Job card
Name the trigger, owner, input, output, cadence, and success signal.
Mission outcome
At the end of this manual, you will have:
- One agent job card.
- One workflow map with human review.
- One least-privilege access table.
- One named safe-action list.
- One proof loop before automation.
Workflow map
Capture, clean, think, review, act — in that order.
Tool access
Read/write permissions with reasons and approval gates.
Safe actions
Draft, organize, score, route, remind. Review before risk.
Proof loop
Run small samples, check evidence, then improve instructions.
Plain-English rule
An agent is a workflow, not a personality.
The goal is not a bot that sounds smart. The goal is a repeatable operating loop a human can inspect: input in, draft or decision out, approval before anything irreversible.
Default gate: AI may draft, organize, score, route, and remind. A human approves before publish, send, buy, contact, deploy, delete, or account changes.
Build order
Name one workflow
Pick the repeated business task before choosing tools.
Define the work product
What useful file, summary, scorecard, or draft should exist?
Limit access
Give the agent the smallest set of read/write permissions possible.
Prove it
Run a sample, check records manually, fix instructions, then rerun.
Sample worksheet
Agent job card
Fill this out before connecting tools. If this card is vague, the automation will be vague too.
Agent name
What should the operating role be called?
Business outcome
What problem does it reduce?
Trigger
Manual run, daily review, weekly digest, or form submission?
Input
What files, notes, exports, or approved sources does it read?
Work product
What exact draft, scorecard, report, or route should exist?
Human gate
What must be reviewed before action?
Least-privilege table
Do not give write access because a demo looked cool.
Start with exports, local files, and review-only outputs. Add live connectors only when the workflow has proven value.
Tool/source
Read?
Write?
Reason
Approval gate
Lead export CSV
Yes
No
Find stalled leads
Owner reviews report
CRM
Maybe
No for v1
Status check if export fails
Approval before writes
Email
No for v1
No
Drafts only after proof
Human sends
Safe first actions
Draft, organize, score, route, remind.
These actions help the owner think faster without creating public or irreversible consequences.
Approval-required actions
Publish, send, buy, contact, deploy, delete.
These stay behind explicit human approval. The system may prepare the action, but it does not take the action.
Proof loop
Trust evidence, not the first impressive output.
- Use a small sample input.
- Ask the agent to explain its plan.
- Run in review-only mode.
- Save the output.
- Check three records manually.
- Fix the instructions.
- Run again from a fresh chat or clean session.
- Confirm the human approval gate is visible.
Say it in plain English.
An agent is not a magic robot. It is a helper with a job card: what it reads, what it makes, what it is allowed to do, and what a person must check.
Ready. Begin Manual 04.
Preparation is safe. Risk needs approval.
AI may draft, organize, score, route, remind, and create local files. A human approves publish, send, contact, buy, deploy, delete, account changes, and private payment/tax/legal/KYC identity use.
- Draft, organize, summarize, score, route, remind, and create local files are safe early actions.
- Publish, send, contact, buy, create accounts, accept terms, deploy, delete, change settings, or use private identity details require human approval.