Documentation Index Fetch the complete documentation index at: https://docs.ephor.ai/llms.txt
Use this file to discover all available pages before exploring further.
Optimising BrainLifts (Quick Take)
In a nutshell BrainLifts work like temporary brain upgrades for AI.
Too much content creates noise. Too little misses impact. The sweet spot? 500-1000 tokens of distinct, atomic insights that don’t tell stories.
Why It Matters Your BrainLift isn’t just providing information — it’s rewiring how AI thinks. Every token counts, but not equally. Get this right and AI becomes genuinely smarter in your domain. Example
Bad: 5000-token narrative about CloudFlare’s evolution from CDN to AI platformGood: 500 tokens of distinct SPOVs like “Durable Objects eliminate cold starts” and “CPU billing aligns with agent economics”
The shorter version performs 10x better. The Science Behind BrainLift Structure
How BrainLifts Work When you inject a BrainLift, you’re creating temporary weight updates in the model. Each piece competes for computational attention — and there’s only so much to go around.
The Critical Thresholds Research shows clear performance boundaries:
Peak Power Zone: 500-1000 tokens
Maximum influence on model behavior
Every insight has computational space
Optimal signal-to-noise ratio
Functional Range: 2000-2500 tokens
Still effective but diminishing returns
Some insights start blending together
Requires more careful curation
Danger Zone: 5000+ tokens
Performance collapse
Adding noise, not signal
Model can’t distinguish insights from background
Doubling your BrainLift size doesn’t double impact — it dilutes it.
The Counterintuitive Truth: Shuffled Beats Narrative Your instinct: Tell a coherent story
Research says: Don’t Why Stories Fail
Models perform worse with logical narratives
Similar content blends together invisibly
Model can’t distinguish insights from noise
What Works Instead
Distinct, Atomic Insights
Each SPOV stands completely alone: ✓ "Durable Objects eliminate cold starts for AI agents" ✓ "CPU-based billing aligns with agent economics" ✓ "JavaScript constraints accelerate time-to-market"
Semantic Orthogonality
Maximum distance between concepts — no overlap or similarity.
Shuffled Ordering
Random arrangement prevents narrative flow and improves processing.
What Fails: ✗ "CloudFlare started with CDN, evolved to Workers, which led to Durable Objects, enabling AI agents..."
Avoiding the Echo Chamber Trap Models are trained to please, not tell truth. Your BrainLift can amplify this problem or solve it. Signs of Echo Chamber BrainLifts
Only supporting evidence included
Confident, absolute language
No uncertainty or qualifications
Missing contradictory perspectives
Building Truth-Seeking BrainLifts
Include 'Except When' Clauses
“Durable Objects are ideal for state management, except when cross-agent queries are needed”
Document What You Don't Know
“Unclear: Performance impact at 10,000+ concurrent agents”
Reward Uncertainty
“Hypothesis: Edge computing may not suit ML training workloads”
Show Negative Cases
“Failed attempt: Company X tried edge-first and reverted to centralized”
The Curation Imperative Every addition comes at a cost. New content doesn’t just add — it competes with and potentially corrupts existing signals. Before Adding, Ask:
Is this semantically distinct? Similar content becomes noiseCan I remove something instead? Deletion often improves performanceDoes this deserve early placement? First tokens do 80% of workAm I telling a story or providing tools? Stories fail; tools succeed Better BrainLift = (Fewer SPOVs × Higher Distinctiveness) / Total Tokens
Practical Optimization Guidelines
Audit weekly — Remove more than you addTest in isolation — Does each SPOV stand alone?Shuffle insights — BrainLift should work in any orderCut connecting tissue — Remove anything bridging ideasMeasure tokens — Stay under 2500 absolute maximum
Adding every new article — Quality over quantityBuilding narrative bridges — Let insights stand aloneSimilar variations — One strong version beats three weak onesGrowing without editing — Brutal curation is essentialIgnoring token count — Size directly impacts performance The Bottom Line
Remember Your BrainLift isn’t a knowledge repository — it’s a precision instrument for temporarily rewiring AI behavior.
Research shows BrainLifts work best when:
Short (under 2500 tokens)Shuffled (no narrative flow)Sparse (maximum semantic distance)Specific (concrete examples, not abstractions) Optimization Checklist Before each BrainLift use: Why It Fits Into Ephor Optimization transforms your BrainLift from expensive noise into a powerful lens that genuinely shifts how AI models see your domain. Master these principles, and every conversation becomes more intelligent, grounded, and valuable. 
