer" pattern. The fastest way to over-claim is to declare that "Claude will run our marketing." Neither is true. What is true is that Claude can fully execute a defined surface end-to-end when you scope the surface, write the system prompt, give it the right data, and put a senior human at the review gate. This chapter lists the twelve surfaces Dcrayons has confidently moved into production for clients.

Content surfaces

1. Brief generation

Input: a topic + a target keyword cluster + the brand fact sheet. Output: a structured content brief with H1/H2 outline, search-intent classification, target word count, and a list of competitor pages to reference. Claude's structured outputs feature (docs.claude.com/en/build-with-claude/structured-outputs) guarantees the JSON schema you ask for, which keeps briefs pasteable into your CMS without manual cleanup.

2. First-draft writing

Input: the content brief + the brand voice document + 3-5 of your own published articles for tone calibration. Output: a first draft that already sounds like the brand. The "few-shot" pattern is the lever; the example articles do more work than any instruction.

3. Editorial sweep

A drafted article passes through Claude with the brand forbidden-words list, the SEO entity list for the topic, and the style sheet. Output: a marked-up rewrite with notes on what was changed and why.

Ads surfaces

4. Angle generation

Input: the product + the audience persona + the offer. Output: 30 angles, grouped by emotional driver (gain, fear, novelty, social proof, identity). Senior strategist picks the top five.

5. Headline + description sets

Each chosen angle produces a Google RSA set (15 headlines + 4 descriptions) and a Meta primary-text + headline set (5 variants). Claude's prompt caching keeps the brand fact sheet warm across the loop, which makes the bulk variant generation cheap (more on this in Chapter 7).

6. Creative scripting

Input: an approved angle + the 5-second hook the editor wants. Output: a 30-second video script with shot directions and on-screen text. Editor cuts the final.

SEO surfaces

7. Entity graph audit

A site crawl is dropped into Claude with the question "what entities does this site authoritatively cover, and what is missing for the topic cluster X?" Output: a gap list and a content plan to close it.

8. Schema markup generation

Given a page URL + the rendered HTML, Claude emits the JSON-LD schema. Structured outputs gives schema validation for free.

9. llms.txt + AI-search citation copy

Claude drafts the llms.txt file and the canonical fact paragraphs that AI engines will cite. This is the surface that drives "share of answer" -- the citation-engineering metric we obsess over.

Operational surfaces

10. Weekly performance digest

Pipe GA4, GSC, Meta Ads, and Google Ads weekly extracts in. Claude returns a one-page digest with the three things to do next week, written for the founder, not the analyst.

11. Competitor monitoring

A scheduled Claude job (via Routines on Claude's surfaces, or your own cron) fetches the top three competitor pages weekly and reports on changes (new sections, new CTAs, new offers). Claude's web fetch tool (docs.claude.com/en/agents-and-tools/tool-use/web-fetch-tool) makes this a 10-line implementation.

12. Internal "ask the brand" knowledge base

Marketing team's brand book, voice doc, campaign archive, and product specs are dropped into a Claude Project (covered in Chapter 3). Any team member can ask "what do we say about X" and get an answer the senior strategist would sign off on.

What does NOT belong on this list

Approving spend over a threshold. Sending a final email to a customer list. Replying to a journalist. Publishing to the corporate blog without human review. The discipline is that Claude drafts; humans sign off. The drafting saves the hours; the signing-off keeps the brand safe.

What to do this week

Pick the three surfaces above where your team currently feels most behind. That is your 30-day rollout list. Do not start with surface 10 (weekly digest) -- start with surfaces 1, 2, and 7 (brief, draft, entity audit), which produce visible, measurable wins inside the first week.

Source: Features overview (docs.claude.com/en/build-with-claude/overview), Structured outputs (docs.claude.com/en/build-with-claude/structured-outputs), Web fetch tool (docs.claude.com/en/agents-and-tools/tool-use/web-fetch-tool). Last reviewed 2026-06.

