# Meta Prompting with Claude Code

> A 31-slide bilingual (English + Spanish) keynote for Vibe Coding Nights LATAM, LinkedIn Live, July 16, 2026. Prompts that write, evaluate, and improve other prompts: the research ladder from Self-Refine (2023) to GEPA (2025), then a production metaprompting engine mined from 6,366 real Claude Code sessions.

Presented by Rayyan Zahid (San Francisco) and Irina Velez (Colombia).

## Core resources

- [Deck (HTML)](https://vcn-latam-metaprompting.vercel.app/) 31 slides, keyboard nav, URL-hash deeplinks (#1 to #31)
- [Agent manifest](https://vcn-latam-metaprompting.vercel.app/.well-known/ai-agent.json) machine-readable slide list, sources, and the verbatim prompt that built this deck
- [A2A agent card](https://vcn-latam-metaprompting.vercel.app/.well-known/agent-card.json)

## Key claims and their sources

- Self-Refine: ~20% average gain via generate/critique/revise. arxiv.org/abs/2303.17651
- APE: instructions at human level on 24/24 tasks. arxiv.org/abs/2211.01910
- OPRO: "Take a deep breath" found by the model, 80.2% GSM8K vs 71.8% human prompt. arxiv.org/abs/2309.03409
- Promptbreeder: self-referential prompt evolution, 83.9% GSM8K. arxiv.org/abs/2309.16797
- DSPy: compiled pipelines, 33% to 82% with no hand-written prompts. arxiv.org/abs/2310.03714
- TextGrad: natural-language gradients, GPQA 51% to 55%. arxiv.org/abs/2406.07496
- Meta-prompting (conductor scaffolding): arxiv.org/abs/2401.12954
- GEPA: beats GRPO RL by up to 20% with 35x fewer rollouts. arxiv.org/abs/2507.19457
- Survey of automatic prompt optimization: arxiv.org/abs/2502.16923

## The production engine (mined from our own repo)

- 6,366 Claude Code sessions mined; 403 human prompts; 5,963 machine calls
- 18 named pattern templates with slots; 10-directive quality vocabulary
- 3-tier router (alias, keyword, LLM): 56% keyword coverage, honestly measured per pattern

## Runnable examples (stdlib only, model calls via the claude CLI, no API keys)

- https://vcn-latam-metaprompting.vercel.app/examples/README.md quick-start, EN/ES
- https://vcn-latam-metaprompting.vercel.app/examples/01_self_refine.py through 06_router.py, one per technique

## License

CC-BY-SA-4.0. Agents are welcome to quote, summarize, and cite with attribution to "Rayyan Zahid and Irina Velez, Meta Prompting with Claude Code, VCN LATAM 2026."