A personal agent can help organise research. The discipline of deciding what matters still needs an owner. Use this guide to shape work that another person can inspect.
Start with a decision.
“Research connectivity” is a topic. “Compare three connectivity options for a second office, using these budget and resilience requirements” is a task. State who will use the answer, the constraints and the date by which the evidence must be current.
Write the decision or learning goal in one sentence.
List the criteria that would change your choice.
Define the sources the agent may use and what it must not access.
Specify the output: a brief, comparison, question list or draft.
Agree the stopping point and the human reviewer.
Exercise: rewrite one broad question as a bounded research task with three criteria and one explicit uncertainty.
Make the evidence traceable.
A citation is a pointer, not proof. Open the source. Check that it supports the precise claim, refers to the right period and applies to the case you are considering. Prefer original documentation for product capabilities and original research for research claims.
Several articles may repeat the same announcement. Trace them back before treating them as independent agreement. Record disagreements, missing information and whether a claim is verified, inferred or still a hypothesis.
For every material claim, keep the supporting URL and a short explanation.
Record when the source was checked; a publication date alone does not establish current accuracy.
Separate a supplier’s claim from an independently tested finding.
Ask what evidence would make you change your mind.
Keep useful context. Set useful boundaries.
Memory should save effort without quietly mixing clients, private information and old assumptions. Make the relevant project, owner and permitted use clear. A saved preference can become wrong; context needs a way to be corrected or removed.
The ask.pa workbench offers a small example: memory is off by default, stores only three questions locally and has an explicit forget control. It is not account-level memory or a secure store for private material.
Exercise: identify which parts of a research brief should survive into the next task and which should expire when the task ends.
Review before the next action.
Check the answer against the original question and constraints. Verify the consequential claims. Look for missing alternatives, outdated evidence and assumptions that slipped into the summary.
Turning a brief into an email, purchase or system update changes the task. A production agent should show the proposed action, destination, inputs and authority before executing it. A research result does not grant permission by itself.
Can another person follow the source trail?
Are uncertainty and conflicting evidence visible?
Is the result useful within the agreed budget?
Is the next step a draft, recommendation or executed action?