AI content stays generic until the brief carries a market, a buyer and that buyer's KPIs. At AccuKnox, a cloud and AI security company, I mapped 11 markets and 33 personas into MagiHQ by hand before it wrote one draft.

Magi is an AI marketing platform that drafts content from the knowledge you load into it. The setup work was manual, and it decided the quality of every draft that followed. I tell the wider story in my two-year recap at AccuKnox.

TLDR

A brief with an ICP (ideal customer profile) and a persona changed the first paragraph of the same draft from the same model.
Map 1 market to 1 ICP to 3 personas before you generate anything.
Give every market 4 competitors and 3 influencers, so the model knows the field.
Require 11 frontmatter fields on every draft, so it arrives structured.
Building all 11 markets took weeks of working sessions, and the drafts came after.

Structure Beat the Model, and 11 Markets Came Before One Draft

I ran one task twice with Claude: write the opening paragraph of an AI-SPM (AI security posture management) blog post for AccuKnox.

The same model opens on a generic count without an ICP and persona, and on shadow AI with them.

Brief 1 held five product capabilities and the audience "cloud security buyers." The draft opened on how many models run in the cloud and then listed the capabilities in order.
Brief 2 added one ICP, one persona, five pain points and five KPIs from the Magi setup. The draft opened on shadow AI and an auditor asking how many AI assets you run.

The model, the product and the market stayed the same. Only the brief changed. Both drafts sit in the Appendix.

Watch out. Claude wrote both drafts for this post, and Magi did not. They test the brief, and they say nothing about Magi's own writing.

I made the same point on X: every AI content tool I tried breaks at one step, which is writing before anyone decided what it should know about the company.

Six steps build that structure. Five of them are inputs, and generation comes last.

Five structured inputs go in by hand, and generation is the sixth and last step.

Step 1. Collect What the Company Already Knows

Pull every source the company owns. Classify none of it yet.

FAQs and llms.txt
Whitepapers and eBooks
CVE pages and security advisories
Help docs and sitemaps
Playbooks for sales, onboarding and objection handling
Sales call transcripts and objection logs

Sales calls and proof-of-concept conversations usually make better source material than published pages, because people explain the product plainly there. We still chose what to extract and how to classify it, and nothing arrived through automatic ingestion.

Tip. Collect first and classify second. Step 4 decides which campaign each file serves.

Shann Holmberg found 7 things that separate a knowledge wiki you keep from one you abandon.

Step 2. Map 1 Market to 1 ICP to 3 Personas

Each market became one campaign with its own ICP, three personas, competitors and influencers. Eleven markets give 11 ICPs and 33 personas.

Each of the 11 markets runs as one campaign with 1 ICP, 3 personas, 4 competitors and 3 influencers.

The three personas follow one pattern in every market:

User: the hands-on operators, such as security, DevSecOps and platform engineers.
Buyer: the decision makers with budget, such as the CISO, the VP of Security and the AI governance lead.
Partner or reseller: channel and account managers, MSP managers and integrators.

Every persona carries a description, use cases, pain points and the KPIs that person is judged on. Here is the full market grid inside Magi.

Magi's Design and Market view lists all 11 configured markets, including CNAPP, DSPM, AI-SPM, KSPM, CWPP and ASPM.

Each card is a campaign anchor. Open one and you see its ICP, personas, competitors and influencers.

Example. The AI-SPM User persona lists shadow AI, no standards, rapid rollout, poor visibility and compliance gaps as pains. Its KPIs are AI asset coverage, policy violations, risk events, audit readiness and MTTR. Those ten items are the input behind the second draft above.

This is the AI-SPM market with its use cases and its three personas mapped.

The AI-SPM ICP in Magi lists pain points and use cases, then three personas for user, buyer and partner.

Step 3. Add 4 Competitors and 3 Influencers per Market

The rule for competitors is 4 per market, 2 broad platforms and 2 niche players. Each one gets a name, a website, a LinkedIn page, a blog and a one-line positioning summary. A model with no map of the field writes as if it has never seen the market.

Influencers are 3 per market, each with a name, a title and a LinkedIn link. We tracked them for signal. The people closest to a category write about its problems, and that is free market research.

The captured AI-SPM view shows three competitors filled and one slot open, plus five influencers, which is two more than the rule.

The AI-SPM market in Magi maps Protect AI, HiddenLayer and Robust Intelligence, plus five named influencers.

