Writing good prompts¶
An assistant works best when your request is specific. These habits save time and make the result easier to trust.
Say where things are¶
Name the folder and the files. The assistant can only use folders you gave with --allow-root, and a vague request such as "my caffeine data" costs it several calls to find out which files you mean. Write the files in C:\MSData\batch1 and say which are standards, blanks, QC samples and unknowns. If a sample sheet exists, give its path.
Give the numbers that decide the result¶
The assistant cannot know your method. State:
- the target m/z and the tolerance, with the unit:
m/z 195.0877, 10 ppm; - polarity and MS level, and the acquisition type for targeted work;
- the expected retention time and a search window;
- for calibration: the units, the level of every standard, the fit and weighting you want, and any dilution;
- for QC: the limits. Do not let the assistant choose acceptance limits for you.
If you leave something out, a good assistant asks. If it does not ask and picks a value, read its answer to see which value it used.
Ask for a preview before a change¶
Integrations, batch runs and loads change what is in your window. Say so in the prompt: "Preview the integration and show me the area before you record anything." The tool integrate previews by default and records only when told to. A request such as "Run the method on all samples, but wait for my OK before you export" keeps you in the loop.
Ask for the evidence¶
Add one of these to a request that matters:
- "Show me the points or the picture you looked at."
- "Give me the job ID and the settings you used."
- "List every sample you changed and every file you wrote."
- "Say what is measured and what is your interpretation."
The evidence is in the window too: View log in the status bar opens the AI call log, which lists every call with its exact request while the window is open. For a record that lasts, ask for session_audit.
Ask for a proposal when you want to decide¶
When you want to look at a change before it happens, say so: "If the boundaries look wrong, propose new ones with your reason and evidence, but don't apply them." The assistant then makes a proposal that appears in your window under Review AI proposals, and only your Approve changes the result. See Reviewing AI proposals.
Start small¶
Run the new workflow on two samples before you use it on sixty. Check one area by hand. Look at one chromatogram yourself.
Keep the permissions tight until you trust it¶
Start with only --allow-root. Add --allow-changes when you are happy with what the assistant reads, --allow-exports when you want files, and --allow-project-writes last. See Setup.
Review flagged peaks yourself¶
An assistant can correct boundaries, but it cannot mark a peak as reviewed; that is your decision, made with ✓ Accept peak in the window. Treat the assistant's reading of a picture as a suggestion. The numbers come from Chromascope; the explanation of why a peak looks odd is the assistant's opinion.
Use the right words¶
| Say | Avoid |
|---|---|
| "XIC for m/z 258.1109 at 10 ppm" | "the caffeine peak" without a mass |
| "integrate between 3.11 and 3.36 minutes" | "integrate the peak" when the file has several |
| "preview only" | "check the area" (it may record it) |
| "save to a new file C:\MSData\Reports\x.csv" | "save it" (it needs a path inside an allowed folder, and a new name) |
| "use smoothing 0 for the reported numbers" | "clean up the trace" |
When you get a wrong answer¶
Ask the assistant to name the tool call that produced the number, then compare it with the call in the AI call log (View log). If the settings differ from what you meant, say so and repeat the step. Nothing the assistant did to the window is permanent until you save it; File, then Open viewer session… returns to an earlier state.