You paste a 40 message email thread into Claude. You ask for the summary and the next action. The model reads the first six messages carefully, skims the middle, and bases its analysis on the most recent three. The buried decision at message 23 is missed. The actual disagreement at message 14 is missed. The full context is in the thread. The model just did not read it properly. The fix is one line of setup.
The trick
Before pasting any long thread, transcript, or document chain, add this line:
I will paste a long thread with messages numbered [1] through [N]. In your analysis, cite the message number for every specific claim you make. If you reference an event or decision, tell me the message number where it appears.Then number the messages in your pasted content. The numbering forces the model to anchor its analysis to specific positions in the thread, which forces it to actually read all of them.
Why it works
Large context window does not equal full attention. Models tend to over weight the start and the end of long inputs and under weight the middle (the technical name is the "lost in the middle" effect, and it is real across every frontier model). When you require message number citations, you implicitly require the model to track every message. The model cannot cite a number it did not read.
When to use it
Use it on any long input where the important information might be buried: email threads over 10 messages, meeting transcripts over 30 minutes, document chains, Slack threads, customer support histories, anything where the decision or signal might not be at the top or bottom. Skip it on short inputs where the model can attend fully.
Three sharper variations
For decisions buried in threads: "Find every decision or commitment made in the thread. For each, cite the message number, the person who committed, and whether the commitment was acknowledged by anyone else."
For disagreements: "Identify every point where two people disagreed in the thread. Cite message numbers for both sides. Tell me which disagreements were resolved and which were left open."
For long meetings: "Number the speaker turns in the transcript [1] through [N]. List every action item with the speaker turn where it was committed. Flag any action item that was raised but not assigned."
Before and after
Without the numbering, asked to summarise a 40 message thread about a vendor selection: "The team is considering Vendor A based on the most recent discussion. There was some initial interest in Vendor B but the conversation moved on. The decision is leaning toward Vendor A."
With the numbering: "The team explicitly chose Vendor B at message 14 based on the procurement constraints raised at message 11. The discussion that appeared to move toward Vendor A at messages 32 to 37 was about a different evaluation criterion (technical fit) which was framed as a tiebreaker, not a decision reversal. The actual decision per message 14 still stands, though it has not been acknowledged since message 18."
The second output catches what you would have missed. The first one would have walked you into a meeting where you were the only person who thought Vendor A was the choice.
The deeper version
For an AI workflow that runs this analysis automatically on every long thread, you set up a Layer 3 workflow with the numbering trick built into the template. For the Layer 3 workflow examples that use this exact pattern, read Inside Five Layer 3 Workflows.
The reframe
Large context windows make operators trust that the model is reading everything. The model is not reading everything. It is reading the start, the end, and skimming the middle. One line of setup forces the model to anchor its analysis to specific message positions, which forces it to actually read all of them. The cost is twenty seconds. The cost of not doing it is the decision you missed at message 23.
PromptLeadz publishes battle tested AI prompt packs for operators across all functions. All prompts are LLM agnostic. Pricing is in USD.
Deja un comentario: