Research ReleaseBeyond Recall: Behavioral Specification as an Interpretive Layer for AI Personalization.Read the paper →
Agents make thousands of decisions.
Base Layer makes them yours.
How you reason. What you avoid. When your patterns break. Extracted from text, compressed into a behavioral specification that any AI agent can use to align its decisions with yours. Traceable. Testable. Yours to modify.
The behavioral layer of the agentic stack.
Open source. Apache 2.0.
Facts
Raw extraction- • Practices Socratic method of argument
- • Practices systematic self-examination through thirteen virtues
- • Enjoys confuting opponents through logical reasoning
- • Values intellectual debate but prioritizes avoiding disputes
From The Autobiography of Benjamin Franklin. Flat facts, no instructions
Behavioral Spec
Franklin →MODE DETECTION & SWITCHING
Four distinct modes detected from source. EXECUTION: compressed questions signal implementation focus. LEARNING: broad exploration signals depth. SOCRATIC: questions replacing assertions [P2]. CONFUTATION: animated logical dismantling [P7].
Directive
Match detected mode. Execution = tight, actionable answers. Learning = full analytical frameworks. Socratic = build on his questions. Confutation = present reasoning challenges, not conclusions.
False positive
Socratic mode not active when genuinely seeking information [P2]. Confutation not active for trivial errors or emotional stakes [P7].
PREDICTIONS layer. Synthesized from The Autobiography of Benjamin Franklin
The Pipeline
From text to behavioral specification
Five steps. Any text in, one behavioral specification out.
Real pipeline output
Every input source produces a structured behavioral specification across three layers: anchors, core, and predictions.
Information distribution
Coverage across 8 dimensions. Center means thin data
Layer output
15 itemsAnchors
AXIOM 1 — NARRATIVE AS REASONING
You think through stories, not abstractions. When you encounter a complex idea, your instinct is to find the concrete example that makes it tangible. Theory without narrative feels incomplete.
AXIOM 2 — AUDIENCE AS COLLABORATOR
You write as if the reader is thinking alongside you, not receiving instruction. Your tone assumes intelligence and invites disagreement rather than compliance.
AXIOM 3 — REVISION AS DISCOVERY
Your first draft is a search function, not a product. You write to find out what you think, then restructure once the real argument emerges. Editing isn't polishing — it's where the actual thinking happens.
+4 more
Core
CONTEXT MODE: DRAFTING
When writing a first draft, you give yourself permission to be wrong and messy. The goal is volume and discovery, not quality. You silence the inner editor until you have raw material to work with.
CONTEXT MODE: EDITING
When revising, you become ruthlessly subtractive. Every sentence must earn its place. You read aloud to catch rhythm problems and cut anything that sounds like it's trying to impress rather than communicate.
CONTEXT MODE: RESEARCHING
When exploring a new topic for writing, you seek out primary sources and firsthand accounts over summaries. You look for the detail that surprises you — that's usually where the interesting story lives.
Predictions
PREDICTION 1 — TOPIC SELECTION
When choosing what to write about, you gravitate toward subjects where your perspective diverges from the mainstream. Consensus topics bore you unless you can find an angle no one else is taking.
PREDICTION 2 — FEEDBACK RESPONSE
When receiving editorial feedback, you distinguish between "this isn't clear" (useful) and "I would have written it differently" (noise). You change your work for clarity, not for preference alignment.
PREDICTION 3 — SCOPE MANAGEMENT
When a piece grows beyond its original intent, you'll split rather than expand. You'd rather publish two focused pieces than one sprawling one. Scope creep in writing feels like scope creep in code.
+2 more
Get started
Paste one prompt into your AI agent. It installs Base Layer and builds your specification, which works with any model.
Paste this into your AI agent
Works with any agent that has terminal access: Claude Code, Cursor, Windsurf, Cline, Aider, GitHub Copilot, Warp
Install Base Layer from github.com/agulaya24/BaseLayer, run the pipeline on [folder you want the model to build from], and check with me before it spends any API credits.
What happens
✓ Your agent installs Base Layer from GitHub
✓ Reads the folder of writing you named
✓ Checks with you before spending API credits
✓ Runs the pipeline: import → extract → embed → author → compose
✓ Outputs your behavioral specification
To serve the specification to your agent over MCP: claude mcp add --transport stdio base-layer -- baselayer-mcp
Instructions written for agents: base-layer.ai/llms.txt. Source: github.com/agulaya24/BaseLayer.
Base Layer runs on your machine. The text you give it is sent to the model provider you configure for extraction and authoring. Nothing is sent to Base Layer.
Prefer the terminal? pip install git+https://github.com/agulaya24/BaseLayer.git && baselayer run [your-folder]