Negative-skills / antipattern files for AI coding agents
Problem
AI coding agents pick skills via regex-based reading and regularly do 'exactly opposite of what is intended' because there is no standard way to express negative constraints. Practitioners are improvising NEGATIVESKILLS.md files and asking whether an accepted pattern exists — it doesn't.
Opportunity
A linting/management layer for agent instruction files that supports explicit antipatterns and 'do-not' rules, validates them against model behavior, and ships as a standard format plus a registry of curated negative-skills packs.
Market analysis
The failure mode is real and recent research finds that negative-versus-positive phrasing of agent instructions is measurably effective, but the format layer is owned by Anthropic and Vercel, and hooks already provide deterministic enforcement.
Market · Agent power users and teams maintaining large SKILL.md / AGENTS.md collections; skills.sh hosting tens of thousands of skills shows the ecosystem scale.
Pricing · Ecosystem tooling here is free and open source; realistic monetization is a paid CI validator or curated packs, not the format itself.
Pros
- + Documented pain: practitioners improvise NEGATIVESKILLS.md files because no accepted pattern exists.
- + Empirical backing: a 2026 paper on agent rules found the negative-versus-positive split effective for coding-agent instructions.
- + A distribution channel already exists: skills.sh and 'npx skills add' make curated packs installable in one command.
Cons
- − Anthropic (Agent Skills spec) and Vercel (skills.sh) control the standard and can absorb negative constraints natively.
- − Hooks (PreToolUse blockers) already enforce 'never do X' deterministically, regardless of how the model reads instructions.
- − Validating do-not rules against actual model behavior requires expensive eval runs per model and version.
Existing / similar tools
Source
Hacker News (Ask HN)
The opening is narrow but real: nobody validates instruction files the way ESLint validates code, and a checker that lints SKILL.md/AGENTS.md for unenforceable phrasing (missing negatives, conflicting rules, trigger-regex drift) fits naturally into CI where hooks cannot. Note that hooks already answer “stop the agent from doing X” deterministically, so the product must target the authoring side (helping humans write constraints models actually follow) rather than the enforcement side, where the platform has won. Standards risk is the killer: if SKILL.md frontmatter grows a native ‘forbid’ field next year, a standalone competing format is dead, while a linter that reads whatever the standard becomes survives.