33 evidence-backed checks

    The 33 checks AgentLint runs on agent harnesses.

    AgentLint checks whether CLAUDE.md, AGENTS.md, CI, hooks, and adjacent rule files are correct, specific, structured, enforceable, and clean enough for agents to use.

    01 / Full check list

    Correctness

    8 checks
    • Referenced files must exist at the paths the harness names.
    • Commands must run from the working directory the instruction implies.
    • Package-manager guidance must match the repo lockfiles.
    • Build, test, and lint rules must match actual project scripts.
    • Tool names and agent entrypoints must match installed configuration.
    • Local documentation links must resolve.
    • Environment variable names must be real and consistently spelled.
    • Harness rules must not contradict CI, hooks, or repository policy.

    Specificity

    10 checks
    • Vague advice must be rewritten as observable action.
    • Risky file classes need explicit ownership or approval rules.
    • Common workflows need exact verification commands.
    • Ask-versus-act boundaries must be stated.
    • Forbidden operations must name the command, path, or data class.
    • Expected outputs must name the format or destination.
    • Generated-change limits must state scope.
    • Decision rules are preferred over personality traits.
    • Communication requirements must be measurable.
    • Loaded adjectives must include examples or be removed.

    Structure

    6 checks
    • The main instruction file should stay short enough to scan.
    • Stable headings should separate facts, rules, and enforcement.
    • Mandatory rules must be separated from preferences.
    • Project facts should appear before workflow policy.
    • Tool-specific rules should stay in the tool-specific layer.
    • Duplicate rules need a named source of truth.

    Enforcement

    5 checks
    • Critical rules need CI, hooks, permissions, or required checks.
    • Unsafe operations need a mechanical approval boundary.
    • Formatting rules need the formatter command.
    • Review expectations need a checklist, test, or handoff step.
    • Failing lints need a stated owner or escalation path.

    Hygiene

    4 checks
    • Historical notes should be removed from active instructions.
    • Secrets and private values must not appear in the harness.
    • Rules need periodic review after tool or workflow changes.
    • Examples must use current repo names, paths, and commands.

    Run the full check set against your repo.

    The point is not a higher score in isolation. The point is a harness that gives agents current facts, clear decisions, and verifiable constraints.

    Install AgentLint