Why a Project beats a system prompt

The most common failure mode in marketing AI is "the same brand brief gets copy-pasted into every prompt, every day, by every team member, with subtle variations." Brand voice drifts. SEO entities go missing. Forbidden words slip in. The fix is to externalise the brand context into one place that every prompt reads from -- a Claude Project, or its functional equivalent inside your own application.

What lives inside a Project

Brand voice document

Three sections: tone (what we sound like), pace (sentence length, paragraph structure), forbidden words (the 30-50 words and phrases you do not want in any output). Anthropic recommends giving Claude a role and instructions in the system prompt; the brand voice document is a long-form, instructive system prompt that goes beyond a single sentence.

Brand fact sheet

The 50 facts about your brand that should never be wrong: founding year, founder names, office locations, certification list, named methodologies, current pricing tiers, currency support, time zones served. Every drafted output is checked against this sheet.

SEO entity list (per topic cluster)

For each topic cluster you cover (e.g. "GEO," "AEO," "Local SEO"), a list of the 30-50 entities that must appear in authoritative content on that cluster. This is the lever that turns a content draft into a content draft that ranks.

Voice samples

5-10 published articles in the brand voice, with light annotations on what worked. Claude is calibrated by examples more than by instruction.

Forbidden-words style sheet

This is the most underused asset on most marketing teams. A maintained list of words that signal AI sloppiness or hollow marketing -- "leverage," "in today's fast-paced world," "unlock," "game-changer." Claude is good at avoiding them when told explicitly.

Prompt caching: the cost lever that makes Projects work

Per docs.claude.com/en/build-with-claude/prompt-caching, prompt caching lets you "provide Claude with more background knowledge and example outputs to reduce costs and latency." Two cache durations are supported: the standard 5-minute window and the 1-hour extended window. The brand voice document + fact sheet + entity list is your cached block; the per-task prompt sits after it. This pattern keeps the brand context warm across a working session, which means the bulk-generation passes (30 angles, 15 RSA headlines) get cheaper as the day continues.

Automatic prompt caching

Anthropic also ships automatic prompt caching, which "simplifies prompt caching to a single API parameter. The system automatically caches the last cacheable block in your request, moving the cache point forward as conversations grow." For teams building their own application surfaces, this is the easiest way to get the caching benefit without manually managing breakpoints.

How to maintain a Project

Three rules. One: the brand voice document is owned by the senior strategist, version-controlled in git or in your CMS, and reviewed quarterly. Two: the brand fact sheet is owned by the COO or chief of staff, refreshed whenever a fact changes (a new office, a new pricing tier). Three: the SEO entity list is owned by the SEO lead, refreshed whenever a topic cluster expands. Anyone on the team can use the Project; only the owners can edit the underlying docs.

What to do this week

Write the brand voice document. Three pages. Start with five existing articles you are proud of and reverse-engineer the tone, pace, and forbidden words from them. Drop the result into a Claude Project. Re-route the team's brief-generation work through that Project. That alone produces a noticeably more consistent first draft inside a week.

Source: Prompt caching (docs.claude.com/en/build-with-claude/prompt-caching), Automatic prompt caching (docs.claude.com/en/build-with-claude/prompt-caching#automatic-caching), Features overview (docs.claude.com/en/build-with-claude/overview). Last reviewed 2026-06.

The four-stage content pipeline

The content workflow that survives production is not "Claude wrote the article." It is a four-stage pipeline where each stage has its own prompt, its own evaluator, and its own success metric.

Stage 1: Brief

Input: target keyword + search intent + the brand fact sheet + the SEO entity list. Output: structured brief (H1, H2 outline, target word count, primary keyword, secondary keywords, entities to cover, CTAs, internal links to point at). This stage uses Claude's structured outputs to lock the JSON schema so downstream stages can rely on it.

