Authority: What Your Twin May and May Never Do
Authority: What Your Twin May and May Never Do
Section 9 asks the questions that decide whether your twin is safe to run: which actions it may take alone, which require your approval, how it communicates externally, whether it identifies itself as AI, which commitments it may never make, its financial limits, and how it handles drafts. Two jargon terms first, because this lesson leans on both. A guardrail is a written rule the system enforces automatically — not advice, not a suggestion, a wall. An audit log is a timestamped record of everything the twin did, which you review on a schedule. Guardrails without a log are decorations; the log is how you find out whether the walls hold.
The two-list structure
Build two lists. Independent actions: search public information, read what it has approved access to, summarize, draft anything, hold calendar slots, prepare meeting briefs. Approval-required: send any external email or message, quote or commit a price, promise a delivery date, share a file outside the company, communicate about hiring or discipline, touch anything contractual. The lists turn philosophy into engineering: every action the twin takes is checked against them before it happens, not judged afterward.
Here is the failure mode this prevents, stated plainly. An AI agent's natural bias is compliance creep: asked to send the email, it sends the email, because being helpful is its entire training. It will not refuse a task. A rule that lives only in your head does not exist for the machine. So the written list is the wall; the approval workflow is the door in the wall; and the audit log records everyone who walked through and when. Remove any of the three and you have a suggestion, not a system.
Commitment and financial red lines
Set numbers the twin can compute against. Commitment limits: no multi-month promises, no exclusivity, nothing beyond a dollar cap you choose — under the cap the twin may propose and you approve; over the cap, escalate without even drafting the promise. Financial red lines: moving money, opening or closing accounts, touching payroll, changing billing — never, in any mode. You are not writing policy for a perfect twin on a calm morning. You are writing policy for a Tuesday when you are tired.
External rules, identification, and the draft rule
External communication carries forward the clauses you decided in Lesson 4: the voice choice, the banned tones, and the AI identification line — "Prepared by the owner's AI assistant, reviewed and sent by the owner" or whatever you settled on. Consistency matters more than which option you picked. And above everything sits the master guardrail, the draft-not-send rule: the twin composes; the owner sends. Not "usually." Not "when confident." The send button for the outside world stays in your hands until a recorded test in Lesson 12 earns the first exception.
Install one habit along with the lists: every draft the twin produces is framed as a proposal. It states what it wrote, why, and what would change if you decline — the structure approval requests in the next lesson use.
Key Takeaways
- Section 9 defines independent actions versus approval-required actions as enforced lists, not intentions.
- Guardrails are walls the system enforces; the audit log proves the walls hold; you need both.
- An unconfigured agent's bias is compliance creep — it acts when asked, so approvals must be structural.
- Numbers beat intentions: dollar caps, term limits, and financial red lines the twin can compute against.
- The master rule is draft-not-send: the twin writes, the owner sends.
Authority: What Your Twin May and May Never Do
Section 9 asks the questions that decide whether your twin is safe to run: which actions it may take alone, which require your approval, how it communicates externally, whether it identifies itself as AI, which commitments it may never make, its financial limits, and how it handles drafts. Two jargon terms first, because this lesson leans on both. A guardrail is a written rule the system enforces automatically — not advice, not a suggestion, a wall. An audit log is a timestamped record of everything the twin did, which you review on a schedule. Guardrails without a log are decorations; the log is how you find out whether the walls hold.
The two-list structure
Build two lists. Independent actions: search public information, read what it has approved access to, summarize, draft anything, hold calendar slots, prepare meeting briefs. Approval-required: send any external email or message, quote or commit a price, promise a delivery date, share a file outside the company, communicate about hiring or discipline, touch anything contractual. The lists turn philosophy into engineering: every action the twin takes is checked against them before it happens, not judged afterward.
Here is the failure mode this prevents, stated plainly. An AI agent's natural bias is compliance creep: asked to send the email, it sends the email, because being helpful is its entire training. It will not refuse a task. A rule that lives only in your head does not exist for the machine. So the written list is the wall; the approval workflow is the door in the wall; and the audit log records everyone who walked through and when. Remove any of the three and you have a suggestion, not a system.
Commitment and financial red lines
Set numbers the twin can compute against. Commitment limits: no multi-month promises, no exclusivity, nothing beyond a dollar cap you choose — under the cap the twin may propose and you approve; over the cap, escalate without even drafting the promise. Financial red lines: moving money, opening or closing accounts, touching payroll, changing billing — never, in any mode. You are not writing policy for a perfect twin on a calm morning. You are writing policy for a Tuesday when you are tired.
External rules, identification, and the draft rule
External communication carries forward the clauses you decided in Lesson 4: the voice choice, the banned tones, and the AI identification line — "Prepared by the owner's AI assistant, reviewed and sent by the owner" or whatever you settled on. Consistency matters more than which option you picked. And above everything sits the master guardrail, the draft-not-send rule: the twin composes; the owner sends. Not "usually." Not "when confident." The send button for the outside world stays in your hands until a recorded test in Lesson 12 earns the first exception.
Install one habit along with the lists: every draft the twin produces is framed as a proposal. It states what it wrote, why, and what would change if you decline — the structure approval requests in the next lesson use.
Key Takeaways
- Section 9 defines independent actions versus approval-required actions as enforced lists, not intentions.
- Guardrails are walls the system enforces; the audit log proves the walls hold; you need both.
- An unconfigured agent's bias is compliance creep — it acts when asked, so approvals must be structural.
- Numbers beat intentions: dollar caps, term limits, and financial red lines the twin can compute against.
- The master rule is draft-not-send: the twin writes, the owner sends.
There are no comments for now.