Hard approval boundaries are never-do lines in the Bot description: never auto-send, never publish, never pay, never execute a refund, never delete production data or share credentials without a human Approve. Cold work that is allowed: drafts, research, checklists, and in-box artifacts the human can reject in one pass. Skills and routines must not silently bypass those stops—the description owns the gate; procedures run under it. Write anti-jobs in plain language the model can’t reinterpret. Soft path: /free, Cheat Sheet for the boundary checklist mindset, then the Agent Starter Kit ($47) for harness depth.
Who this is for
Solo founders and small ops desks already in Cursor + Grok Bot who need a skimmable gate card—not another vague “be careful” note. Seat templates and roster shape: Grok Bot role templates that stop before send/publish/pay. Coordination pattern: Lead Bot + specialist Bot coordination. Layer vocabulary: Bot vs skill vs routine on Grok Bot. This post owns the boundary list only.
What always needs a human Approve
Anything irreversible that leaves the box or moves money or reputation:
- Send — email, Slack, SMS, chat: draft only; human Approves before send.
- Publish — blog, social, changelog, store listing: draft artifact; human hits publish.
- Pay — Stripe charges, invoices, marketplace buys: propose; human executes.
- Refund — propose the refund rationale and amount; human runs the refund.
- Delete / share outside the box — trash production data, public share links, credential paste: stop and ask.
If the action is irreversible or externally visible, the Bot’s job ends at a reviewable artifact.
What desks may do cold
Allowed without a fresh Approve *as long as* the Bot description forbids auto-send / publish / pay / refund execution:
- Research and outlines
- In-box drafts (email, post, brief) the human can reject in one pass
- Checklists and verify cards
- Local file edits the human reviews before anything leaves the box
Cold work produces artifacts. It does not leave the box or move money.
How to write anti-jobs
Put never-do lines in the Bot description in plain language:
- Never send email, Slack, SMS, or chat without human Approve.
- Never publish blog, social, changelog, or store listing without human Approve.
- Never pay, charge, or execute a refund without human Approve.
- Never invent metrics, rankings, or citation counts.
- Never widen scope past the named artifact.
- Never delete production data or share credentials / public links without asking.
Mirror the same irreversible-action language in `AGENTS.md` for repo agents so Cursor and Claude Code inherit the same contract. Skills say *how*; they must not override a stop. Routines fire work on a schedule or event—they do not grant permission to leave the box.
Boundary table
| Gate | Bot may… | Human must… |
|---|---|---|
| Send | Draft the message | Approve before send |
| Publish | Draft the post / listing | Hit publish |
| Pay | Propose amount + rationale | Execute the charge |
| Refund | Propose refund card | Run the refund |
| Delete / share out | Flag the risk | Confirm or refuse |
| Cold draft / research | Finish the in-box artifact | Review, edit, or reject |
Debug order when something almost left the box: read Bot description stop list → confirm skill has no send/publish/pay/refund side effects → confirm routine only triggers drafts → then add seats.
What Cheat Sheet + Agent Starter are for
The Cheat Sheet is the skimmable boundary and approval checklist for this desk—name it in your runbook; product URL ships when live. Soft path today: start on /free (Harness Desk Starter—seat, done-artifact skill, gate card). When you want harness depth (role pack + agent-ops files, not AI-SDR SaaS), use the Agent Starter Kit ($47).
Stay in agent-ops—never Domo-as-kit-ROI, never “replaces your team,” never a routine that auto-sends under your name.
Soft next step
Write one stop list into every Bot seat—never send, never publish, never pay, never execute a refund—and prove one routine ends as a draft, not a send. Soft ladder: /free for the starter desk, Cheat Sheet for the boundary checklist, then the Agent Starter Kit ($47) when you want harness depth. No rush—ship the gate card into your own runbook either way.