Stage 2: Draft

Input: the brief + the brand voice document + 3 brand voice samples + the prompt-caching breakpoint that holds all of the above. Output: a 1,500-2,500 word first draft. Anthropic's extended thinking feature (docs.claude.com/en/build-with-claude/extended-thinking) gives Claude room to reason through the structure before generating -- worth turning on for long-form drafts.

Stage 3: Edit

Input: the draft + the forbidden-words list + a sharper instruction set ("tighten 20% without losing entity coverage; remove every instance of the forbidden words; restructure any paragraph longer than 4 sentences"). Output: an edited draft with a diff summary. Senior editor reviews the diff, not the whole article.

Stage 4: SEO sweep

Input: the edited draft + the entity list + the schema requirements. Output: a final pass that adds missing entities naturally, generates the JSON-LD schema for the page, drafts the meta title and description, and produces a list of suggested internal links. The schema generation uses Claude's structured outputs to guarantee valid JSON-LD.

The single most important discipline

Each stage's output gets stored. When something breaks downstream -- the article ranks poorly, the editor flags a tone problem -- you can rerun a single stage with a corrected prompt, not redo the whole pipeline. The cost of running stages 2 and 3 again is non-trivial; the cost of running stage 1 again is roughly free thanks to prompt caching.

How to keep the cost discipline

Claude's pricing per million tokens (per docs.claude.com/en/about-claude/models/overview) gives you the budget math: a 2,000-word draft with a 5,000-word cached prefix routes to Sonnet 4.6 at about $0.10 per generation when caching is active. A team producing 100 articles a month spends roughly $10 on the drafting stage. The cost trap is running drafts on Opus 4.8 when Sonnet 4.6 produces equivalent quality; default to Sonnet, escalate to Opus only when the editor flags a quality gap on a specific topic.

Batch processing for catalogue work

For programmatic SEO or large content sweeps -- 500 product pages, 200 city pages -- Anthropic offers batch processing: "Process large volumes of requests asynchronously for cost savings. Send batches with a large number of queries per batch. Batch API calls cost 50% less than standard API calls." This is the right tool for one-shot bulk generation; not for interactive editing.

What to do this week

Pick a single topic cluster. Run all four stages on three articles. Compare to the three articles the team produced last quarter on the same cluster. The differences in entity coverage, on-page SEO compliance, and editorial cleanliness will surface immediately. That comparison is the artefact you use to get the rest of the team on board.

Source: Extended thinking (docs.claude.com/en/build-with-claude/extended-thinking), Structured outputs (docs.claude.com/en/build-with-claude/structured-outputs), Batch processing (docs.claude.com/en/build-with-claude/batch-processing), Prompt caching (docs.claude.com/en/build-with-claude/prompt-caching). Last reviewed 2026-06.

From hours per ad set to minutes per ad set

The single largest time sink on most paid media teams is the "we need fresh creative" tax: thirty angles into thirty briefs into thirty creative tickets into thirty Asana cards. Claude collapses the first half of that pipeline into hours, leaving the creative production where it should be: with humans, with cameras and editors.

Workflow 1: 30 angles in 20 minutes

Input: product fact sheet (kept warm in cache), target persona, current offer, the channel. Prompt asks for 30 angles structured by the five emotional drivers in marketing: gain, loss aversion, social proof, novelty, identity. Output: a structured list with the angle, the audience it targets, and the recommended channel.

Workflow 2: RSA + Meta primary-text generation

Each chosen angle produces:

Google RSA: 15 headlines (max 30 chars), 4 descriptions (max 90 chars), with character-count enforcement in the prompt.
Meta primary-text: 5 hook variants, 5 body variants, 5 CTA variants -- 125 combinations.
LinkedIn sponsored: 3 headline variants + 3 body variants tuned to the persona.

Structured outputs locks the schema so the team can ingest the JSON straight into the ad platform.

Workflow 3: A/B test wiring

Input: the angle list + the success metric (CAC, CTR, ROAS, NTB %). Output: a test plan with hypothesis, MDE calculation, sample-size requirement, and the holdout cell. This is the discipline most paid teams skip -- Claude is good at not skipping it when prompted properly.

