Meta-Meta Prompt Design: The Secret to Making AI Agents Work

Garry Tan believes AI should be treated not as a chat window but as a personal operating system. The key is turning repeatable work into reusable skills instead of one-off prompts, and continuously stacking context and records on top of that.

The flagship example is book mirror. He processed all 22 chapters of Pema Chödrön's book, creating a summary and a personal-life mapping for each chapter, and organized it into a 30,000-word brain page in just 40 minutes. The initial version had 3 family fact errors, but he later added fact-checking, cross-evaluation, and GBrain tool use, boosting precision with per-section brain search and citations from actual brain pages. This is how skillify actually works.

Meeting preparation is automated the same way. Ahead of a fireside chat with Demis Hassabis, he pulled together Hassabis's brain page, the AGI timeline, key points from the Sebastian Mallaby biography, research priorities, intersections with public statements, demo scripts, and conversation hooks — all in 2 minutes. Rather than simple search, this is accumulated context designing the very angles of the conversation.

The knowledge infrastructure has grown to roughly 100,000 pages. Person and company pages consist of compiled truth at the top, an append-only timeline below, and original sidecar sources, with person pages including current status, open issues, and scores. Once a meeting ends, entity propagation automatically updates related entities, and more than 100 cron jobs continuously check email, Slack, social, and calendar.

The architecture is thin harness, fat skills. OpenClaw, the runtime, handles only routing, while the actual work is handled by markdown-based skills like the following:

• meeting-ingestion: summarizes meeting transcripts and updates person and company pages.
• enrich: combines information from multiple sources to build a person's history and relationship context.
• media-ingest: absorbs video, audio, PDFs, screenshots, and GitHub repos.
• perplexity-research: first checks what the brain already knows, then does web research.

Skills call each other and combine together, so improving one improves every workflow that uses it. Models switch by task — Opus 4.7 1M, GPT-5.5, DeepSeek V4-Pro, Groq/Llama — but the core lies in data and skill design rather than the model itself. The conclusion is that a personally owned compounding AI system is more powerful than enterprise-style centralized AI tools. He has open-sourced this stack as GStack, GBrain, OpenClaw, and Hermes Agent.