Patterns and Riffs

Two product concepts that coexist in every pipeline:

• Pattern — the reusable structure the AI extracts from an input (hook, tone, format, triggers).
• Riff — the new draft the AI generates grounded in one or more patterns, in your voice, without copying content.

How is a pattern decoded?

When a pipeline processes an input (post, URL, paste), the chosen LLM extracts four dimensions:

Dimension	What it captures	Example
hook	The opener's shape — what makes someone keep reading.	numbered-list + stake, contrarian opener, story-arc opener, question opener
tone	The author's voice.	confessional, operational, analytical, contrarian, narrative
format	The physical structure.	single post, 5-post thread, 2-post reply chain
triggers	Rhetorical / numeric elements that drive engagement.	dollar-amount, time-anchor, personal-loss, nobody-says-this

These fields are persisted in the structured artifact and form the basis for riffing later.

How does riffing work?

A riff takes a pattern + the operator's input (voice profile, specific instructions) and generates prose in the operator's voice while respecting the pattern's structure.

It is not a copy. The LLM has explicit instructions to:

• Keep the shape of the hook (numbered, contrarian, etc.).
• Keep the tone within the requested range.
• Change content completely — topics, examples, metrics.
• Use the voice profile (operator's linguistic traits) if available.

Genealogy and refinement

Every artifact preserves its genealogical tree via parent_id. When you trigger a refinement (shorter, change_tone, etc.), a child artifact is created with:

parent_artifact
  ├── child_a · action=shorter
  │     └── grandchild_a1 · action=change_tone (target=punchy)
  └── child_b · action=match_voice

You can navigate the entire tree from the UI or by requesting the artifact://{id} resource via MCP.

Multi-source riff (synthesis) — roadmap

:::caution Roadmap Today each riff takes one source post. The multi-source synthesis with citation markers described below is not implemented yet: in taxonomy.ts the multi-source combinations are marked as supported: false. :::

Future idea: when a pipeline has multi-source, the riff will combine patterns from multiple inputs and generate a draft with citations:

Input 1: @growth_dr post about pricing
Input 2: @bootstrap_fi post about 4 decisions
Input 3: blog URL about pricing anatomy

→ "Most founders obsess over their pricing page [3]. The teams that
   scaled fastest skipped them — until they couldn't [1] [2]…"

The [1] [2] [3] markers will reference each source's post_id / url. In the meantime, the only way to combine inputs is to trigger a brief with multiple source_post_ids (up to 50) from the Agent API, which synthesizes prose but without citation markers.

Voice model — roadmap

:::caution Roadmap There is no persistent voice profile yet. The operator's voice enters today as style_reference (free text, ≤4000 chars) that the agent passes explicitly on every call to the MCP tool start_draft or on every match_voice refinement. There is no Settings → Voice UI nor an API endpoint to save the profile. :::

The product intent is to support a voice profile per operator with:

• Samples of the operator's own text (3–10 paragraphs).
• Opt-in traits: do you use contractions? Swear words? First-person plural?
• Default preferred tones per pipeline.

Meanwhile, the recommended pattern is to keep your voice samples in a local file (or in an agent memory you have at hand) and pass them as style_reference when you trigger a draft.

Edit and publish

After generation, the operator can:

• Accept the draft as-is (publish_artifact).
• Edit it inline (override final_text).
• Trigger a refinement for a child variant.
• Archive it without publishing.

Any publication requires human confirmation — Whet does not auto-publish.

See also

• Pipelines — the container that defines which sources feed riffs.
• Operators — the role that operates refinement loops and publishing.
• MCP · refine_post — the prompt template that guides refinement via agents.