Agent Planning

Before asking an agent to do heavy work, make a short plan. A good plan is easier to review than a long answer, and it gives the workspace the right shape before papers, PDF evidence, and manuscript drafts start to accumulate.

Planning prompt

Use Chat or Side chat for a plan when the task has more than one step:

Make a short research plan before answering.
Keep it source-grounded.
List the papers or files that need to be checked.
Do not draft final claims until the evidence is reviewed.

Re-plan when direction changes

If a search strategy, reading plan, or draft structure turns out wrong, summarize what you learned and start a new Side chat for a cleaner second pass instead of patching a confused thread.

Example Demo

Scenario: you want an agent to help with a mini-review, but you do not want it to directly draft unsupported prose.

Ask:

Plan first; do not write the review yet.
Topic: PCAT and coronary CTA prognosis research.
List search terms, priority paper types, evidence-table fields,
and the human review checkpoints required before Document Writing.

A good output is a reviewable plan: search strategy, inclusion criteria, reading order, evidence-table fields, manuscript outline, and human checkpoints. After approval, move into Literature, Paper reading, or Document Writing.

Use stronger workflows for high-stakes work

For high-stakes reviews, systematic searches, and manuscript work, use the most capable model and most specialized workflow available. The practical metric is not response speed; it is how often you need to intervene, correct unsupported claims, or redo the task.

Review output before reuse

Treat agent output as a working draft, not final evidence. Return to original papers before relying on conclusions, and move stable outputs into Deliverables or Document Writing only after checking the source trail.