Performance-Grounded Context Refresh
Use output performance to continuously improve the shared AI context layer.
- Difficulty
- Moderate
- Time to result
- ~months to results
- Steps
- 6
- Confidence
- 98%
Performance-Grounded Context Refresh is a feedback loop for improving the foundational layer with evidence from real outputs. At the end of a review period, the team supplies the AI system with what it produced and the corresponding performance results. The system compares strong and weak outputs, identifies repeatable patterns that may explain the difference, and proposes updates to the relevant audience, style, positioning, or journey files. Because many skills consult those files, the learning propagates across the system rather than remaining inside one corrected prompt. The framework should operate on meaningful groups of outputs and be repeated periodically, such as every quarter. Its purpose is not merely to imitate winners but to refine the shared assumptions that guide future work.
Origin
Extracted from Marketing Against The Grain, where the host recommends quarterly foundational updates and describes using content performance data to reinforce the patterns associated with stronger results.
Core principles
- 01Shared context should evolve as new performance evidence appears.
- 02Patterns across outputs matter more than reactions to one isolated result.
- 03Improve the foundation so successful learning propagates to every dependent skill.
- 04Preserve more of what performs and remove patterns associated with weak results.
- 05Regular refreshes prevent context from becoming stale.
How to run it
- 1
Assemble the evidence
Collect the outputs from a defined period along with consistent performance measures and relevant context.
Pro tip Use comparable metrics for outputs serving similar goals.
Watch out Combining incomparable formats or objectives can create false patterns.
- 2
Classify results
Separate clearly strong, average, and weak performers without relying on a single vanity metric.
Pro tip Include qualitative signals such as replies or sales conversations when they reflect the intended outcome.
Watch out A high-impression outlier may not represent effective marketing.
- 3
Find recurring patterns
Compare groups to identify repeated differences in audience insight, style, positioning, journey fit, or other foundational context.
Pro tip Require a pattern to appear across multiple outputs before promoting it into the foundation.
Watch out Do not overfit the system to one exceptional success or failure.
- 4
Update the correct files
Translate validated patterns into concise changes within the foundational files responsible for those decisions.
Pro tip Keep each update in the file with the clearest ownership of that context.
Watch out Adding the same lesson everywhere creates overlap and contradictory guidance.
- 5
Retest dependent skills
Generate representative outputs with the revised foundation and check whether the changes improve relevance without causing regressions.
Pro tip Compare old and revised outputs on the same briefs where practical.
Watch out A context update can improve one workflow while unintentionally degrading another.
- 6
Repeat periodically
Run the cycle on a regular cadence and refresh outdated market or audience assumptions even when performance is stable.
Pro tip Use quarterly reviews as a starting cadence and adjust to the rate of change.
Watch out Updating after every small fluctuation creates a noisy, unstable foundation.
In the wild
A team gives Claude its quarter's social posts, newsletters, and performance results. The analysis finds that concrete workspace screenshots and audience-native phrases repeatedly outperform generic productivity advice, so those patterns are strengthened in the audience and style files.
→ Every dependent content skill is more likely to reproduce the successful contextual patterns next quarter.
A company finds that content centered on a common AI claim draws impressions but few qualified conversions, while cross-functional workflow stories produce stronger demos. It updates the positioning map to downgrade the contested claim and strengthen the evidenced alternative.
→ Future skills emphasize a position correlated with meaningful customer action.
Common mistakes
Overfitting one winner
A single high-performing output may reflect timing, distribution, or chance rather than a reusable contextual pattern.
Updating the wrong layer
Changing individual skills when the lesson belongs in shared context prevents the improvement from propagating.
Optimizing vanity metrics
Performance data must reflect the intended business or audience outcome rather than superficial reach alone.
Is it for you?
Best for
It is best for teams producing enough measurable AI-assisted output to detect repeatable patterns over time.
Not ideal for
It is not ideal when outputs lack reliable performance measures or when a tiny sample would encourage overfitting.
From the transcript
“And you update those MD files every quarter.”
“Based upon this, how can I improve all of the files on my foundational layer to get better performance?”
“Update my foundational layer with the patterns that would get more of the great stuff.”
From the episode
The Real Reason Your AI Content Is Average (It's Not Your Prompts)