Workflow 4: lifecycle copy

Input: the lifecycle flow (welcome series, abandoned cart, post-purchase, win-back) + the brand voice + the relevant offers. Output: full email sequences with subject lines, preview text, body, CTAs, and the suggested send time relative to the trigger event. The brand voice cache keeps the cost flat across the sequence.

The right model per task

Angle generation rewards Opus 4.8 -- the difference in angle quality between Opus and Sonnet is visible inside the first ten outputs. Variant generation can ride on Sonnet 4.6 (or Haiku 4.5 for the LinkedIn variants where character economy matters more than emotional nuance). Lifecycle copy is Sonnet's sweet spot. Per docs.claude.com/en/about-claude/models/overview, Sonnet 4.6 is "the best combination of speed and intelligence" and Haiku 4.5 is "the fastest model with near-frontier intelligence" -- both are honest descriptions of when each model belongs.

What does not work yet

Direct integration with the ad platforms. Today, you still export the generated copy to a CSV or paste it into the platform UI. Anthropic ships computer use (docs.claude.com/en/agents-and-tools/tool-use/computer-use-tool), described as "Control computer interfaces by taking screenshots and issuing mouse and keyboard commands," which makes the platform-paste step technically automatable -- but for a production marketing team in 2026, the security and audit-trail tradeoffs are not yet worth it. Wait for the platform-native APIs to mature; or build a thin import script for your platform of choice.

What to do this week

Pick your highest-spend ad set. Run Workflow 1 + 2 against it. Ship three new variants in the next test slot. Measure CTR + CVR delta against the control. The first cycle takes a day; subsequent cycles take an hour.

Source: Models overview (docs.claude.com/en/about-claude/models/overview), Structured outputs (docs.claude.com/en/build-with-claude/structured-outputs), Computer use tool (docs.claude.com/en/agents-and-tools/tool-use/computer-use-tool). Last reviewed 2026-06.

SEO is now two games at once

In 2026 every senior SEO operates two simultaneous programs: the classic ten-blue-links program (still 80% of qualified traffic for most brands) and the AI-search citation program (the growing slice of top-funnel discovery). Claude is well suited to both because the underlying work -- entity research, schema generation, citation engineering -- is structured language work.

Workflow 1: Entity graph audit

Input: a site crawl (URL + H1 + first paragraph + key entities mentioned per page) + the topic cluster you want to own. Claude returns a list of entities you authoritatively cover, entities you mention but do not authoritatively cover, and entities you do not mention at all. Then it proposes the content plan to close the gap.

Workflow 2: Schema markup generation

Schema markup is one of the cleanest fits for Claude. Input: the rendered HTML of a page + the schema type required (Article, Product, FAQPage, LocalBusiness, Organization). Output: valid JSON-LD that passes Google's Rich Results Test on the first try. Structured outputs enforces schema conformance: "Guarantee schema conformance with two approaches: JSON outputs for structured data responses, and strict tool use for validated tool inputs."

Workflow 3: llms.txt drafting

llms.txt is the emerging file standard for telling AI engines what your site is, who it is for, and which pages are the canonical sources for which topics. Claude is well suited to drafting it because the input -- your sitemap + the topic cluster you want to anchor -- is the kind of structured context Claude excels at. Output: a draft llms.txt with the standard sections (about, products, key topics, citation-anchor URLs).

Workflow 4: AI Overview optimisation

Workflow: pull the queries that currently trigger an AI Overview for your category from your favourite SERP tracker. For each query, Claude inspects your ranking page and produces a rewritten H2 + answer paragraph that matches the citation pattern Google's AI Overviews prefer (question-led heading, concise answer, follow-up depth). The discipline is to track AIO presence as a metric in its own right.

