Research a question into a sourced brief

Ask a question once and get back a written brief in which every claim is tied to a page this workflow actually fetched and read. It searches the web for your question, reads the top results one at a time, and writes the brief from that text alone - then lists each source with its link underneath, numbered so the [1] and [2] markers in the prose point somewhere.

The honesty of the result is the whole design. The search engine's own written answer is deliberately not requested, so nothing enters the brief without a page behind it; the model is told to use only the supplied text and to say plainly when the sources do not answer the question rather than filling the gap from memory. If the search finds nothing, or if none of the pages can be read, the run stops and says so instead of producing a confident write-up with nothing underneath it.

'maxSourcesToRead' caps how many pages are fetched, so one broad question cannot run up an unbounded bill. Because that cap can bite, the brief ends with a line reading 'Sources used: 3 of the 8 search results found (1 could not be read)', which is what stops a partial run from looking like full coverage. A page that will not load costs one source and is reported; the rest of the list still goes through.

The brief is saved as a new Google Doc each run, so nothing you already have is overwritten. Set it up by typing your question into the 'researchQuestion' step - that is the only blank. Worth tuning: 'maxSearchResults' widens or narrows the pool the reader draws from, 'charactersPerSource' decides how much of each page reaches the model, and the search step's depth can be raised from Balanced to Thorough for harder questions at twice the search credits.