Naming the four ways agents build fortresses instead of features

What it does

HERO is a block of text you paste into your agent’s config file (CLAUDE.md, AGENTS.md, .cursorrules, etc.) that tells the agent to stop padding work with defensive ceremony. It names four recurring shapes of over-defense — Hashing, Edge cases, Rubrics, Overbuild — so instead of arguing vibes, you can say “that’s an H” and quote the matching case back. It works with Claude Code, Codex, Cursor, Copilot, Windsurf, Antigravity, and Gemini CLI; there’s nothing to install.

The interesting bit

The framing: these behaviors look like an agent optimizing for not being blamed rather than the work being good. The README is honest that this is a hypothesis, not established training lore — but the shapes are real and observable. The rules are also carefully two-sided: they bound what the agent proposes, never what it looks for, and two “looks like over-defense but isn’t” cases are marked ✓ to report rather than dismiss.

Key highlights

• Nine rules covering hashing, scaffolding, corner-case obsession, judgment-by-checklist, defensive prose, stalling, and generation traces
• A cases/ catalogue of observed behaviors — what was asked, what the agent did, what proportionate looks like — meant for quoting back, not pasting in
• Bilingual (English/Chinese) contract, distilled from the author’s ~13.7k★ ARIS project
• A changelog that flags when the block changed and your paste is stale — with a warning that the install command won’t replace an old copy
• Documented failure mode: the block fades on long runs, with three causes and how to tell which one you have

Caveats

• It’s a prompt, not a mechanism — the README itself admits effectiveness thins over long sessions and under compaction
• The block has changed several times since publication; early adopters need to manually delete and re-paste

Verdict

If an agent regularly hands you checksums nobody reads or a twelve-stage audit loop around a two-line feature, this is a cheap, well-documented fix. If your agents already stay in scope, it’s a taxonomy you’ll enjoy reading and never need.