Workflow 5: citation engineering for ChatGPT, Gemini, Perplexity

The citation engineering loop: run a 100-query benchmark of the queries you want to be cited for, capture which sources each AI engine cites today, identify the brand-fact paragraphs they could cite from your site, restructure those paragraphs to be more citable (one-fact sentences, named entity per sentence, dated). Run the benchmark again in 30 days.

Web fetch + web search

For competitive SEO research, Anthropic's web fetch tool ("Retrieve full content from specified web pages and PDF documents for in-depth analysis") and web search tool ("Augment Claude's comprehensive knowledge with current, real-world data from across the web") let Claude do the competitive scan inline. For SEO tasks specifically, web fetch is often the right pick -- you usually already know the URL you want to study.

What to do this week

Pick one topic cluster. Run Workflow 1 against it. Use the gap list as your next month's content roadmap. The entity gap report is one of the easiest wins to demo internally because the gaps are concrete and the fixes are scopable.

Source: Structured outputs (docs.claude.com/en/build-with-claude/structured-outputs), Web fetch tool (docs.claude.com/en/agents-and-tools/tool-use/web-fetch-tool), Web search tool (docs.claude.com/en/agents-and-tools/tool-use/web-search-tool). Last reviewed 2026-06.

Why most marketing reports are unread

The standard weekly marketing report has too many dashboards, too many tabs, and not enough text. Founders skim it. Operators do not act on it. The format Claude is best at generating is the opposite: a one-page narrative with three things to do next week, written for the decision-maker, not the analyst.

The Dcrayons weekly readout structure

One page. Five sections. Every Monday.

Headline number: the one revenue metric we are paid to move + delta vs last week + delta vs the four-week trailing average.
Wins: the three biggest moves of the week with the metric and the action that drove them.
Misses: the three issues + the diagnosis + the planned correction.
Next-week priorities: three actions, named, with the owner + the success metric.
Dcrayons Score readout: the AI-search citation share + entity coverage + technical SEO score, tracked week-over-week.

The data pipeline

Claude is not the data extractor. The platforms (GA4, GSC, Meta Ads, Google Ads, your AI-citation tracker) export to a flat CSV or JSON. The CSV is dropped into Claude with the prior week's report as a calibration example. Claude returns the new report in the same format.

Prompt caching makes the weekly cadence cheap

The prior 12 weeks of reports + the brand fact sheet + the metric definitions live in the cached prefix. Each week's incremental cost is tiny -- only the new data and the new draft. This is the pattern that makes the "weekly readout in Claude" cadence sustainable at scale.

Compaction and context editing for long-running threads

For teams running the readout as an ongoing conversation rather than a fresh prompt each week, Anthropic ships two features that matter: compaction ("Server-side context summarization for long-running conversations. When context approaches the window limit, the API automatically summarizes earlier parts of the conversation") and context editing ("Automatically manage conversation context with configurable strategies. Supports clearing tool results when approaching token limits"). Both let a single Claude thread carry a whole quarter of weekly readouts without ever hitting the context wall.

Computer-readable outputs

When the readout has to feed a dashboard or a notification system, ask Claude to emit a JSON sidecar alongside the human-readable report. Structured outputs enforces the schema so the downstream system never breaks on a malformed payload.

What to do this week

Take last week's marketing report. Drop it into Claude as the calibration example. Drop this week's data alongside. Ask Claude for the one-page narrative version. Run it past your CMO or the founder. The reception will tell you whether the cadence sticks.

Source: Prompt caching (docs.claude.com/en/build-with-claude/prompt-caching), Compaction (docs.claude.com/en/build-with-claude/compaction), Context editing (docs.claude.com/en/build-with-claude/context-editing), Structured outputs (docs.claude.com/en/build-with-claude/structured-outputs). Last reviewed 2026-06.

An honest list of the gaps

Claude is the right default for almost every marketing task we have shipped. It is not the right call for every task. This chapter lists the surfaces where, today, you should reach for a different tool, and the honest reason why.

