Trendwell grew out of my effort to make health data easier to review over time. One way to use its output is to bring a report into an external AI conversation. The useful part of that workflow is being able to inspect the evidence behind an answer.
A structured input helps keep dates, values and limitations together. It does not make an AI response medically authoritative, and asking a model to act as a specialist does not give it that specialist’s judgement.
Keep the report and the interpretation separate
Trendwell generates descriptive reports on your device using defined rules. The values, classifications and stated limitations in that report are Trendwell’s output.
When another tool explains those findings, compares reports or suggests an explanation, it creates a separate analysis. It should preserve the original findings and make its own reasoning visible. Trendwell has not validated that response.
Begin with a question the material can answer
“What changed between these two reporting periods?” is a useful starting point. So is “Which parts of this comparison have weak coverage?”
“How much should I train tomorrow?” asks the tool to make a decision that a descriptive wearable report cannot establish. A confident answer does not bridge that gap.
You can copy text from a Trendwell report into your chosen service. Give it the relevant periods, retain the dates and units, and add only the dated context you want to share. Separate your notes from the report so the model does not mistake an observation you supplied for a measurement from the app.
Ask for the limitations before the explanation
A week with fewer recorded nights is not automatically comparable with a week of fuller sleep coverage. Two weekly reports can overlap. A value included in a report may be a latest-known measurement from an earlier date.
Those details can change what an apparent trend means. Ask the model to identify them first, preserve the reported metric definitions, and distinguish a change in the data from a possible explanation for it.
For each material conclusion, it should be possible to identify:
- what the report actually states;
- any calculation made from the supplied values;
- the model’s interpretation;
- what would be needed to test a possible explanation.
Conflicting signals should remain visible. The model should not average them into reassurance, treat missing data as normal or quietly replace a Trendwell classification.
Check the answer against the source
Keep the original report. Check the reporting periods, units and important numbers in the response. Look for invented history, unsupported causes and interpretations presented as measurements.
For example, if a report shows a change in heart rate alongside a change in recorded activity, a tool can describe that coincidence. It cannot establish from the coincidence alone what caused the change or whether it is medically important.
The same applies to stability. An absence of supported instability in the available recordings is not a claim that every moment was observed. A separate medical test describes its own observation period; it does not retrospectively validate months of wearable interpretation.
Choose deliberately what to share
Trendwell’s report generation is on-device. Pasting text or attaching a file to another service is your separate choice, subject to that service’s data handling. A report may contain sensitive health information. Share only what is useful for the question and review the recipient’s privacy settings.
An external model can help organise a comparison and expose unanswered questions. The value is in an answer you can check, with uncertainty intact, not in a medical persona, a reassuring tone or a specific instruction about what to do next.