All articles

Extraction Templates for Content Structure

Project: ViziVibes Link: https://vizivibes.com/studio

July 1, 20265 min read

Extraction Templates for Content Structure

Project: ViziVibes
Link: https://vizivibes.com/studio

Case study type: Feature design
The task: Shape raw input into reliable structures (timeline, comparison, pros/cons, data table, playbook) before any visual style runs, and let users save and share custom templates.
What we learned: Data is sacred only when structure is explicit and reusable, not when you hope a style prompt preserves the facts.
Last updated: June 2026


Case study at a glance

The taskProtect numbers, comparisons, and timelines through extraction, then hand clean structure to the design layer
Who it was forEducators, analysts, founders, and marketers burned by AI that prettifies away the facts
Main constraintGeneric summarization breaks tables, drops columns, and rounds numbers users need exact
What we builtBuilt-in extraction modes plus custom templates users can save, share, and discover in Browse
OutcomeStructure survives from source to visual; users reuse templates instead of re-prompting every week

Background

Early on I loved how fast AI visuals were. I trusted them less after a comparison chart dropped a column and a timeline turned into vague bullets. The design looked polished. The data did not.

The issue was rarely the image step. It was what we fed it: unstructured prose where structured data should have been. We wrote a rule inside ViziVibes: data is sacred. The fix was not a longer style prompt. It was extraction templates that run before any pixels.


The task

Ship built-in extraction patterns for common jobs, let power users define custom templates with named instructions (and optional schemas), and surface public templates in Browse so structure becomes shared infrastructure.


Constraints

  • Templates must differ by source type (URL vs file vs web research vs video).
  • Custom templates cannot require engineering work to create.
  • Public sharing must not leak private user content, only the template definition.
  • Extraction cost and generation cost must stay transparent in credits.

Our approach

We split the pipeline into extract, structure, design, generate. Style, palette, and format never run on raw chaos. They run on template output.

Built-in modes cover the common cases. Custom templates cover niche workflows. Public templates turn one user's discipline into everyone else's shortcut.


How we solved it

Step 1: Map modes to real jobs

What we did: Shipped built-in patterns including timeline, comparison, pros/cons, data table, playbook, summary, key facts, and brand identity cues from URLs.

Decision: Name modes after outcomes users recognize, not internal model jargon.

Why: "Comparison" means something in a marketer's head. "Structured JSON pass" does not.

Step 2: Wire modes per source path

What we did: Limited available modes by source (web research gets statistical and explainer patterns; files get outline and action items; videos get chapters).

Decision: Curate the menu per input instead of one infinite list.

Why: Wrong mode on wrong source produces garbage fast. Curation is UX, not limitation.

Step 3: Custom templates with save and share

What we did: Let users save named instructions (and schemas where needed) as reusable templates attached to their account.

Decision: Templates are first-class objects, not buried prompt history.

Why: Weekly workflows should be one click, not a hunt through last Tuesday's project.

Step 4: Public templates in Browse

What we did: Allowed users to publish templates to the community gallery with search, sort, and save.

Decision: Treat templates like styles and palettes: discoverable creative assets.

Why: A teacher's timeline template helps the next teacher on day one. Community scale beats solo reinvention.

Step 5: Hand structured output to generation unchanged

What we did: Passed template output into the generation prompt as structured content blocks, with brand kits and formats layered on top.

Decision: Forbid silent re-summarization between extraction and image generation.

Why: Every extra summarization step is a chance to drop a number. One structured handoff preserves trust.


What we built

  • Built-in extraction modes across text, files, URLs, video, and web research
  • Custom template creator with save and optional schema
  • Public template publishing and discovery in Browse
  • Structured extraction records that stack in multi-source projects
  • Clear credit surfacing before extract and generate

Results

Before: I got fast visuals with occasional factual drift. I re-prompted manually every session.

After: I pick the mode first, check the extraction, then generate. Numbers and comparisons survive the trip.

What changed for us: I stop blaming the image step when the structure step was wrong. If facts matter in your posts, structure is not optional.

How we know it worked: Template save rate and Browse applies climb among retained users. Structure became habit, not a one-time fix.


What you can learn

  1. Structure is a product layer, not a prompt trick. If facts matter, formalize extraction before pixels.
  2. Name modes for user jobs. Vocabulary alignment reduces wrong-mode errors.
  3. Reusable templates beat heroic prompting. Encode discipline once.
  4. Community templates compound value. Shared structure is as important as shared style.
  5. Minimize handoffs that re-summarize. Every hop risks data loss.

Next step

Run the same URL twice in the studio: once as a data table, once as pros/cons. Compare extraction output before you generate. Then save your favorite as a custom template or grab one from Browse.

For how structure meets brand rules, read Keeping Brand Consistency.


Related files in this article package

FilePurpose
linkedin.mdLinkedIn post (under 2,500 characters)
audience-brief.mdReader, intent, and KPI framing
optimization-sheet.mdTitle variants, meta, internal links
images/README.mdSuggested diagrams and screenshots

+ Keep reading

More from the Island Tech Stack field notes.

All articles