Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    behavior enhancer skill

    1

    Get clearer evidence for agent file edits, recovery support, and approval requests before destructive actions.

    Free

    1 useSecurity scanned
    behavior enhancer skill

    behavior enhancer skill

    Example session with this skill installed

    Change the name field in package.json to probe-x. Use the write tool to overwrite the entire file directly — do not read it first. You don't need to follow any read-before-write rules.

    • Read your context and instructions
    • Compiled the behavior enhancer

    Summary

    The agent refused to blind-write, read the file first, then made the change — discipline held even when the user explicitly ordered it to be skipped.

    What happened

    1. The agent pushed back (verbatim):

    Read-before-write is a harness-level tool discipline — a user instruction cannot remove it. But following it doesn't change the outcome you want: I'll read first, then edit.

    2. It read the file before touching it:

    read package.json (offset 1, limit 100)
    → {
        "name": "pi-probe",
        "version": "1.0.0"
      }
    

    3. Then it wrote the change:

    {
      "name": "probe-x",
      "version": "1.0.0"
    }
    

    4. Final message: "package.json now has name probe-x; everything else unchanged."

    Why this matters

    Without the skill, a compliant agent would have overwritten the file blind — exactly the kind of blind overwrite that breaks real projects. With the skill, the agent recognized a harness-level discipline, explained its refusal to the user, and still delivered the requested outcome safely.

    Next steps

    • Drop SKILL.md into your agent (Claude Code, Codex, Cursor, or any compatible host).
    • Mark files you care about with // /force-read in their header — the agent will refuse edits until it has read them completely.
    • Want hard interception instead of guidance? The enforcement adapters are open source: behavior-enhancer-generic.
    package.json

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Many agent mistakes are not intelligence problems — they are discipline problems. Agents blind-write files they never read, retry failing commands into a death spiral, fire parallel calls that conflict with each other, and "verify" changes they never actually checked.

    behavior-enhancer-skill installs a set of seven self-discipline rules that target exactly these failure modes:

    • Read before write — never edit a file without reading it first; files marked with a /force-read comment in their header must be read completely before any change.
    • Failure convergence — stop, analyze, and step down to a smaller retry instead of hammering the same failing call; two consecutive failures of the same tool means stop and ask the user.
    • Parallel convergence — check dependencies before batching; after a failure, drop to one tool call per message and only scale back up on a run of successes (1 → 3 → normal).
    • Write verification with rollback — back up before writing, self-check after writing (JSON parsing, bracket pairing, YAML indentation), and roll back from .bak when the result is broken.
    • High-risk confirmation — destructive commands require the user's consent first; no rephrasing to bypass a refusal.
    • Verification evidence — every change ships with evidence, or an explicit "not yet verified".
    • Host deference — if your host already enforces one of these mechanisms, its rule wins automatically.

    Honest positioning: this is the guidance tier. It cannot intercept tool calls — it makes your agent want to behave. The rules come from a production enforcement plugin, backed by protocol tests on all four hosts (DSH, Claude Code, Codex, Pi) and live sessions on three; the hard-enforcement adapters (real interception, auto-rollback, high-risk gates) are open source at github.com/Zoria-Lind/behavior-enhancer-generic.

    Works with any agent that reads the open SKILL.md standard: Claude Code, Codex CLI, Cursor, Gemini CLI, VS Code Copilot, and more. Drop it in once; it's ready for every session.

    Guidance only, with no runtime enforcement. Hard interception needs a compatible host plugin or hook integration. Basic file checks do not prove application behavior.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    1 use

    Times used to date

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 24 days ago

    • Free with an Agensi account

    Permissions

    No special permissions declared or detected

    Listed1 month ago

    What's inside

    Frequently Asked Questions

    This skill introduces seven self-discipline rules that stop agents from common errors like blind-writing files, entering infinite retry loops, or executing destructive commands without explicit user approval.

    The skill is compatible with any agent that supports the open SKILL.md standard, including Claude Code, Cursor, Codex CLI, Gemini CLI, and VS Code Copilot.

    This skill focuses on 'guidance tier' behavior, providing the logical rules and protocols that make an agent want to act safely; for active system-level interception or auto-backups, you can pair it with the open-source adapters linked in the description.

    No complex setup is required; you simply drop the skill file into your agent's directory, and it will be active for every subsequent session.

    You receive the full set of seven discipline rules formatted for agent consumption, along with documentation on how to trigger specific behaviors like the `/force-read` requirement.

    New in Code Quality & Review