# DOE map worksheet

Split your workflow into Directives, Orchestration, and Execution before you build it.

Every reliable agent keeps 3 things separate. Directives are the written rules, a text file a human edits. Orchestration is the decision the AI makes by reading those rules. Execution is deterministic code that carries the decision out. Errors compound when these blur, and a 90% accurate step run 5 times in a row is 59% accurate end to end. Keep the boxes clean and the failure always has an address.

## Completed example: email triage

| What is written down (Directives) | What the AI decides (Orchestration) | What code executes (Execution) |
| --- | --- | --- |
| The directive file: 4 categories, priority order, auto-reply rule, cold-pitch rule, if-unsure default | Which of the 4 categories this email is. Nothing else. | The 4-way switch: post to Slack, save a draft, archive, or label and forward to bookkeeping |

## Your workflows

| What is written down | What the AI decides | What code executes |
| --- | --- | --- |
| | | |
| | | |
| | | |

## The 2 audit questions

Answer both for every row. A no on either one means redraw the map before you build.

1. Does the model perform any action itself? It should not. If the model sends the email or writes the row, move that step into the Execution column and have code do it.

2. Can an operator change behavior without a developer? They should be able to. If changing what the system does means editing code, business logic has leaked out of the Directives column. Pull it back into the text file.

---

From Agents That Work, lesson 4. advizracademy.com