The AI Crypto CMO Stack

// JUDGMENT LAYER

Operator-grade rules applied on top of every model output.

// 01 — RESEARCH & INTEL

Perplexity Nansen

Real-time competitive monitoring. On-chain wallet behaviour. The two research surfaces nobody can replicate without operator-grade inputs.

// 02 — CONTENT PRODUCTION

Claude ElevenLabs

Long-form drafts, positioning docs, email sequences, structured briefs. Multilingual audio for community ops.

// 03 — COMMUNITY OPS

Telegram tooling Discord tooling

Community pulse, KOL outreach shortlist, narrative tracking. Where most crypto marketers either over-automate or under-engage.

// 04 — PAID ACQUISITION ANALYTICS

Dune Analytics

On-chain attribution dashboards. The only post-acquisition behavioural layer no off-chain stack can replicate.

// 05 — EXPERIMENTATION

PostHog

Product analytics, feature flags, A/B testing across web and app. Self-hostable for teams that need behavioural data without handing it to an ad platform.

// 06 — REPORTING & SCRIPTING

Cursor

Internal data pipelines, automation scripts, custom reporting interfaces. At solo-operator scale, this eliminates the engineering dependency that used to gate every custom tool request.

// INSTALL

Install the stack.

Shipped as a Claude skill. Works in Claude Code, Cursor, and any agent supporting skills.

$npx skills add jukkablomberg/northpoint --skill crypto-cmo-ai-stack

Source: github.com/jukkablomberg/northpoint

// JUDGMENT LAYER

Five rules.

If AI output reads as generic to a native crypto user in 3 seconds, kill it — don't edit it.

Generic output is a brief problem, not a drafting problem.
Every AI-drafted piece gets one human-only contrarian pass before shipping.

"What would the most credible critic say in three words?" If you can't answer, it's not ready.
Never let AI write the first 20 words of anything.

Coherent is not compelling. The hook comes from a human who has felt the problem.
Volume is not a strategy.

Ship less, with higher conviction. Trust is burned faster than it's built.
AI output inherits the quality of the brief, not the model.

80% effort on the brief, 20% on the output. The reverse loses.

// NOT FOR AI

Three things we don't automate.

// 01 — CRISIS COMMS

First draft of crisis communications.

The first words from the brand carry the weight of whether trust survives. Competent is the wrong register for a crisis.

// 02 — FOUNDER VOICE

Founder-voice content.

The slightly-off sentence, the unhedged opinion — these are the signal a human with convictions wrote this. AI corrects them out.

// 03 — TASTE CALLS

Taste calls on community health.

Whether a meme is in-group or cringe, whether a narrative is signal or noise — pattern recognition only built by being inside crypto communities.