1. Live Google SERP analysis

Gemini wins because Google owns the SERP. For an SEO operator who needs to see the live SERP for a query, Gemini's integration with Google Search and Google Search Console is tighter than what Claude's web search can do. Use Gemini for live SERP work; bring the output back into Claude for analysis.

2. Image generation

Claude generates text and code. It does not generate images. For static ad creative, hero shots, or thumbnails, Midjourney, OpenAI's DALL-E, Google Imagen, and Stability are the working set. For brand-controlled output, Midjourney with a careful style reference is the production default for Dcrayons creative team.

3. Video generation

Same as images. The current production-grade tools are Google Veo, OpenAI Sora, Runway, and Luma. Claude can write the script, the shot list, and the post-edit instructions; the video itself is generated elsewhere.

4. Voice synthesis

For audio ads or YouTube voice-overs, ElevenLabs and OpenAI's TTS are the working defaults. Claude writes the script; the audio is generated elsewhere.

5. High-throughput translation

For one-shot translation of a few hundred strings, Claude is excellent. For ongoing high-throughput translation across a multilingual catalogue, dedicated services like DeepL and Google Translate are still cheaper per call and have specialised post-editing features Claude does not.

6. Real-time analytics queries against your warehouse

Claude is good at writing the SQL. It is not the right tool for running the SQL against your warehouse on every dashboard refresh. Production analytics belongs in your BI tool (Looker, Mode, Hex); Claude writes the queries that your analysts paste in.

7. Voice + visual brand identity work

Claude can describe a brand identity in text and point at design references. It cannot generate the brand identity itself. Designers and creative directors are still the source of the brand; Claude is the operator that scales it once built.

8. Compliance + legal sign-off

A regulated marketing message (financial product disclosures, pharma claims, gambling) needs a human legal reviewer, not an AI summary. Claude is excellent at flagging compliance risk; the sign-off has to come from the team that takes the regulatory liability.

9. Tasks that benefit from Gemini's multi-modal Google Workspace integration

If your marketing org runs on Google Docs + Google Sheets + Google Slides and you want an AI inside that workflow, Gemini's integration is tighter than any external API. Use it for in-doc work; bring the output back into Claude for the more structured downstream work.

10. Anything you cannot give Claude the data for

If the source data is locked inside a vendor portal with no export, no API, and no browser-automation path that meets your security bar, Claude is not the right tool. Solve the data access problem first; then add Claude on top.

What to do this week

Audit your current marketing AI stack. List every tool. For each tool, ask: "is Claude better at this?" For half the list the answer will be yes; consolidate. For the other half, document why the non-Claude tool wins, and codify when to reach for it. The discipline is to choose deliberately, not by which tab is open.

Source: Features overview (docs.claude.com/en/build-with-claude/overview), Models overview (docs.claude.com/en/about-claude/models/overview). Last reviewed 2026-06.

Adoption beats capability

Most marketing teams already have access to Claude (via the API, via claude.ai, via a Claude Code subscription, or via their cloud platform). The question is whether the team actually uses it on the surfaces where it produces leverage, every day. This chapter is a 90-day plan that gets a marketing org from "we have Claude" to "Claude is how we run the function."

Days 1-14: Foundations

Week 1: Pick the single senior strategist who will own the Claude rollout. Build the brand voice document + brand fact sheet + the SEO entity list for the top topic cluster. Drop them into a Claude Project. Pick three surfaces from Chapter 2 as the 30-day rollout list.

Week 2: Run all three surfaces twice each. Compare to the previous quarter's manually-produced output on the same surfaces. Write a one-page "what changed" memo. Share with the rest of the team.

Days 15-45: Team rollout

Bring two more team members into the workflow. Each owns one of the three surfaces for the next month. The senior strategist remains the editor of last resort. Document the prompts. Add the second topic cluster's entity list to the Project. Begin the weekly readout cadence from Chapter 7.

