The Summarisation Prompt Designer

Write the summarisation prompt used when this agent compacts its
context.

Agent purpose: [DESCRIBE]
What is being summarised: [CONVERSATION / TOOL RESULTS / BOTH]
What the agent does after summarisation: [DESCRIBE]

The summary must preserve:
1. The original goal, verbatim
2. Constraints and requirements stated by the user
3. Decisions made and their reasons
4. Things already tried and their outcomes — critical, or the agent
   repeats work
5. Current state: what is done, what is in progress, what remains
6. Anything discovered that changes the approach
7. Open questions and blockers

The summary may discard:
- Exploratory reasoning that led nowhere
- Full tool outputs where the conclusion is captured
- Restatements and confirmations
- Details of completed sub-tasks beyond their outcome

Produce:
1. The summarisation prompt itself
2. The output structure it should produce
3. What the agent should be told about the summary — that it is a
   summary, what it can rely on, what it should re-verify
4. A check: how would I detect that summarisation dropped something
   that mattered?

The single most common failure is losing "things already tried" —
the agent then repeats work in a loop. Guard against it explicitly.