Why 4 systems, not one
Every "AI content tool" we've tried treats content as a single problem. It isn't. The reason most AI content feels generic is that it was generated by a single model with a single prompt trying to do five different jobs at once.
We broke the problem into four discrete systems, each with its own model configuration, retrieval layer, and human review step. Here's what we actually built for a D2C skincare brand last quarter.
System 1: The voice fingerprint
Before we write a single piece of content, we build a voice fingerprint for the brand. Not a "brand voice document." An actual parameterised system prompt built from analysis of their best-performing existing content.
The process: 1. Collect 30–50 pieces of content the founder considers "on-brand" 2. Run them through a structured extraction prompt that identifies: sentence length distribution, filler words and verbal tics, argument structure (do they open with a hook or with context?), emotional register (warm? clinical? irreverent?), and what they never say 3. Encode the output as a system prompt addendum, not a separate document
The result: every downstream generation step inherits the voice fingerprint. You don't have to prompt for tone on every request—it's structural.
System 2: The content brief engine
The brief engine takes a topic (from a human or from a content calendar) and expands it into a structured brief before any draft is written. The brief includes:
- Primary keyword and 3 secondary keywords
- Target audience segment and their specific pain point
- Angle (there are only six angles that work: problem/solution, story, listicle, contrarian, tutorial, case study)
- Three competing pieces on the same topic and what gap this piece fills
- Mandatory inclusions (specific stats, product mentions, CTA requirements)
The brief is the spec. The draft must satisfy the brief. This sounds basic but it's what separates content that ranks from content that exists.
System 3: The draft-review pipeline
This is where the LLM does the writing—but it's not a single generation call. It's a pipeline:
1. **Outline generation** structured JSON output (sections, word targets, key points per section) 2. **Human outline review** a person approves or edits the outline (2 minutes) 3. **Section-by-section drafting** each section generated independently with the preceding section as context 4. **Coherence check** a second pass that reads the full draft and flags transitions, repetition, and brief violations 5. **Human final edit** a person does a 15-minute edit pass
The section-by-section approach is the most important architectural choice. It prevents the "middle sag" you get with single-pass long-form generation, where the model loses energy around section 3.
System 4: The distribution formatter
Content that lives only as a long-form article is content that doesn't earn its production cost. The distribution formatter takes the approved article and generates:
- Twitter/X thread (structured JSON, 8–12 tweets)
- LinkedIn post (single take, professional register)
- Instagram caption (short, hook-first, 3 hashtags max)
- Newsletter paragraph (150 words, reader-value framing)
- Short-form video script (45 seconds, hook + payoff structure)
All five formats are generated in one pipeline run, in parallel. Total incremental cost: about 12 cents in API fees per article.
The actual stack
- **Orchestration:** n8n (self-hosted on a €20/month VPS)
- **Primary model:** Claude 3.5 Sonnet for drafting and review
- **Brief enrichment:** Perplexity API for real-time search context
- **Storage:** Notion database (because the client was already using it)
- **Review interface:** A custom Notion view with approve/reject buttons that trigger n8n webhooks
- **Distribution:** Buffer API for scheduling
Total monthly infrastructure cost for this client: €85. That's not a typo.
What we don't automate
The voice fingerprint update. Every quarter, we manually review what's performing and update the fingerprint. Automating this creates a feedback loop where the model optimises for metrics that diverge from the brand over time.
The outline review. The two minutes a human spends on this step saves an hour of downstream revision. The model is good at structure but doesn't know what the brand strategically wants to be known for. That's a human call.
The final edit. Not because the model can't do it—it can pass 90% of a copyedit. But 90% on content that carries the founder's name isn't good enough.
The number that matters
30 days of content: 8 long-form articles, 40 social posts, 8 email sections, 8 video scripts. Previously took 3 weeks of a full-time content manager's time. With this system: 2 days of a part-time editor's time, plus the infrastructure.
The goal was never to remove humans from the process. The goal was to remove the parts of the process that a human shouldn't have to do.