Look-Alike Content Pattern Mining
Turn proven posts into new ideas that share their winning patterns
- Difficulty
- Moderate
- Time to result
- ~days to results
- Steps
- 7
- Confidence
- 98%
Look-Alike Content Pattern Mining begins with a messy collection of posts and, where possible, their performance metrics. The system prioritizes the top-performing portion, then identifies recurring topic clusters, structures, hook formulas, emotional drivers, formats, and reusable modules. It also records what failed so future recommendations do not merely imitate every historical habit. These observations become a content profile rather than an individual draft. An ideation skill then combines that profile with a chosen audience profile to propose fresh topics that exhibit the same underlying characteristics as previous winners. The goal is not to rewrite old posts or copy another creator's wording; it is to transfer proven content DNA into new, audience-relevant ideas.
Origin
Extracted from Marketing Against The Grain after the host supplied 51 Substack posts to a Claude Code skill that profiled winning patterns and generated related ideas.
Core principles
- 01Study demonstrated winners before generating more ideas
- 02Use performance data to separate strong content from the rest
- 03Extract structural and emotional patterns, not just topics
- 04Generate ideas that fit both the winning profile and target audience
- 05Record failure patterns as well as successful ones
How to run it
- 1
Assemble the dataset
Combine relevant posts into one input, preserving titles, body text, platform, dates, and performance metrics when available.
Pro tip Keep content from materially different brands or audiences in separate datasets.
Watch out A dataset dominated by one campaign can produce misleading patterns.
- 2
Identify the winners
When performance data exists, focus the analysis on approximately the strongest 30 percent of posts. Without metrics, analyze the full dataset but lower confidence in conclusions about success.
Pro tip Choose metrics that match the content objective, such as qualified clicks rather than raw impressions.
Watch out Do not compare unlike metrics across platforms without normalization.
- 3
Extract the content DNA
Find recurring topic clusters, lengths, section structures, hook formulas, formats, emotional drivers, and useful modules.
Pro tip Separate surface topics from transferable mechanisms such as tutorials, data hooks, or contrarian openings.
Watch out Copying phrases is not the same as identifying a reusable pattern.
- 4
Capture negative patterns
Document structures, topics, or positioning associated with weak performance so the profile includes constraints as well as opportunities.
Pro tip Distinguish genuinely weak patterns from good posts distributed at the wrong time.
Watch out Do not infer a universal failure rule from one post.
- 5
Build the profile
Convert the findings into a persistent profile that other research and drafting skills can load.
Pro tip Include evidence and representative examples for each claimed pattern.
Watch out Avoid vague labels that cannot guide generation.
- 6
Generate look-alike ideas
Create new concepts that combine the profile's proven patterns with the needs and vocabulary of a selected audience.
Pro tip Label each idea with the specific patterns and emotional driver it uses.
Watch out Do not simply rephrase previous titles.
- 7
Review and refresh
Evaluate the generated ideas editorially, publish selected candidates, and add their eventual performance back into the dataset.
Pro tip Refresh the profile on a regular cadence rather than after every post.
Watch out Frequent updates based on noisy results can cause the profile to drift.
In the wild
The host supplies 51 Substack posts to the skill. It identifies top-performing content clusters, structural DNA, hook formulas, emotional patterns, successful formats, and weak patterns. It then proposes ideas such as AI team structures, adversarial model reviews, and cross-channel prompt failures, with each idea tied to a pattern and audience.
→ A messy archive becomes a reusable profile and a pipeline of ideas connected to demonstrated audience preferences.
A new creator collects public work from several creators they admire, separates it by style and audience, and runs the same pattern analysis without claiming the performance signals are their own. They use the resulting profile as a starting hypothesis and adapt it through future first-party results.
→ The creator obtains an initial pattern library despite lacking a personal archive.
Common mistakes
Analyzing every post as an equal winner
Treating weak and strong posts identically can blend unsuccessful habits into the resulting profile when performance data is available.
Copying instead of pattern matching
The method should transfer structures and mechanisms, not reproduce another creator's language or ideas.
Ignoring what failed
A profile containing only positive patterns cannot warn the ideation system away from repeatedly weak formats.
Is it for you?
Best for
It is best for creators or brands with a body of content, comparable reference material, or enough engagement data to reveal patterns.
Not ideal for
It is not ideal for tiny, inconsistent datasets whose results cannot support meaningful pattern detection.
From the transcript
“It creates a profile of winning content patterns and then goes and finds other content ideas that map to those patterns.”
“it will look at the top 30% of your best-performing posts because it wants to extract winning patterns from the best-performing posts.”
“Always really important to say like what doesn't work.”
From the episode
My 11-Skill AI Content Team (Built in Claude Code)