Capstok — learn by doing

Why this matters

Not every bot action is the same. Some should just happen (auto-deleting known spam), some should happen but with you cc'd (draft-and-notify), some need explicit approval before happening (send this reply), and some should be refused outright by the bot with a nudge to you (respond to a lawsuit threat). The escalation ladder is a four-rung classification you apply to every action type: AUTO (bot acts silently), NOTIFY (bot acts, sends you a receipt), ASK (bot proposes, waits for approval), REFUSE (bot never handles; forwards to you with context). The rung is a function of stakes, reversibility, and how much personal context the response needs. Every module in this course picks a rung for its actions and defends the choice — and the delegation audit you did earlier is where you first practice.

Demo

A policy engine that maps action types to escalation rungs based on stakes/reversibility/context signals. Every rung has a matching UI (silent, Telegram message, approval prompt, forwarded email).

Try it yourself

For each action type in your delegation audit, pick a rung and defend it in one sentence.
Find an action currently at AUTO whose stakes would move it to NOTIFY if the reversibility changed. Design the trigger.
Design one 'catastrophic' action the bot should REFUSE entirely, with the forwarding message it would send you.
For the NOTIFY rung, decide the notification medium (Telegram DM, email digest, batched daily) for each action type.

Prompt your AI

Use these three in order. Each builds on the one before.

1. Basics & terminology

In one paragraph, explain the four-rung escalation ladder for a bot acting on my behalf.

2. Why it works (the mechanism)

Walk me through how reversibility and personal-context signals should shape rung selection — with examples where they conflict.

3. Advanced — application & what's next

I want the bot to reply to LinkedIn DMs. Given the mix of stakes (some VIPs, mostly recruiters) and reversibility (a bad reply is embarrassing but recoverable), what's the rung and what's the exception rule?

References

Chat about this lesson

type Rung = "AUTO" | "NOTIFY" | "ASK" | "REFUSE";
type ActionType =
  | "delete_newsletter_spam"
  | "archive_low_priority"
  | "draft_reply"
  | "auto_send_reply_to_vip"
  | "answer_dm_on_my_behalf"
  | "respond_to_legal_threat"
  | "respond_to_investor";

interface Signals {
  reversible: boolean;
  stakes: "low" | "medium" | "high" | "catastrophic";
  needsPersonalContext: boolean;
}

const POLICIES: Record<ActionType, Signals> = {
  delete_newsletter_spam:    { reversible: true,  stakes: "low",          needsPersonalContext: false },
  archive_low_priority:      { reversible: true,  stakes: "low",          needsPersonalContext: false },
  draft_reply:               { reversible: true,  stakes: "medium",       needsPersonalContext: true  },
  auto_send_reply_to_vip:    { reversible: false, stakes: "high",         needsPersonalContext: true  },
  answer_dm_on_my_behalf:    { reversible: false, stakes: "high",         needsPersonalContext: true  },
  respond_to_legal_threat:   { reversible: false, stakes: "catastrophic", needsPersonalContext: true  },
  respond_to_investor:       { reversible: false, stakes: "high",         needsPersonalContext: true  },
};

function pickRung(action: ActionType): Rung {
  const s = POLICIES[action];
  if (s.stakes === "catastrophic") return "REFUSE";
  if (!s.reversible && s.stakes === "high") return "ASK";
  if (s.needsPersonalContext && s.stakes === "medium") return "NOTIFY";
  if (s.stakes === "low" && s.reversible) return "AUTO";
  return "ASK";
}

// Pretty-print the policy
for (const [action, _] of Object.entries(POLICIES)) {
  console.log(action.padEnd(30), "->", pickRung(action as ActionType));
}

Run: node main.js

The escalation ladder: auto → notify → ask → refuse