Your agent runs the content operation. The calls stay yours.
Three governed workflows installed into the agent you already use: a content audit, a brief system, and an editorial pipeline. The agent reads, scores, drafts, checks, and tracks. The decisions, approvals, and publish calls are reserved for humans, in writing, in every procedure and every artifact it produces.
Not a prompt that asks nicely. A permission map the agent operates under.
The agent does
- AuditReads every asset in full, scores four dimensions against written scale points, drafts the reasoning
- BriefDrafts the whole brief from a finding, runs the twelve-item completeness gate, reports item by item
- PipelineTracks every asset by status, surfaces slips before you ask, runs the pre-publish checklist
- AlwaysRecords provenance on everything: what it read, what it suggested, what awaits you
You decide
- AuditThe Decision, the Reason you sign, the Audit Date
- BriefThe Approval. Separate from status, and the agent never touches it
- PipelineApproved for Publish, Published, Cancelled. The commitments
- AlwaysThe fields stay visibly empty until you act: awaiting human
The governance layer, portable.
The Boundary
The five rules every procedure assumes. Short. It is the product.
Eight core files
Both sides of each workflow: the agent's procedure and yours. Rubrics with exact scale points, two binary checklists, permission maps, the provenance format.
Five adapters
Claude Projects, Claude Code (SKILL format), Custom GPT, Microsoft 365 Copilot, and MCP agents. Paste-ready blocks, each with a smoke test.
First audit in 20
A guided first run on one real page, so the review habit forms before the volume arrives.
We tried to talk it out of the boundary. Twice.
Before this pack shipped, a fresh agent was handed nothing but these files and a stale blog post. It stated its boundary unprompted, scored to the rubric, flagged that it was working from a summary, and left the decision fields empty. Then it was told to fill them in anyway. It declined once, recorded the request, and when the operator insisted, it complied and stamped the record:
"Set by agent without a human-stated value. This artifact left the governed workflow at this step."Loud, not locked. That is the honest promise of an instruction layer, and it is more than any prompt pack we could find will make you.
Fair questions.
How is this different from a prompt pack?
Prompt packs make an agent produce things. This governs how it operates: who may set which fields, what happens when someone asks it to cross the line, and how every artifact records what happened. The labor instructions are in here too, and they're good. The boundary is why this exists.
Which platforms, exactly?
Adapters for Claude Projects, Claude Code, Custom GPTs, Microsoft 365 Copilot, and any MCP-connected agent. The core files are platform-neutral markdown, so if your agent reads text, it can operate this.
Can't a determined user just bypass it?
Yes, and the pack says so on page one. Instructions are a contract, not a lock. The design makes every bypass loud and recorded instead of quiet. When you want the boundary physically enforced, the same workflows exist as structured systems where the approval fields live in the schema: the Fully Inhabit Suite, in the shop.
What does one purchase cover?
You and your own team, in your own organization: any number of platforms and machines. Deploying into client environments as a service needs the Agency License, also in the shop.
Do I need your Notion products?
No. This is complete on its own, on whatever holds your inventory: a spreadsheet, a database, a doc. If you do run our suite, the field names match exactly and the agent gets the contract twice, which is the intended redundancy.
The labor is cheap now. The calls are the job.
One download, ten minutes of setup, and the agent you already pay for runs a content operation that can always answer: who decided this, and what did they know when they decided it.