Evil Martians Recommends Agent Hooks and nano-staged for LLM Safeguards

Per an Evil Martians blog post published May 26, 2026, the authors argue that teams should stop encoding behavioral rules and style constraints in AGENTS.md and instead run those checks through automated pre-commit / agent-hook tooling. The post, authored by Irina Nazarova, provides a concrete recipe: install nano-staged, create a .nano-staged.json that maps file globs to linters (the example shows oxfmt, oxlint, and stylelint), add a Stop hook in .claude/settings.json that runs ./node_modules/.bin/nano-staged --unstaged --quiet --bail || exit 2, and optionally add husky to invoke nano-staged on git commit.

Per the post, the .nano-staged.json example runs linters only on changed files rather than the whole repository. The post shows inline command examples and recommends writing custom linter rules (for example, an ESLint rule) if an LLM repeats the same error, noting that doing so pushes the constraint into a deterministic check rather than relying on the agent to remember documentation text.

Moving checks from doc-based instructions into deterministic tooling is a common engineering pattern for reducing human error and variability. Running linters on changed files only cuts compute and latency compared with re-sending long documentation to an LLM, and integrating checks into pre-commit hooks or CI yields faster, reproducible feedback for developers and agents alike.

For teams building agent-driven developer workflows, the blog frames a pragmatic cost and reliability tradeoff: token and latency savings from not re-transmitting large AGENTS.md files, plus the lower flakiness of deterministic linters. This matters for practitioners who embed agents into developer loops because small, frequent feedback cycles are typically cheaper and easier to automate than relying on an LLM to enforce style and safety rules from free-text docs.

Observers should watch adoption signals in open-source agent integrations and pre-commit templates: more templates that include nano-staged or equivalent per-file linters, example .claude/settings.json hook snippets, and community recipes for converting repeated LLM failure modes into custom linter rules. Also watch whether agent platforms add first-class support for running repository-scoped linters as agent hooks.