What we built

Castmint (castmint.studio) is an autonomous social-content engine — it learns a brand's voice, generates posts on schedule, and runs in "unattended-30" mode where it can ship content for a full month without a human touching it. Live, in production.

The shape of what we shipped

The pipeline is four stages:

1. Voice training. The operator hands us 20-30 real posts in their voice (LinkedIn, X, blog snippets). We extract a voice profile — sentence length distribution, signature phrases, banned patterns, hooks they use. 2. Idea generation. A weekly Claude job produces 10-15 post candidates from the operator's pillar topics + current events in their industry. 3. The judge. Every candidate runs through a strict brand-voice judge prompt. Binary pass/fail. ~40% pass on the first run; the rest are dropped (NOT auto-fixed — auto-fix drifts the voice). 4. Per-channel envelope. Each passing post is reshaped into platform-native envelopes (X = 280 char + thread; LinkedIn = hook + body + question; etc.) — same idea, different shape.

The whole thing runs through a 12-gate anti-ban policy cascade before publish.

Why "unattended-30" is the bar

Most "AI content tools" need daily babysitting. That's just an autocomplete with a UI. The real test is: can the operator step away for 30 days and the brand still sound like them, without getting banned? That's what "unattended-30 mode" certifies.

The honest result

Castmint is a real product. Whether it's profitable is a separate question — the curriculum is about the technique, not the revenue. The technique works; the business work (sales, distribution, pricing) is the operator's job.

What the next lessons cover

• Lesson 2: the voice training profile — what to extract, what to ignore, how to test it captured the voice.
• Lesson 3: the judge prompt + envelope code — copy-pasteable artifacts in production today.