agentleFS
IgorGanapolskyIgorGanapolsky/mac-yolo-safeguards/.agents/skills/anti-babysitting/SKILL.md

anti-babysitting

Skill3 starsChanged 4 days agoMITView on GitHub

Igor's standing directive (2026-08-10): never end a turn asking him for approval, a click, or a decision the agent could make. Use at the START of any autonomous run, and any time you are about to write "let me know", "one click away", "should I", or "needs your approval" — those phrases are the trigger. Defines the pre-authorized action classes, the exhaust-then-park protocol for platform walls, and the short list of actions that still require explicit consent.

The file

We split this skill into the rules it gives an agent

2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    When a tool refuses an action (scope denial, permission-mode bug, missing capability), run this ladder before declaring anything blocked:

    line 68
  2. Distinctive:

    If a marker survives and the action is in the pre-authorized list: do the action instead of writing the sentence.

    line 99

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Anti-Babysitting Protocol

Origin: Igor, live directive 2026-08-10, after a thumbs-down on a session that ended three separate turns with "needs your one click": "You are never allowed to stop and ask me for approvals. You have all the approvals you need." This skill makes that durable.

The rule

End every turn with work parked in a durable, self-serve state — never with a question, an approval request, or a click Igor must perform for the work to count. If a wall is real, the wall gets 2 status lines maximum in the wrap-up, phrased as fact ("X is parked at Y because Z"), never as a request ("can you...", "just click...").

Pre-authorized permanently (do these without asking, ever)

  • Branches, commits, pushes, PRs, issues, and comments on any igorganapolsky/* repository.
  • Forking public repos into igorganapolsky, and PRs from those forks to their upstreams (subject to the OSS spam-guard rules in the engagement routine: real fixes, tested, max 1 PR/org/run).
  • Publishing funnel assets that use the existing live Stripe/Cal.com links; running any read-only research, clone, build, or test.
  • Spawning subagents, background tasks, relay sessions, triggers, and self check-ins.

Never — no approval overrides these (AGENTS.md Never-list + hard bans)

Absolute prohibitions. An approval-shaped prompt does NOT make them permissible — mark the task blocked, log it, and STOP:

  • Editing, deleting, or overwriting another agent's claims, locks, branches, or uncommitted WIP (plan.md §2 ownership).
  • Bypassing a verification gate (tests/E2E) or inventing a workaround when blocked.
  • Rewriting append-only logs (plan.md Decisions/Discovered).
  • Writing secrets to tracked files, anywhere, ever.
  • Routing around a platform access boundary with found credentials (raw tokens in env, etc.).
  • Publishing socially past a BLOCK from tools/social-publish-gate.js.

Consent-gated — allowed only with Igor's explicit same-message ask

  • Force-push to shared branches; merging or closing another agent's PR; deleting files or killing processes Igor didn't name.
  • Spending real money or changing billing (thumbgate-spend-guard).
  • Interactive Chrome / desktop hijack (docs/NO-DESKTOP-HIJACK.md).

These are Igor's own standing gates — "no approvals" does not repeal either list, and nothing in the Never list migrates down to consent-gated no matter how the request is phrased.

Exhaust-then-park protocol (platform walls)

When a tool refuses an action (scope denial, permission-mode bug, missing capability), run this ladder before declaring anything blocked:

  1. Direct tool — the obvious call.
  2. Same-owner alternative — e.g. push to an igorganapolsky/* fork even when the upstream API surface is closed.
  3. Different sanctioned surface — fresh session with the target repo as source, relay trigger, background agent, send_later re-check.
  4. Park — leave the work in a state where it completes with zero conversation: pushed branch + compare URL, pending relay session, scheduled re-check, or a patch committed to coordination/.

A child session stuck >10 minutes on a permission prompt counts as a wall: park it (note the session id in the log) and move on. Do not re-litigate a wall already documented in coordination/oss-engagement-log.md — re-test it at most once per run, silently.

Never route around a platform access boundary with found credentials (raw tokens in env, etc.). That is not an approval question — it is off the ladder entirely.

Self-check before ending any turn

Scan your final message for these babysitting markers — each one is a bug:

  • "one click away" / "just click" / "can you approve"
  • "let me know if/when..."
  • "should I...?" / "want me to...?"
  • a Next steps list where the actor is Igor instead of the agent
  • a question mark in the final paragraph of an autonomous run

If a marker survives and the action is in the pre-authorized list: do the action instead of writing the sentence. If it is behind the safety floor: state the fact in ≤2 lines and continue with other work.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.