Man, I don't even know how I'd function without AI. So I made it work properly: specs instead of vibes, evidence instead of flattery, memory I own instead of memory a vendor owns.
Play LEDRING in your browser. You watch an AI builder make a web page and decide what gets through. The free built-in AI needs no key; the game is in English and 中文. Under the hood it is a toolkit that replays agent traces as branch trees and scores how far a run drifted from its instruction. Repo →
My agents run the repetitive parts of work and life. I stay in the loop where judgment matters: what to build, how to test it, and whether it actually works.
| Agent layer | What it takes off my plate |
|---|---|
| research-thinking-ai | Turns a vague research interest into evidence-tiered directions · live |
| n8n-workflow-collection | Lark/Feishu bot messages → Bitable records and PRD drafts; social metrics tracking |
| tech-news-digest (private) | Papers, GitHub Trending and Stack Overflow, ranked to my interests every morning |
| lakeboard-converter | Yuque Lakeboard / XMind → Markdown that agents can read |
| dev-work-grok-bot-skill | A Grok Bot skill that sets up a careful developer assistant: repos, tracker, team chat, preview tests, with a secret/PII checker |
| ledring | Status lights for my agents: replays many runs as one branch tree with divergence %, and scores how far a run drifted from its instruction. Claude Code + OpenTelemetry GenAI traces |
Sense → think → act. Built daily from GitHub Actions runs and committed eval results only; failing runs show as failing. Agents: read llms.txt or status.json.
A personal memory layer for agents that is portable (plain files, any model), private (stays on your disk), and correctable (you can see and fix what it believes about you).
Pulled from my public GitHub activity by a daily Action. Real events only.
| date (UTC) | repo | latest commit |
|---|---|---|
| 2026-10-15 | ledring | Make the zh play flow Chinese end to end (#91) |
| 2026-10-14 | research-thinking-ai | fix(watch): degrade gracefully without the private digest (#12) |
| 2026-10-14 | dev-work-grok-bot-skill | docs: README in house style with For agents section (#21) |
| 2026-10-14 | lakeboard-converter | fix: canonical GitHub link in web app (#8) |
| 2026-10-14 | n8n-workflow-collection | feat: agent-callable workflow manifests, n8n MCP docs and manifest check |
| 2026-09-04 | fitness-guide | ci: use pnpm in github actions |
How I build with agents
spec problem, user stories, out-of-scope; written before any code
slice tracer-bullet tickets, each demoable in one agent session
test red → green at agreed seams; LLMs and APIs behind adapters
review two axes: does it follow the standards, does it match the spec
Stack I reach for: TypeScript · React/Next.js · Node · Python · n8n · Lark Open Platform · MCP · Claude Code / Codex / Cursor. I learn by building and by arguing with docs, so most of what's here started as a practice project.
The banner scene rotates daily across focus, recovery and fun: deep work, dialogue, markets, recovery, live set, premiere. Nobody needs the same day twice.