Format-and-Style Training Stack
Train AI on an asset’s structure and voice before asking it to create
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 6
- Confidence
- 97%
The framework trains an AI system with two complementary instruction sets. The first is an asset-format guide that explains the function of each component, such as an opening scene, internal conflict, turning point, actions, outcome, reflection, and lesson. The second is a platform-appropriate writing-style guide derived from examples of a creator or brand, with principles and detailed replication rules. The user supplies both guides alongside a topic, platform, length, and exclusions such as no emojis or hashtags. Separating structure from style makes each component reusable and allows formats and voices to be mixed deliberately. Strong examples and precise instructions improve the initial draft, while platform-specific guides reduce mismatches such as applying a philosophical X voice to a conventional LinkedIn post.
Origin
Extracted from Marketing Against The Grain, where Kieran Flanagan demonstrated Claude training documents built from successful posts and creator-specific writing samples.
Core principles
- 01Separate content structure from writing style
- 02Derive instructions from proven examples
- 03Build platform-specific training documents
- 04Give the model line-level context
- 05Treat generated drafts as starting points
How to run it
- 1
Define the asset and platform
Choose a narrow target such as a LinkedIn personal-achievement post, an X thread, a paid ad, or a podcast episode. Specify the intended length, audience, and outcome.
Pro tip Start with short-form content because its structure is easier to observe and reproduce.
Watch out Do not assume one template or style transfers cleanly across platforms.
- 2
Collect proven examples
Gather multiple high-performing examples of the target format and, separately, examples of the desired creator or brand voice. Use examples representative of what you genuinely want to publish.
Pro tip Segment creator examples by platform so the model learns the relevant version of the voice.
Watch out Weak or inconsistent examples will produce weak or contradictory instructions.
- 3
Extract the format guide
Ask the AI to explain the purpose and required content of each sentence or section in the examples. Convert the analysis into a repeatable template with an ordered sequence.
Pro tip Request active verbs, vivid details, transitions, and the intended reader effect for each component.
Watch out A vague outline will not provide enough context to outperform a generic prompt.
- 4
Extract the style guide
Feed examples iteratively and instruct the AI to acknowledge each one until you say finish. Then have it produce key principles and detailed guidelines another assistant can follow.
Pro tip Use a continue-until-finish protocol so the model analyzes the full sample rather than drafting after the first example.
Watch out Do not confuse recognizable verbal tics with the deeper principles that make the writing effective.
- 5
Generate from both guides
Provide the topic, format guide, style guide, word range, and any explicit exclusions in one creation brief. Ask the AI to follow both uploaded documents.
Pro tip Use a bounded word range rather than only a generous maximum to control draft length.
Watch out Mixing a style guide from one platform with a format from another can create tonal or structural friction.
- 6
Refine the instruction set
Inspect where the output missed the format, voice, or desired level of punchiness. Revise the guides or add a targeted instruction, then regenerate only the weak section when possible.
Pro tip Ask for precise transformations such as condensing a conclusion into two thought-provoking lines.
Watch out Repeatedly adding broad prompts can cause the model to abandon earlier structural requirements.
In the wild
Kieran combined a sentence-level LinkedIn personal-achievement template with a writing guide derived from Naval Ravikant’s X posts. Claude generated a post arguing that AI amplifies rather than destroys marketing creativity. After a request for punchier language and a tighter ending, the draft produced several lines the hosts considered publishable.
→ A strong LinkedIn draft was created in minutes after the reusable preparation work had been completed.
A marketing team collects ten effective problem-solution ads and ten examples of its brand voice. It extracts a scene-by-scene ad template and a separate voice guide, then combines them with each new product brief rather than asking for an ad from scratch.
→ The team produces more consistent first drafts while retaining control over claims, positioning, and final language.
Common mistakes
Using a two-sentence prompt
Requests such as writing a funny LinkedIn post omit the line-level context needed for a distinctive result. The model must infer too many structural and stylistic decisions.
Ignoring platform differences
A voice or content type that performs on X may not work on LinkedIn. Build platform-specific templates and style guides instead of treating social channels as interchangeable.
Publishing the first draft
Even a well-trained model can produce awkward openings, excessive length, or generic conclusions. Human editing remains necessary to make the result authentic and publication-ready.
Is it for you?
Best for
It is best for marketers repeatedly producing short-form posts, ads, newsletters, videos, or podcast formats.
Not ideal for
It is not ideal for one-off requests where collecting examples and preparing reusable training documents would cost more time than writing directly.
From the transcript
“you have to train it to figure out what is the right format of the thing and then train it to create the you know…”
“the platform the style those two things can be used for pretty much most marketing assets”
“it's unbelievably fast once you've done the prep”
From the episode
Write Viral LinkedIn & X Posts With Claude 3.5 Sonnet (Tutorial)