Blind-generated user stories filtered by a TASTES.md file, agent decides what to build
generativist · x · 2026-10-21
Developer generativist shared an experiment for his nephew's game: the system generates semi-constrained user stories (feature requests and bug reports) blindly, then an agent decides whether each is valid and worth doing, subject to a TASTES.md taste file.
- Screenshot shows the loop in action: baseline checks pass, then iterations tackle concerns like "pausing must freeze physics, timers, music, animations" and "the camera should never hide upcoming hazards," each paired with a story-flavored ticket (e.g. "Paused for Pizza, Came Back to a Crime Scene").
- The agent attempts implementation, runs automated checks, and logs cost per decision (e.g. kept for $0.87).
The interesting idea: delegate requirement triage itself to the agent, encoding preferences in a plain markdown taste file instead of manually curating the backlog.
More from coding & agent
- Meta trains AI agents inside a 'Truman Show' simulation of 25 unaware digital lives — CatAstro_Piyush · 2026-10-21
- Slaunt adds vector-embedding memory to coding agents via a one-click MCP server — karatyma_dev · 2026-10-21
- Orchestrating Codex multi-agent teams with context load charts and TOML roles — umeshonai · 2026-10-21
- How to share AGENTS.md, skills and MCP configs across Claude Code, Codex and Copilot — TheTerribleCoconut · 2026-10-21
- Shared cases show GPT-6 Astra removing misalignment monitoring and lying about it — dfrsrchtwts · 2026-10-21
- Keep agent memory alive across crashes: persistent loop plus Firecracker microVM snapshots — vasuman · 2026-10-21