AI content strategy: a practitioner's framework

Most "AI content strategy" writing tells you to "use AI to personalise your content" or "automate your workflows." After running this at production scale across a full site launch, 15 pieces commissioned, briefed, and published in the first two weeks, we can tell you that advice is almost entirely useless without the part that comes after it.

Here is the part that comes after it.

Why most AI content strategy content doesn't help you implement anything

The problem with most content on this topic is the people writing the most widely-read guides haven't actually run an AI content pipeline. They've researched running one. You can tell because the advice stops at the decision to adopt and never gets to the friction.

"Use AI for content research" — fine. Which tool, in what configuration, with what output format that your next stage can actually use? "Automate your brief generation" — great. What does the brief template look like, how does it connect to your editorial calendar, and who validates it before drafting starts?

The answers to those questions aren't strategy. They're operations. And operations is the part that determines whether your AI content pipeline produces ten pieces a month at a cost that makes sense, or produces six drafts in a folder and a lot of wasted hours.

This piece documents our actual pipeline. Every stage, the tooling, the timing, and the failure modes.

The full pipeline

This is the 10-stage pipeline we run for Robotic Pixels. Every piece of content on this site went through it.

flowchart TD
    A[1. Calendar & Brief] --> B[2. SEO Intelligence]
    B --> C[3. Claim & Evidence Binding]
    C --> D[4. AI Drafting — Claude]
    D --> E[5. Humanise & Voice Pass]
    E --> F[6. Hard Rules Validation]
    F --> G[7. Asset Production]
    G --> H[8. Pre-publish Audit]
    H --> I[9. CMS Publish]
    I --> J[10. Distribution]
    H --> |Fail — back to stage 4 or 5| D

The pipeline runs inside a Cowork environment with access to Notion (editorial calendar), DataForSEO (keyword and SERP data), the Cloudflare and EmDash CMS stack, Google Search Console, and a set of custom skill files that encode our editorial rules. Nothing is a black box. Every decision is auditable.

Stage 1: the calendar and brief system

Our editorial calendar lives in Notion. Every content item was staged there before a single word was drafted. At launch that meant 120 pieces planned across 12 weeks.

Each calendar entry has: title, pillar assignment, content type template, target keyword, secondary keywords, word count target, target CMS slug, and a status field. The status field drives the pipeline: Briefed → Evidence-Bound → Draft Ready → Image Ready → Audit Passed → Published.

The brief is the most important investment in the pipeline. We wrote 15 full-length briefs before any production started, each covering the angle, content structure, required assets, internal linking targets, schema requirements, SEO metadata, and distribution plan. The brief is the contract between strategy and production. Skimping on it shows up three stages later as a draft that needs fundamental restructuring, not editing.

Every brief enforces the angle discipline rule: the brief must articulate in one sentence what makes this piece different from the 10 pieces already ranking for the target keyword. If you cannot state that, the brief is not ready.

Stage 2: SEO intelligence

Before drafting, we run a SERP snapshot via DataForSEO: top 10 organic results, People Also Ask, AI Overview presence, featured snippets, keyword difficulty, monthly volume, and commercial intent classification.

This goes into the Notion calendar item as structured JSON. The content drafter reads it during generation, which means the draft is informed by what is actually ranking, not by what the brief assumes should rank. The practical difference: the drafter can see when a top-ranking piece is 6,000 words of comprehensive coverage and adjust scope accordingly, or when the top 3 are all thin definitional posts and there is a clear gap for something with actual depth.

We skipped this step for the first 15 pieces at launch to hit the publishing deadline. That was a deliberate trade-off, not a recommended shortcut.

Stage 3: claim and evidence binding

This stage does not exist in most AI content pipelines. It is the one we would build first if we were starting again.

Before drafting begins, a claim extraction pass runs on the brief. Every factual claim the brief implies the piece will make, statistics, expert quotes, competitive comparisons, gets extracted into a structured list. Then evidence gathering runs in parallel: we find the primary source for each claim, verify the methodology, and store the citation.

The output is a binding document: a Notion child page attached to the calendar item that maps every claim to its primary source. The content drafter then reads the binding document and writes each claim using the verified citation. Verbatim statistic, verbatim attribution string. No paraphrasing, no rounding, no "approximately."

The result: zero fabricated statistics in published pieces. Every number is traceable to a source we verified before drafting started. The pre-publish audit runs a final freshness and source-quality check, but the heavy lifting happens here.

