Planning long tasks

How to break complex multi-step work into reliable agent instructions

Why long tasks go wrong

Agents fail at long tasks for predictable reasons: the goal is too vague, the steps aren’t specific enough, there’s no error recovery plan, and there are no checkpoints to catch problems before they compound. Good task planning prevents all of these.

The task brief format

TASK: [Clear, specific name]
GOAL: [What success looks like — the end state, not the process]
INPUTS: [What data/context the agent needs to start]
STEPS: [Ordered sequence of specific actions]
CHECKPOINTS: [Where to pause and report back]
OUTPUT: [What to produce at the end]
ERROR HANDLING: [What to do if something goes wrong]
OUT OF SCOPE: [What NOT to do — explicitly]

Example: weekly client reporting brief

TASK: Generate weekly status reports for 5 active clients

GOAL: Each client has a drafted status update email ready for my review by Monday 8am

INPUTS:
- Client list with emails (in /clients/active.md)
- Project status for each client (in their Drive folders)
- Last week's sent updates (in Gmail, label: "client-updates")

STEPS:
1. Read the client list and identify the 5 active projects
2. For each client:
   a. Find their project folder in Drive
   b. Read the latest status notes
   c. Read last week's update (for continuity)
   d. Draft a status update following the template in /templates/status.md
   e. Save the draft to Gmail (do not send)
3. Create a summary doc in Drive listing all 5 drafts with links

CHECKPOINTS:
- After step 1: confirm client list is correct
- After completing all 5 drafts: report what was created

OUTPUT: 5 Gmail drafts + one summary document in Drive

ERROR HANDLING:
- If a client folder is not found: note it and skip, flag in summary
- If last week's update is missing: proceed without continuity context

OUT OF SCOPE:
- Do not send any emails
- Do not modify the project status notes