The Brand Voice Engine Playbook
How B2B operators build a brand voice AI agent that actually sounds like them.
Most “brand voice GPTs” are system prompts pasted into a custom GPT. The output is your content, flattened. This Playbook walks through the three-layer system we run at Flywheel to keep our own content recognizable, plus the two-week build plan you can run yourself.
What’s inside the Playbook
The three-layer Voice Engine framework
Inputs (the library that goes in before the engine), the engine itself (the five-layer system prompt), outputs (scoring, human-in-the-loop review, feedback loop). Skipping any layer produces AI slop.
Inside the engine: POV, voice, tone, lexicon, guardrails
What goes in each layer, what to write down explicitly, and the rules that separate human-readable cadence from the LinkedIn-AI fingerprint.
The output loop that makes the agent compound
The scoring rubric, the human-in-the-loop workflow, and the feedback mechanism that funnels rejected drafts back into the input library — plus the ongoing improvements the agent pulls automatically from new inputs each week.
The four failure modes we see most often
The four common failures are treating the system prompt as the whole agent, skipping call transcripts, running without guardrails, and shipping without a feedback loop. Each one comes with the specific fix we learned in production.
The Flywheel production system as a worked example
What is in our 310-line system prompt, what the input library looks like, what the scoring rubric checks, and what we have spent running it.
A two-week build plan you can run yourself
Day-by-day build sequence for weeks one and two, plus the ongoing maintenance cadence to keep the feedback loop alive.
Get the Playbook
The three-layer framework, the five-layer engine, and the two-week build plan that turns your custom GPT into a brand voice agent that compounds.
We use your work email to send the Playbook and the occasional Flywheel update. No spam, no list reselling, unsubscribe in one click. Privacy policy.
The worked example
We run our own brand voice agent in production.
Every piece of content the Flywheel agent swarm ships goes through a 310-line brand voice system prompt, an input library of roughly 200 published pieces and 100 call transcripts, a scoring rubric that runs on every draft, and a human-in-the-loop review before anything publishes.
310 lines
System prompt
~$0.04
Per LinkedIn draft
5 LI + 1 blog
Shipped each week
~75¢ / wk
Total runtime cost
Frequently asked questions
What is a brand voice AI agent?
A system that produces content matching a specific company's voice at scale. It has three parts: an input library (existing content, call transcripts, customer language, rejected drafts), a five-layer system prompt engine (POV, voice, tone, lexicon, guardrails), and an output loop (scoring rubric, human-in-the-loop review, feedback mechanism). It is different from a custom GPT in that it is version-controlled infrastructure, not a per-user chat tool.
How is a brand voice agent different from a custom GPT?
A custom GPT is a chat interface with a system prompt. A brand voice agent is versioned infrastructure with structured inputs, an explicit architecture, an automated scoring rubric, a human-in-the-loop workflow, and a feedback loop that updates the system over time. Custom GPTs drift; brand voice agents compound.
How long does it take to build a brand voice agent?
About two weeks of focused work to ship version one, then a few hours per week to maintain the input library and feedback loop. The agent gets meaningfully sharper in the first three months as rejection patterns accumulate.
What does a brand voice agent cost to run?
Our Flywheel system runs at roughly four cents per LinkedIn draft and 25 cents per blog draft on Claude. Total weekly runtime cost for our output volume is about 75 cents.
Can a non-developer build one?
Yes. We built ours without writing production code by hand. The engine lives as structured markdown, the inputs are a folder of text files, and the dispatcher is built with Claude Code.
Already running content through a custom GPT?
If you can already tell the output is sliding toward generic, book a Phase 0 discovery call. We will walk through the live Flywheel system on the call and map what it would look like inside your stack.
Book a Phase 0 discovery call