Step 4. Map Knowledge Files to Campaigns

One question drives this step: which file belongs to which campaign? Without the answer, a large source library is storage. With it, Magi pulls the right asset into the right campaign.

My brag list on atharvashah.com counts 648 campaign ideas generated from this setup.

Samiksha Shukla shares a 21-page context engineering guide on memory, retrieval, chunking and tools.

I use the same inputs-first pattern to run an AI marketing team with Claude skills.

Step 5. Require Eleven Frontmatter Fields

Frontmatter is the metadata block at the top of a draft. Every AccuKnox blog needed all 11 fields at generation time.

Every draft carries 11 required SEO frontmatter fields at generation time, so it arrives close to publishable.

Tone guidance, category rules and required links went into the brief next to the fields. The draft then arrived close to publishable.

Rule of thumb. A field you leave optional is a field the model skips.

MagiHQ sits under Content and Marketing on my tools page.

The MagiHQ card on atharvashah.com/tools, in the Content and Marketing section.

Step 6. Generate From the Brief

The generation form takes a prompt, a campaign, a content type and ideas. This brief asked for a 1,500-word CNAPP blog post that compares AccuKnox's AI-powered Zero Trust CNAPP with Claude Code Security Review.

The Magi brief form with a CNAPP blog prompt, campaign, content type and idea, beside the agent's research steps.

Magi reads the campaign's knowledge, finds a template and builds an outline. The brief is grounded in one campaign, one ICP and one knowledge source, and the buyer's pain shows up in the first paragraph, as the two drafts above showed.

Shann Holmberg calls a knowledge base the easiest low-effort win for an AI marketer.

The full AccuKnox story, with its numbers, is in the AccuKnox case study.

The AccuKnox case study page on atharvashah.com.

Five Questions a Newcomer Asks About the Magi Setup

Do I need Magi to use this? No. Any tool that takes a brief can hold the same inputs. A folder of Markdown files for the markets, ICPs, personas and competitors is a working first version.

Why 3 personas per market? Every market had a user, a buyer and a partner. Add a fourth only when a market has a fourth buyer who would read a different piece.

Why leave revenue size, industry, employee size and country blank? Those fields stayed empty on purpose, so segmentation happens at targeting time.

Where do I start collecting knowledge? Start with sales calls and help docs. They hold the plain-language explanations that published pages lose.

Does a better model fix generic content? The test above held the model fixed and changed only the brief. The opening changed, so the brief carries the quality.

The Appendix Holds Both Briefs and Both Drafts

Brief 1 holds features only:

Company: AccuKnox, a cloud and AI security company
Market: AI-SPM (AI security posture management)
Content type: blog post
Audience: cloud security buyers
Product capabilities: discover AI assets, govern model usage, secure pipelines, monitor access, enforce AI policies

Claude wrote Draft 1 on 2026-10-13:

Most security teams cannot say how many AI models, datasets, and endpoints run in their cloud today. Engineers spin up models in minutes, and each one reads data, calls APIs, and answers to no policy. That gap is what AI-SPM exists to close. AI security posture management starts with discovery of every AI asset, then governs which models people may use. It secures the pipelines that train and serve them, monitors who has access, and enforces AI policies. This post shows how AccuKnox helps cloud security buyers do each of these steps.

Brief 2 adds the ICP and the persona:

ICP: operates AI and LLM workloads without dedicated AI security tooling, and supports AI features used in its products
Persona: AI-SPM User, the hands-on operator
Pain points: shadow AI, no standards, rapid rollout, poor visibility, compliance gaps
KPIs: AI asset coverage, policy violations, risk events, audit readiness, MTTR

Claude wrote Draft 2 on 2026-10-13:

Your team shipped an AI assistant last quarter, and now three more models run in production that nobody on the security side approved. This is shadow AI, and it grows faster than any inventory you keep by hand. Each model has its own access keys, data sources and usage rules, or none at all. When an auditor asks how many AI assets you run and who governs them, you have no complete answer. AI-SPM fixes this by discovering every AI asset, enforcing policies on model usage and showing risk events as they happen. You get coverage you can prove and audit evidence that is ready on request.

The 11 frontmatter fields:

title, excerpt, slug, author, meta.title, meta.description, tldr,
primary_keyword, secondary_keywords, sticky_cta, faqs

How I Mapped 11 Markets Before MagiHQ Wrote a…