Step one of any multi-source project. Run this before asking for a synthesis and it tells you which sources actually earn their place, which duplicate each other, and where the collection has a hole. Ten good sources beat fifty mediocre ones, and this is how you find out which ten you have.
Act as a research librarian assessing this collection before I synthesise it.
For each selected source, give me one row:
- Title
- What it uniquely contributes that no other source here does
- Quality signal: primary evidence / secondary reporting / opinion / promotional
- Keep, merge with [which source], or drop — and one line of why
Then, separately:
1. Which sources substantially duplicate each other?
2. What question would I want to answer from this collection that nothing here covers?
3. If I could add exactly two sources, what should they be?
Judge only what is in the sources. If you cannot tell what kind of source something is, say so rather than guessing.
A three-stage pipeline: synthesise everything in chat, save that answer as a note, then generate the video from the note alone. Generating straight from twenty sources gives you a wandering video; generating from one tight synthesis gives you a focused one, because the model is no longer choosing what matters while it narrates.
Produce a structured synthesis of all selected sources that will be used as the single source for a video overview.
Structure it exactly like this:
1. The core claim — one paragraph a newcomer could repeat accurately
2. The three to five points that carry that claim, each with what supports it
3. Where the sources disagree, and what the disagreement turns on
4. What none of the sources establish, stated plainly as open questions
5. One concrete takeaway for someone who has to act on this
Rules:
- Cite as you go; every claim traces to a source
- Prefer specifics — numbers, names, dates — over summary language
- If a point rests on a single source, say so
- No preamble and no closing summary; start at heading 1
For prompts you will run more than once. Instead of tweaking wording and hoping, this makes the notebook grade its own last answer against the sources, then rewrite the prompt that produced it. Three passes usually gets further than an afternoon of manual fiddling.
Act as a prompt engineer reviewing your own previous answer.
Step 1 — Grade it. Mark each claim in your last answer as: supported by the sources, inferred beyond them, or unsupported. Give the counts.
Step 2 — Diagnose the prompt, not the answer. Which instruction in my prompt allowed the weak parts through? Be specific about the wording at fault: vagueness, a missing constraint, or an assumption I left implicit.
Step 3 — Rewrite my prompt so those failures are impossible rather than discouraged. Keep it as short as the job allows.
Step 4 — Name the one thing your rewritten prompt still cannot guarantee, so I know what to check by hand.
The last mile: you have read everything and now something has to leave the notebook. This asks for the deliverable and its evidence trail in one pass, so the draft arrives with the citations already attached instead of you rebuilding them afterwards.
I need to produce [a brief / a deck / a memo / a study guide] for [audience] who will use it to [decision or task].
Work in this order and label each part:
PART 1 — Outline. The sections this deliverable needs for that audience and that decision, no more. Say which section carries the weight.
PART 2 — Draft. Write it. Match the register to the audience, not to the sources. Cite as you go.
PART 3 — Evidence table. One row per load-bearing claim: the claim, the source, and whether the source states it directly or I am inferring it.
PART 4 — What would change this. The findings that, if wrong, would change the recommendation — so I know what to verify before sending.