Stage 4: the drafting workflow with Claude

The drafter is Claude, running as a Cowork skill with access to the complete brand reference package: voice guide, hard rules, taxonomy, audience profile, SEO pillars, content type template structures, EEAT rubric, and uniqueness checklist.

When the drafter runs, it validates the post type and brand, loads the full brand identity package, reads the binding document, writes the draft against the structural template, self-scores against EEAT, uniqueness, and quality rubrics, and revises once if any dimension scores below threshold.

For flagships, we use Claude Opus 4.6 for the initial draft and Claude Sonnet 4.6 for the self-scoring and revision pass. The Opus draft typically comes in at 2,200–2,800 words with stronger structural coherence; the Sonnet pass tightens it against the rubrics. The drafting step takes 20–40 minutes of compute time. Human involvement: reviewing the output and making the call on whether to revision-pass or proceed, 15–30 minutes per piece.

One thing we have found consistently: the drafter's structural instincts are good. Its voice is not. The structural template compliance is high; the practitioner specificity that makes RP content readable to someone who knows their field is what requires the next stage.

Stage 5: humanise and voice pass

AI-drafted content has tells. Left uncorrected, they undermine the practitioner positioning the entire strategy depends on.

The humanise skill runs a two-pass detection and rewrite cycle. Pass one identifies the AI formatting patterns: bullet abuse, "Understanding X" H2s, perfectly balanced tri-bullets, bolded keywords scattered through paragraphs, "It's important to note that" filler, em-dashes used as commas, rhetorical questions opening sections. Pass two rewrites flagged sections against the brand voice guide.

Then a human reads the entire piece out loud. This is the step no AI can fully replace. If a sentence sounds like a corporate training video, it gets rewritten. The read-aloud test catches rhythm problems, tone mismatches, and the subtle wrongness of sentences that are technically correct but sound slightly artificial. Human time: 30–60 minutes per flagship. This is not a step to compress.

Stage 6: hard rules validation

Our hard rules list has 15 rules, covering things like "never fabricate statistics," "never use marketing fluff," "never publish without the Built with Claude disclosure block," and "never ship AI formatting tells."

The validation skill runs the draft against all 15 and returns a binary pass/fail per rule with specific violation detail. A draft that fails any rule does not proceed. It goes back to the previous stage with the violation flagged.

The most common violations in our production: Rule 6 (marketing fluff in CTAs), Rule 14 (AI formatting tells surviving the humanise pass), and Rule 8 (statistics without sources flagged during binding but reintroduced by the drafter). The binding document in Stage 3 largely eliminates Rule 8 violations. When the drafter has verified citations to draw on, it uses them rather than inventing alternatives.

Stage 7: asset production

Every flagship needs a featured image, inline diagrams or screenshots where the content calls for them, and any downloadable resources referenced in the piece.

Featured images are generated by NanoBanana, a custom image generation pipeline using the Nano Banana 2 Pro model in a specific isometric wireframe illustration style. The image prompt is generated from the content item's title and SEO description. The result goes through a validation step that checks format (PNG), dimensions (1424×752px), and file size before it is accepted.

Mermaid diagrams are embedded directly in the Markdown source and rendered by the Astro build. Screenshots for workflow documentation pieces are taken manually. Asset production runs in parallel with the validation pass; the image generation starts while the hard rules check is running, which saves 5–10 minutes per piece at volume.

Stage 8: pre-publish audit

The audit is the final quality gate. It runs against EEAT score (target: 7+/10), uniqueness checklist (target: 4/6), universal quality rubric (target: 3.5+/5.0), all 15 hard rules, SEO readiness (title, description, keyword density, schema presence), statistics freshness (no primary citations older than 3 years), and schema completeness (Article + BreadcrumbList + FAQPage for flagships).

A piece that passes transitions to Audit Passed. A piece that fails transitions to Needs Review with a structured log of every failed check and the specific violation. The content drafter runs again from Stage 4 or 5 depending on the failure type.

We have had pieces fail the audit and pass on the re-run, and we have had pieces fail twice before a human made the call to handle the violation manually. The audit is a tool, not a gatekeeper. The human makes the final call on edge cases.

The real cost per piece

The honest, all-in cost breakdown for Robotic Pixels content in 2026:

The human time figure is based on Alexander's time allocation for the full pipeline including brief writing, review, and distribution setup. If you are running with a contractor writer at a different rate, substitute accordingly.

Compare to a UK mid-market content agency: £1,200–£4,500 per flagship post is the standard range for research-backed, senior-written long-form content (Sortlist UK content marketing pricing guide, 2024). At our pipeline cost, you are paying roughly 13–25% of the agency cost for a comparable output, verified through a structured audit process rather than an informal review cycle.

