AI Ships Best When Requirements and Context Come First

Had a productive 02Ship AI gathering today — the core message: AI ships best when you treat "requirements + context" as first-class artifacts, not an afterthought.

What We Discussed (High-Signal Notes)

1) Write a req.txt That the AI Can Execute (Bob)

• Keep it simple but complete:
  • Goal / description
  • Tech stack
  • User stories
• Use ChatGPT prompts to optimize the requirement quality before coding starts
• Useful refs for workflow + structure:
  • CLAUDE.md (project rules / working agreement for the assistant)
  • obra/superpowers (prompting + tooling ideas)

2) Customer Calls to REQ to Implementation Pipeline (Michelle)

• Record the customer call
• Speech-to-text → save into notes
• Refine notes, then send into ChatGPT with a structured prompt to generate a proper REQ doc
• Send that REQ to Claude Code to implement
• Key takeaway: this is a repeatable "sales/PM → engineering" conversion flow

3) Learn a Codebase Faster (Leo)

• Fork a GitHub repo
• Ask the model questions to understand the code quickly
• Enforce consistent code style so AI-generated changes don't feel messy or mismatched

4) Multi-Agent + Self-Learning Direction (Bing)

• Looked at OpenClaw
• Interest in self-learning workflows and multi-agent approaches to scale execution

5) Context Is Still the Real Bottleneck (Jack)

• Context handling remains the hardest part
• Idea: aim for ~30% compacting (but avoid losing clarity/intent)
• Practical division of labor:
  • ChatGPT + Claude Code → requirements & planning
  • Codex → implementation/execution

My Takeaway

A practical "AI shipping stack" is emerging:

Record real customer intent → turn it into structured REQ → keep context compact but clear → use the right model for the right stage → ship with consistent code style.

If you're building with Claude Code / Codex / multi-agent workflows, DM me — keen to compare notes and jam on next steps.