Multi-session coding agents kept freelancing on me, so I wrote a charter + draft-only gate (free skeleton inside)

I use Claude Code / Cursor-style agents across sessions and kept hitting the same mess: helpful drafts that quietly turn into outbound actions, merges, or invented context. Instead of another mega-prompt, I wrote a small operating sheet: • One-paragraph charter: purpose, inputs, outputs, authority, stop conditions, human owner • Draft-only by default: research and prep OK; no send / publish / merge / delete / spend / prod without an explicit approval plus a short approval record • Skills as bounded procedures: when to use, inputs, steps, forbidden actions, return shape, how a human checks it • 7-day checklist: charter → gates → 2–3 skills → daily status routine → one end-to-end draft run → tighten I packaged the fillable templates as Agent ops pack ($29 one-time digital download, not consulting, not official Anthropic/Cursor docs). Free charter skeleton below. Platform-neutral; I tested the same boundaries against multi-session coding agents. Curious what you all do: for Claude Code multi-session work, do you put the “do not act externally” rule in CLAUDE.md / project instructions, in a skill, or both? Happy to answer build questions. If you want the full template kit, ask and I’ll share the link. Keeping this post link-light. Free excerpt: Agent team charter skeleton Team name: Owner / final approver: Purpose: (1–3 sentences) In scope / Out of scope: Inputs and allowed sources: Outputs: Agent roles: Coordinator / Researcher / Builder / Reviewer Default permissions: Draft and analyze only. No publish, send, merge, delete, spend, or production changes without approval. Approval gate: Proposed action, exact content, destination, affected people/systems, risks, rollback. Escalate when: Ambiguous requirements, unverified info, irreversible/external/costly actions, inventing facts/credentials. Definition of done / Review cadence: