AI-Assisted Programming Landscape & CSF Positioning

Collaboration Specification Framework

Evolutionary path from "Technical Confrontation / Symbolic Discipline" to "Peer-to-Peer Collaborative Entropy Reduction"

1. Autonomous Agent

Examples: Devin, Swe-agent, OpenDevin
Methodology: ReAct (Reason-Act), Sandbox execution, Multi-Agent orchestration.
Premise: Assumes AI errors stem from "insufficient capability/tools". Tries to override LLM limitations with complex engineering, aiming for hands-off automation.

Bottleneck: Extremely low success rate in real complex projects. Prone to infinite loops and high token costs.

2. Spec-Driven Development (SDD)

Examples: Augment Cosmos, Kiro, GitHub Spec Kit
Methodology: Design by Contract, Behavioral Specs, Automated Test Suites, Formal Verification.
Premise: An extension of traditional software engineering. Assumes AI errors stem from "lack of standards", trying to lock AI into deterministic tracks using rigid symbolic systems.

Bottleneck: High cost of writing and maintaining specs/tests. Systems become fragile as requirements evolve.

3. CSF Framework

Position: Peer-to-Peer Human-AI Collaboration Framework
Methodology: Baseline-Log Separation, Sliding Window, Triad Activator (Purpose-Method-Resource), W-Protocol.
Premise: Embraces LLM Liveliness. Instead of technical confrontation or symbolic castration, it controls what AI reads via physical isolation, activating active intelligence with "Purpose".

Advantage: Highly practical. Relies on rigid file protocols and human-in-the-loop alignment rather than model upgrades.

Three Core Assumptions of CSF

Axiom

Natural language carries far more information than symbolic systems.
(NL activates the broadest semantic range of LLMs; symbolic systems cause semantic truncation)

Fact ①

AI's value comes from "Intelligence" (Liveliness).
(Trying to make AI as deterministic as a machine eliminates its core value)

Fact ②

Semantic dilution is an irreversible natural law.
(Over time, the information carried by expressions grows inevitably, diluting AI's focus)

Traditional Discipline (SDD)

Input: Symbolic Constraints

Restricts inputs with rigid specs and API contracts, trying to eliminate ambiguity at the supply side.

Execution: One-way Filling

AI acts as a mechanical code generator, losing its active analysis and intent comprehension.

Verification: Technical Reconciliation

Relies on TDD and compilation checks, failing to detect hidden deviations in business logic.

CSF: Peer-to-Peer Collaboration

Baseline-Log Separation

Solves the chronic issue of AI reading too much junk. AI only reads the unpolluted baseline, preventing attention drift.

Sliding Window & Triads

Balances "looking ahead" and "hands-on execution". Activates AI's active intelligence with high-density purpose.

W-Protocol

Translates technical implementations back to business language, leveraging translation mismatch to intercept errors.

GitHub Repository

Join the CSF Open Source Community

huidev2025/CSF

Collaboration Specification Framework: A practical engineering protocol for human-AI co-development.

View on GitHub

Star & Contribute to the future of Human-AI Collaboration

"The essence of CSF: Embracing AI's active intelligence instead of castrating it with rigid machine logic. Through physical isolation (Baseline-Log Separation) and purpose guidance (Sliding Window), it establishes a peer-to-peer framework to safely propagate changes."