Teaching AI Agents to Forget Their Rejected Ideas

What it does

no-negative-echo is an Agent Skills-format module that stops iterative AI agents from smuggling discarded ideas into final deliverables. It provides structured prompts and check scripts that force the agent to regenerate titles, commit messages, PR descriptions, and UI copy purely from the final adopted state—so you get “番茄炒蛋” instead of “番茄炒蛋（没有东坡肉）”.

The interesting bit

The project treats prompt hygiene as installable infrastructure. It ships with SHA256-verified installation contracts, a multi-surface scanner for filenames and suspicious Unicode, and a high-assurance finalization mode, all framed around a genuinely annoying LLM behavior: the compulsion to narrate what it didn’t do.

Key highlights

• Ships in two modes: a full Skill install with check scripts, or a lightweight AGENTS.md ruleset for passive guardrails
• Built-in scanner checks text, filenames, and suspicious Unicode residues (though it explicitly warns that PASS is not semantic proof)
• Includes an evaluation protocol under evals/, with CI running deterministic scorer tests
• Explicitly scoped as prompt-layer mitigation: it cannot purge model context or tool logs, and requires explicit invocation before important deliveries

Caveats

• The Skill must be explicitly called for important deliveries; mere discovery by the agent does not activate it
• The scanner only catches textual and filename residues, not deeper semantic echoes
• The AGENTS.md mode depends on the host agent (e.g., Codex) actually reading project instructions at session start

Verdict

Worth a look if you use agentic coding and are tired of commit messages that read like therapy sessions about rejected approaches. Skip it if you aren’t using Codex or similar agents with iterative planning loops.