How I Use Claude Fable 5 for AI UGC Research

Do not ask an AI model what is going viral in your niche.

It will give you a confident smoothie made from old trend reports, generic psychology, and things that sound true.

Give it the posts.

Give it comments, reviews, screenshots, dates, metrics, your own losing creative, and a job narrow enough to verify.

That is where Claude Fable 5 becomes useful for AI UGC research. Anthropic positions Fable 5 for long-running knowledge work and vision-heavy document analysis. Those capabilities fit creative research, where the input is messy and half the evidence is visual. Check Anthropic's current Fable 5 page for availability and product details because model access changes faster than this workflow.

The workflow below is less exciting than "find me viral ideas."

It also produces briefs I would actually let a team make.

The rule: evidence, interpretation, decision

Keep these as separate columns.

LayerExampleEvidence17 of 63 comments mention forgetting renewal datesInterpretationRenewal anxiety may be stronger than monthly price anxietyDecisionTest three renewal hooks against the existing subscription-audit format

Models blur these layers if you let them.

"People hate subscriptions" might be an interpretation.

"17 comments used the words forgot, surprise, or renewal" is inspectable evidence.

"Make a reaction video" is a decision.

Your research output should make it obvious which one you are reading.

Build a research pack, not a mega-prompt

I use a folder:

research/
  00-brief.md
  01-product-truth.md
  02-audience-language.csv
  03-competitor-posts.csv
  04-creative-screenshots/
  05-our-performance.csv
  06-claims-and-rights.md
  output/

00-brief.md

01-product-truth.md

02-audience-language.csv

One row per comment, review, support ticket, or interview fragment:

source_id,date,source_type,verbatim_text,product_or_topic,url
r-018,2026-07-02,app_store_review,"I forgot the annual renewal again",subscriptions,https://...

03-competitor-posts.csv

post_id,account,date,url,format,views,likes,comments,shares,notes
p-044,example,2026-07-10,https://...,slideshow,84000,3100,184,620,"renewal hook"

Do not invent missing metrics. Blank is better than fake precision.

04-creative-screenshots

Name files with their post ID and slide or frame:

p-044-slide-01.jpg
p-044-slide-02.jpg
p-051-frame-00.jpg

Fable can inspect images. The filenames make the visual evidence joinable to the row.

05-our-performance.csv

Use the actions you can actually measure:

06-claims-and-rights.md

What is licensed, who approved likeness use, what commercial claims are allowed, and which source assets are research-only.

This stops a competitor screenshot from accidentally becoming your published creative.

Pass 1: extract observations without strategy

The first model pass should be boring.

Prompt:

Read the attached research pack.

For each source row, extract only observations supported by that row
or its linked screenshot:

- audience problem
- exact phrase
- objection
- desired outcome
- emotional register
- format
- opening mechanic
- proof shown
- CTA

Return CSV with source_id on every row.

Do not recommend content.
Do not infer performance from a screenshot.
Do not fill missing fields.
Put uncertain observations in a separate column and explain why.

Why separate extraction?

Because if the model starts recommending while reading, it will pay more attention to evidence that supports its first clever idea.

Extraction builds a ledger. Strategy comes later.

Pass 2: cluster audience language

Now group the observations.

Prompt:

Using only the extracted audience-language rows:

1. Cluster semantically similar problems.
2. Count distinct source rows in each cluster.
3. Keep the five strongest verbatim phrases per cluster.
4. Separate frequency from intensity.
5. Identify contradictions.
6. Cite source_id for every phrase and conclusion.

Do not merge two clusters merely because the same product could solve both.

Frequency and intensity are different.

Twenty people saying "I forget renewals" is frequent.

Two people saying "this caused an overdraft before rent" is intense.

Both may deserve a test. Do not let a count erase the pain of a smaller segment.

Example output:

ClusterDistinct sourcesStrong phraseTensionSurprise renewals17"I forgot the annual renewal again"Low attention, sudden chargeTiny recurring charges11"None of them are expensive alone"Individually harmless, collectively painfulShared subscriptions6"We both thought the other cancelled it"Ownership ambiguity

That table is already more useful than "money-saving content performs well."

Pass 3: tear down the creative

Do not ask why a post went viral. You rarely have enough evidence to know.

Ask what the post did.

Here is a first-party slideshow frame:

A visual teardown should identify:

Prompt:

Analyze each screenshot set as a sequence.

For every post:
- transcribe visible text
- describe the visual literally
- identify the hook mechanism
- identify the promise made by slide/frame 1
- show how later slides/frames pay it off
- identify the proof object
- identify the CTA
- cite post_id and frame filename

Separate literal observation from interpretation.
Do not call anything viral unless the supplied metrics support that label.

