Getting started
Prompting the agent
How to describe a validation test so you get the right workflow on the first attempt, and how to correct one that came out wrong.
The agent is good at turning a specific test into a sequence and bad at guessing what you left out. Every ambiguity you resolve in the prompt is a step you do not have to fix afterwards.
The pattern that works
A useful prompt states four things. In practice one sentence carries all four.
| State | Example | What happens without it |
|---|---|---|
| The stimulus and its range | “Sweep VIN 1 V to 5 V in 0.5 V steps” | You get a single set-point instead of a sweep, or a step size the agent guessed. |
| What to measure, and with what | “Measure VOUT on the DMM” | You get stimulus with no data. A test that measures nothing is not a test. |
| The conditions | “After a 200 ms settle, with the load at 500 mA” | Readings taken before the rail settles, which look like a failing part. |
| The limit | “VOUT should be 3.3 V ±2%” | Data with no verdict. You have to eyeball the numbers yourself. |
With the load set to 500 mA, sweep VIN from 3.0 V to 3.6 V in 0.1 V steps. Wait 200 ms after each step, then measure VOUT and IOUT. VOUT must stay within 1% of 1.8 V.
Stimulus, instrument, condition, limit. This builds in one turn and needs no correction.
Name instruments by their role
Say “the power supply”, “the DMM”, “the scope”. The agent resolves those against your configured bench. Model numbers work too, and are worth using when you have two of the same type.
- Good, one of each type: “set the supply to 5 V and measure with the DMM”
- Better, two supplies on the bench: “set the E36313A channel 1 to 5 V”
- Bad: “set the source to 5 V” when the bench has a supply and an SMU. Both are sources.
Attach the datasheet
Attaching a PDF changes the quality of the output more than any wording change. With a datasheet the agent takes limits, pin names, and operating ranges from the electrical characteristics table instead of inferring them, and a generated validation plan cites the table it took each limit from.
PDFs, images, and text files all attach from the + button in the composer.
One test per request
Ask for the sweep, check it, then ask for the protection tests. A single message asking for eight tests produces eight approximate tests, and reviewing them is slower than building them one at a time. The exception is a validation plan, where the whole point is breadth: ask for the plan first, then build its tests one by one.
Correcting a step
Say what is wrong, in the same terms you would use with a colleague. Edits are patches, so a correction changes only the step it names.
- “The settle is too short, make it 500 ms.”
- “Measure DC current, not AC.”
- “Add a 100 mA current limit on the supply before you enable the output.”
- “The limit should track the set-point, not be fixed at 3.3 V.”
What not to ask for
Do not ask for SCPI
“Send MEAS:VOLT:DC? to the DMM” works against the design. The catalog action for that measurement already carries the right command for your exact model, plus the parameter handling and the result parsing. Ask for the measurement.
Do not ask for a script
The workflow is the deliverable, and it compiles to Python you can read. Asking for a PyVISA script in chat gets you text you cannot run, edit, or attach limits to.
Do not ask it to make a failing test pass
If a reading is out of spec, that is a result. The agent will report it and will not widen the limit to bury it.
Prompts worth keeping
Line regulation
Set the load to 1 A. Sweep VIN 4.5 V to 5.5 V in 0.1 V steps, settle
200 ms, measure VOUT each step. VOUT within 0.5% of 3.3 V.
Load regulation
Hold VIN at 5 V. Step the electronic load 0 A to 2 A in 250 mA steps,
settle 300 ms, measure VOUT and IOUT. VOUT within 1% of 3.3 V.
Efficiency
Sweep the load 100 mA to 2 A in 100 mA steps at VIN = 5 V. Measure VIN,
IIN, VOUT, IOUT at each point and compute efficiency in a Python step.
Startup
Enable the supply, capture VOUT on the scope with a 20 ms window, and
measure rise time and overshoot. Overshoot under 5%.
Ripple
At VIN 5 V and 1 A load, measure VOUT peak-to-peak on the scope with
AC coupling and 20 MHz bandwidth limit. Under 30 mV.
Thermal soak
Set the chamber to 85 C, wait for it to stabilise, then run the load
regulation sweep. Repeat at 25 C and -40 C.