Built for the 2,000-to-15,000-character middle ground β the analyst note, the policy document, the lecture transcript β where a chat prompt gives you mush and a manual read costs an hour.
The pilot ran across three regional offices from March to July. Support tickets per rep fell 18%, and first-response time improved from 9.4 to 6.1 hours. Two of the three offices reported that the new triage rules conflicted with their existing escalation matrix, and adoption in the third office stalled at 40% because the training sessions landed during their peak season.
The three-office pilot cut tickets per rep by 18% and first-response time from 9.4 to 6.1 hours. Two offices hit rule conflicts with existing escalation paths, and the third stalled at 40% adoption after training was scheduled into peak season. Rollout should resolve the escalation conflict and move training outside peak weeks.
Inputs above roughly 8,000 characters are split into sections. Each section is condensed into notes that keep names, dates, figures and conclusions, and those notes are then merged into one summary. That is why a 15,000-character report comes back as a single piece of prose instead of four stitched-together fragments.
For documents over the limit, split at natural boundaries β chapter, section, agenda item β and summarize each, then summarize the summaries.
| Job | Suggested setting |
|---|---|
| Brief my manager before a meeting | 220 words, Professional, paragraph |
| Post a recap in the team channel | 150 words, bullets |
| Turn a client call into action items | Meeting notes, bullets |
| Draft a LinkedIn takeaway | 80 words, LinkedIn post |
| Decide whether a paper is worth reading | 80 words, Report |
Up to 15,000 characters in one run, roughly 2,500 words. Longer documents are split into sections internally and merged. A PDF's text can be pasted in the same way.
It is instructed never to add facts, and in practice it stays close to the source. It can still compress two similar statements into one and lose a caveat, so check anything you will act on.
Yes β put them in the 'Must include' field. It is the most reliable way to stop a number or a product name from being dropped.
No. It is processed to produce the summary and cached in memory for at most six hours so repeat runs are instant. Nothing is stored long-term and nothing is used for training.
It will summarize them, but the output is English. Quality is highest when the source is already English.