ls -la ~/notes/tools/
TOOLS
11 articles
cd ../An 88-Page Paper in Three Pages: Reading DeepSeek’s Cordis on Spatiotemporal Composability
The Cordis preprint from DeepSeek-affiliated authors is long, but its core fits two dimensions of spatiotemporal composability: revertible side effects and reactive dependencies. This piece critiques length and jargon, then explains the mechanisms, confluence, and validation limits in engineer language.
AI Harness as Function Composition: Orchestrating Systems Around a Frozen M(x)
A first-principles abstraction of AI Harness as function composition: freeze the foundation map M(x), orchestrate the outer chain H. Single-turn is strict multi-layer composition; multi-turn fuses composition with a state machine. Optimize H, not weights by default.
Customizing OpenSpec: Bridging Superpowers into Spec Workflows
Follow OpenSpec’s three customization levels, then use the community superpowers-bridge schema to show how Superpowers skills map onto an Artifact DAG—without forking either tool’s source—plus entry gates and anti-patterns.
The Nine Types of Claude Code Skills: From Personal Prompts to Org-Level Agent Capability
Anthropic's internal experience with hundreds of skills shows: a skill is not a Markdown prompt, it's an explorable capability package — verification skills should come before code-generation skills.
Claude Code vs. Codex: Comparing the Official Best Practices
Both official guides converge on the same operating model: context first, plan when ambiguity is high, verify with runnable checks, persist guidance, control permissions, and scale out only once reliable.
Reading the Coding Agent Market and Its Benchmarks (2026-06)
Artificial Analysis's Coding Agent Index combines DeepSWE, Terminal-Bench, and SWE-Atlas-QnA — a high Index score doesn't mean it fits your workflow; it needs to be read by task type.
Harness Engineering: A Systematic Approach to Steering Agents
A harness isn't a prompting trick — it's a control system made of Guides and Sensors that raises the odds of getting it right the first time, and enables self-correction before delivery.
The AI4SD Open Source Landscape: From SDD to the Outer Loop
The AI software delivery open-source ecosystem is organized into six layers: SDD specs, Harness/Skills, context orchestration, task decomposition, agent platforms, and Outer Loop CI/CD — pick a starting point based on what you mainly want to control.
LLM Cost and Deployment Fundamentals: The Model Gives You Capability, the Workflow Decides the Bill
The LLM bill = model unit price × token volume × workflow amplification factor — the same review feature can cost a few cents in a single-turn chat, or several dollars in an agent loop.
Browser Automation + MCP: How an Agent "Sees" the Frontend
Playwright + an MCP Server lets a coding agent actually "see" and "operate" a browser — one of the most underrated capabilities in the AI4SE toolchain.
The MCP Protocol: A Standard Interface for Agent Tool Chains
The Model Context Protocol is becoming the standard that connects agents to the real world. Understanding MCP's design philosophy is what lets you actually use it well — not just install a few MCP servers.