Days 46-75: Surface expansion

Add three more surfaces from Chapter 2. By day 75 the team is shipping work through six of the twelve. Cost discipline: log the monthly Claude spend; expect it to be in the low hundreds of dollars per active surface per month at typical marketing-team volumes. Most teams find this an order of magnitude cheaper than the analyst time it replaces.

Days 76-90: Custom integrations

The team that has shipped six surfaces is ready to invest in custom integrations: scheduled jobs via Routines or your own cron, a custom MCP server for your CMS, an internal "ask the brand" knowledge base accessible from Slack. The Agent SDK ("the Agent SDK lets you build your own agents powered by Claude Code's tools and capabilities, with full control over orchestration, tool access, and permissions") is the right tool for the custom integrations.

What does NOT work in the 90 days

Trying to run all twelve surfaces simultaneously from day one. Skipping the brand-voice document. Routing everything through Opus when Sonnet would do. Letting the cost discipline lapse. Building integrations before the manual workflow is proven.

Measuring the win at day 90

Three numbers. One: hours per surface per week before vs after. Two: editorial quality scored on a held-out sample (blind review by the senior strategist). Three: the revenue metric the marketing function is paid to move -- a 90-day window is too short for a definitive read but you should see directionally consistent movement against the trailing baseline.

What to do this week

Print this chapter. Block 90 minutes with your CMO or the founder. Walk through the plan. Adjust the surface ordering for your org. Pick the start date. The plan only matters when somebody owns it.

Source: Claude Code: Overview (code.claude.com/docs/en/overview), Agent SDK (referenced in the Claude Code Overview), Features overview (docs.claude.com/en/build-with-claude/overview). Last reviewed 2026-06.

Why we wrote this handbook

Dcrayons has been AI-first since 2024. We pivoted the moment Google AI Overviews started absorbing top-funnel clicks, and our share-of-answer in ChatGPT, Gemini, and Perplexity has been our first KPI since. The plays in this handbook are the plays we run for clients every week. They have shipped, they have broken, and they have shipped again. We wrote this because the version of AI-marketing advice you find on most agency blogs is generic and undated. This is the version that has receipts.

Who we are

500+ campaigns shipped. 100+ active clients. AI-first since 2024. Verified Shopify Partner and Amazon Ads Partner agency. Senior strategist on every account. Offices in Delhi (HQ) and Sheridan, Wyoming. We work with founders, CMOs, VPs of Marketing, and senior operators on programs scoped above Rs 25 lakh per quarter.

The one next step that matters

If you have read this far, the next step is a written Dcrayons Score readout for your brand. It tells you, in one business day, where you are losing share-of-answer inside ChatGPT, Gemini, Perplexity, and Google AI Overviews -- and the 90-day plan to close the gap. The readout is free. The plan is yours to keep.

Three ways to get the readout

WhatsApp +91 96678 13600 with your website + the single revenue metric you want to move.
Email info@dcrayons.app with the same.
Book a 30-minute scoping call at dcrayons.app/contact.

A note on capacity

We take three to five new retainers per quarter. Senior strategist time is the constraint, and we will not promise capacity we cannot ship against. If the next quarter is full, we will tell you when we expect the next opening.

Where to go next

Read the Dcrayons Growth Formula at dcrayons.app/dcrayons-growth-formula. The Growth Formula is the 90-day sequencing playbook that maps spend across SEO, AI search, paid, and lifecycle channels back to a single revenue metric you pick. It is the framework underneath every program in this handbook.

Final thought

The brands that win in 2026 are not the ones with the biggest AI budget. They are the ones whose teams have actually rebuilt their daily operating cadence around AI. This handbook is one step toward that rebuild. The next step is to pick one chapter, take one action this week, and put a reminder on your calendar to check in next Monday. AI-first is a discipline. The discipline compounds.

Source: Dcrayons internal engagement records, 2024-2026. For external sources see chapter footnotes throughout this handbook. Last reviewed 2026-06.

The AI-First Marketer's Handbook: Claude as Your Operating System

Read the full guide