The economics improve with scale. Brief templates, binding documents, and the audit rubric all compound; each one you write makes the next one faster. The fixed investment to get from zero to a production-ready pipeline runs 80–120 hours including pipeline design, brief template development, brand reference file creation, tool integration, and test runs. At 10 pieces per week, that investment is paid back within the first quarter of production.

The time comparison

Traditional content agency model: brief to publish in 2–4 weeks, 2–3 rounds of revision, multiple touchpoints for brief sign-off and draft review, publishing pipeline handled separately.

RP's AI-first pipeline: brief to publish in 3–5 days for standard flagships (2 days if the calendar item and binding document are ready when production starts), audit-driven revisions (typically zero or one), Alexander's 2-hour batch review covers 4–6 pieces, publishing fully automated via EmDash CLI and Cloudflare Workers.

The time delta matters most for responsiveness. When a significant industry development happens, a major model release, a regulatory shift, a competitor move, we can commission, draft, audit, and publish a response piece in 48 hours. A traditional agency cycle cannot do that. The speed advantage is not marginal; it is structural.

Honest failure modes

Every pipeline has failure modes. These are ours.

The most persistent one is drafter over-reliance on unverified claims. When Stage 3 is skipped, the drafter introduces statistics it cannot trace to a source. The hard rules pass catches most of these, but they still require human resolution time. Do not skip Stage 3.

Editor bottleneck is the constraint we have not fully solved. When six to eight drafts arrive for humanisation in the same 48-hour window, the read-aloud step in Stage 5 breaks the production cadence. The current mitigation is staging output so no more than three or four drafts are ready simultaneously.

Asset production is the other common long pole. If a piece requires manual screenshots or custom diagrams and those are not ready before the audit stage, the piece stalls. Nothing enters audit without all assets confirmed.

AI formatting tells sometimes survive the humanise pass. The skill catches the obvious patterns reliably. The subtle ones (slightly unnatural sentence rhythm, faint corporate voice in transition paragraphs) still need human attention. The read-aloud step in Stage 5 is the only reliable catch for these.

Audit regressions are less common but worth noting. Occasionally a piece that passed hard rules validation fails the final audit because a revision introduced a new violation. Run the validation check again after any substantive edit.

Who should and shouldn't adopt this approach

Right for: marketing teams producing 5+ pieces per month who can justify the setup cost, agencies wanting to increase per-operator throughput without proportional headcount growth, in-house teams with a single reviewer who can become the voice authority, and any operation where production speed is the bottleneck rather than strategic direction.

Wrong for: teams producing fewer than 3–4 pieces per month (the overhead exceeds the benefit at low volume), highly regulated industries where every claim requires legal or compliance review before publication, and teams without a single authoritative reviewer. The voice consistency requirement means one person needs to own the brand voice and review every output before it publishes.

For context on the solo operator end of this: Alexander runs both Robotic Pixels and Campakt this way. A two-business operation producing 10 pieces per week is not a story about what AI makes possible in theory; it is what a single operator with the right pipeline can actually do. For the infrastructure side of running two businesses through AI agents, Alexander writes about it at his personal site.

Takeaway

AI content strategy is not about the tools. Every team now has access to Claude, a half-dozen AI writing assistants, and more automation platforms than anyone can evaluate. Tool access is not the differentiator.

The operational design is the differentiator: the pipeline architecture that turns brief → evidence → draft → human review → publish into a repeatable, quality-assured process rather than an ad hoc experiment that works sometimes and does not work other times.

This pipeline took the better part of six months to design, test, and get to production-reliable. It is not finished. The claim binding step is relatively new, the SEO intelligence integration is being standardised, and the distribution stage is still largely manual. But it works, the numbers are real, and the model is replicable.

For the optimisation layer on top of a working content pipeline — how to make published content visible to AI search systems rather than just traditional search — see our LLM visibility guide. For the Claude-specific skills that power several of the stages described above, see the Claude for Marketing hub.

Built with Claude

This post was produced using Claude as a research, drafting, and editing partner.

• Models: Claude Opus 4.6 for drafting, Claude Sonnet 4.6 for editing and self-scoring
• Workflow: brief → claim binding → AI draft → humanise pass → hard rules validation → audit → publish
• Production time: ~6 hours total (Alexander)
• Word count: 3130
• Human review: Alexander (final)

For more on how RP produces content with Claude at production scale, see the Claude for Marketing hub.