Blog & Skills

Notes, opinions, and copy-paste-ready skills.

context-mode: stop your coding agent from re-reading everything it already saw

An MCP server + hooks setup that keeps raw tool output out of your context window — 315 KB to 5.4 KB — and rebuilds session memory from SQLite instead of a context dump.

Aug 6, 2026 · 2 min read

Terminal dashboard showing real-time token attribution with quota bar, burn rate, and per-source breakdown by tool, agent, skill, and MCP server

skill · curated

claude-token-lens: see exactly what's burning your Claude quota, live

A live terminal dashboard that attributes every token to a specific tool, skill, agent, or MCP server — so you know what to fix before you hit the wall.

Aug 3, 2026 · 2 min read

opencode terminal UI showing a coding session with the build agent active

skill · curated

opencode: the open-source terminal agent that isn't locked to one AI model

SST's opencode hit 190K GitHub stars by letting you run Claude, GPT-4, Gemini, or local Ollama through one terminal interface. Here's what's actually good—and what the 3.7k open issues signal.

Jul 30, 2026 · 2 min read

TencentDB Agent Memory's L0-to-L3 semantic pyramid architecture diagram showing how raw conversations get distilled into atomic facts, scenario blocks, and user personas

skill · curated

TencentDB Agent Memory: the four-tier memory pyramid that stops your AI agent from starting fresh every session

Tencent's open-source memory system builds L0-to-L3 structured long-term memory on local SQLite — no external vector DB, no flat conversation dumps, no API bills.

Jul 27, 2026 · 2 min read

Understand-Anything dashboard showing a codebase rendered as an interactive knowledge graph with color-coded architectural layers

skill · curated

Understand-Anything: map your codebase once, stop paying to re-read it every session

A plugin that runs a multi-agent analysis pass on your codebase, builds a persistent knowledge graph, then lets you explore, search, and query it from any AI coding agent — without burning tokens on raw file reads every time.

Jul 23, 2026 · 2 min read

skill

skill · curated

obra/superpowers: the coding-agent framework that makes your AI plan before it codes

Jesse Vincent's obra/superpowers enforces a seven-phase workflow — brainstorming through branch finishing — so your agent actually designs before it ships. 258k GitHub stars, MIT licensed, works across eight coding agents.

Jul 20, 2026 · 2 min read

skill · curated

caveman: why use many token when few token do trick

A one-file Claude Code skill with 90k GitHub stars that cuts output tokens by ~65% by making your AI agent talk like a caveman. Surprisingly effective; honest about where the savings actually come from.

Jul 16, 2026 · 2 min read

dev-browser header image — browser automation for Claude Code

skill · curated

dev-browser: Give Claude a Playwright-backed browser without handing it your filesystem

A QuickJS-sandboxed Claude Code skill that wraps the full Playwright API — so your agent can click, fill, and screenshot without touching the rest of your system.

Jul 13, 2026 · 2 min read

skill

skill · curated

agent-skills: 24 production-grade workflows that stop your AI from shipping shortcuts

Addy Osmani's 74k-star skill collection gives your coding agent the full senior-engineer lifecycle — spec to ship. Here's what it actually does right, and what it still requires of you.

Jul 9, 2026 · 2 min read

skill

AgenticAI: One Goal In, a Whole Delivery Company Out

A skill that turns a single goal into a self-checking delivery loop — a PM, a developer, and an adversarial QA that refuses to believe anything works without proof. The honest pros and cons, plus the file.

Jul 8, 2026 · 2 min read