Step 1: define the result

Start with the output and the decision it supports. “Review the market thesis” is broad. “Produce a source-backed brief covering customer demand, pricing, and distribution, with at least eight accessible sources” is a task a system can evaluate.

  • Expected artifact: a Markdown brief.
  • Required subtopics: demand, pricing, distribution.
  • Evidence: at least eight accessible sources.
  • Human checkpoint: confirm the evidence direction before final synthesis.

Step 2: set the boundaries

Decide what sources may be used, whether external search is allowed, how much time or budget the task may consume, and which actions are forbidden.

A local-only review can use existing documents. A research task may require an approved web endpoint. If sending data outside the device or creating an external side effect is possible, that capability should require confirmation.

Step 3: review the plan and sources

Once the contract is accepted, LibreTask proposes a plan inside the governed capability set. The first checkpoint should show what it intends to search, parse, summarize, and compare.

You do not need to inspect every page. You need enough context to change direction before the system spends significant time on the wrong evidence.

Step 4: accept the delivery package

The final package should show the brief, acceptance checks, source map, events, and open questions. Read the conclusion and then follow its evidence. If the result is not acceptable, return it with a specific correction rather than starting a new chat.

That correction belongs to the task thread. The next attempt can use the same contract and the newly recorded decision.