"Pretty beach image" is description.

"The ordinary-life confession over an aspirational background creates tension" is interpretation.

Keep both. Label them.

Pass 4: find repeated structures

One high-performing post can be luck.

A creator repeating the same structure is more interesting.

Ask:

Group competitor posts by reusable structure.

A structure includes:
- opening mechanic
- sequence
- proof object
- CTA pattern

For each structure:
- list every supporting post_id
- show how surface topics changed
- show which elements stayed fixed
- report median supplied performance only when at least three rows
  have the required metric
- flag outliers instead of letting them dominate

Do not recommend copying a creator's wording or assets.

You are looking for molds:

The words and proof should be yours.

Pass 5: compare the market to your own account

This is where most competitor research goes wrong.

The model finds a structure somebody else uses and recommends it because it exists.

Existence is not a reason.

Compare it against:

Prompt:

For each repeated market structure, score:

- audience-language fit
- product-proof fit
- novelty relative to our last 30 posts
- rights feasibility
- production feasibility
- claim risk

Use High / Medium / Low with one cited reason.

Then recommend:
- test now
- research more
- reject

Every test-now recommendation must name the audience cluster,
proof object, and one variable to change.

This blocks pretty competitor content from becoming a useless brief.

Pass 6: write hypotheses, not ideas

Idea:

Make a slideshow about subscriptions.

Hypothesis:

A first-slide hook naming annual renewals will earn more saves than a generic subscription-saving hook because 17 audience sources describe forgetting the renewal date. Hold the six-slide audit format and avatar constant.

Now the content team knows:

Give the model this schema:

hypothesis_id
audience_cluster
evidence_source_ids
format
hook
proof_object
single_primary_variable
control_or_comparison
intended_action
success_signal
risk_or_unknown

If a field is missing, the idea is not ready.

Pass 7: convert the hypothesis into a Ghostfeed brief

Example:

Hypothesis: H-07
Audience: people who forget annual subscription renewals
Format: six-slide story/list hybrid
Avatar: approved budgeting persona

Slide 1:
"the annual renewals making your monthly budget lie"

Slide 2:
"monthly budgets hide anything that charges once a year"

Slides 3-5:
three concrete places to check, each supported by product behavior

Slide 6:
show the renewal view and ask the viewer to check before the next billing date

Visual direction:
ordinary home-desk and phone moments, warm natural light, no luxury-finance imagery

Do not claim:
savings amount, cancellation success, or guaranteed detection

Intended action:
save

That can go directly into a slideshow or reaction production workflow. In this setup Claude has Ghostfeed MCP connected, so the brief should also name:

For a reaction, Claude searches workspace templates before the inspiration library. If it imports a 30–120 second source, it returns the dashboard link so I can choose a crop or run Smart Crop. It renders the requested first frames and stops. Only after I approve a frame does it use clone or prompt mode.

You: Execute brief H-17 in Launch Lab. Search Ghostfeed for the
specified medium-close opening and cast Stella, Gia, and Ethan.
Frames only.

Claude: I found one owned template and one inspiration candidate. The
owned template matches the gaze and hand position more closely. I will
render three first frames and keep video generation paused.

Research has now terminated in an asset request.

The daily review

Do not upload yesterday's metrics and ask the model to "analyze performance."

Ask narrow questions:

Prompt:

Join yesterday's post rows to the hypothesis table.

For each hypothesis:
- compare the intended signal to its named control or baseline
- report sample size
- distinguish measured result from interpretation
- recommend repeat, revise, or stop
- propose one next variable only

Do not declare a winner from one post.
Do not use views as the success metric when the intended action was saves or clicks.

This gives you a production brief instead of an analytics horoscope.

What Fable 5 is good at here

In my use, the valuable capabilities are:

Anthropic describes Fable 5 as a model for ambitious long-running knowledge work and vision analysis. That supports the shape of this workflow. It does not prove any marketing conclusion for you.

The audience still gets the final vote.

What not to upload

Do not paste raw customer data because a model has a large context window.

Remove:

Anthropic's Fable 5 page currently states that using the model requires 30-day data retention for safety monitoring. Check the current policy and your organization's agreement before uploading client material.

The safest research pack contains anonymized phrases and the minimum evidence needed for the decision.

The no-slop test

Reject the research output when it contains:

The model should reduce the pile without hiding the uncertainty.

That is the standard.

Claude Fable 5 can process more of the research loop in one coherent job. Ghostfeed can turn the resulting brief into reactions, avatars, and slideshows. Neither one supplies the customer truth.

You still have to collect it.