Use cases — what AI got done with PartsPlan
We gave the same AI the same design task twice: once without PartsPlan, once with the same tools plus PartsPlan. The side without PartsPlan still had web search and web fetch, and in the circuit-design and PCB tests also a shell (to download and search whole PDFs) and KiCad. Each result was scored against a checklist sealed before the task, by a grader who did not know which side produced it.
New circuit design — zero major errors, in less than half the time
The AI designed a battery-powered portable pump board with pressure sensing and Wi-Fi from seven specified parts, delivering a schematic, a BOM and a design note.
Major errors 0 vs 3 · 25.6 vs 63.7 min
| Same tools + PartsPlan | Without PartsPlanweb, shell, KiCad | |
|---|---|---|
| Checklist items met (of 30) | 24 | 20 |
| Failed | 0 | 4 |
| Major errors | 0 | 3 |
| Time | 25.6 min | 63.7 min |
What the AI did with PartsPlan
It checked each part's start-up voltage, supply current and monitoring limits against the relevant datasheet pages.
What the AI without PartsPlan missed
The boost converter's minimum start-up voltage (3.0 V), the Wi-Fi module's supply current (0.5 A or more), and the charger ADC's operating limit.
Conditions: Same AI (Claude Opus 5.5) · same task · 90-minute cap · one run per side · the side without PartsPlan could also download and search PDFs from a shell and use KiCad · October 5 (without PartsPlan) and 7 (with PartsPlan), 2026
In a web environment like Claude.ai or ChatGPT — it kept a START ordering rule buried deep in the datasheet
The AI designed a 4-wire Pt100 temperature board with CAN. Its ADC must not have configuration registers written until the START pin is taken high — a rule printed only on pages 28 and 37 of the datasheet.
START order kept · 20 vs 19 items met
| Web + PartsPlan | Web onlyno shell | |
|---|---|---|
| Checklist items met (of 30) | 20 | 19 |
| Failed | 2 | 3 |
| Minor errors | 0 | 1 |
| Same environment, portable pump board — met · failed | 20 · 1 | 18 · 2 |
What the AI did with PartsPlan
It took START high before writing the configuration registers, as the datasheet prints it.
What the web-only AI missed
It wrote configuration registers with START low. On the pump board it sized the Wi-Fi module supply at 350 mA, below the datasheet requirement of 0.5 A or more.
Conditions: No shell on either side — web search, web fetch and file read/write only (the same conditions as the web apps customers use) · same AI (Claude Opus 5.5) · one run per side · October 7, 2026
PCB placement — ahead in all three runs
The AI placed the 49 parts of an STM32 development board. Where each supply pin's capacitor goes, the analog supply filter and the crystal conditions are spread across the datasheet.
Ahead 3 of 3 · average 92.6 vs 81.7
| Same tools + PartsPlan | Without PartsPlanweb, shell, KiCad | |
|---|---|---|
| Average score (of 100) | 92.6 | 81.7 |
| Runs ahead | 3 / 3 | 0 / 3 |
| Average time | 21.4 min | 32.2 min |
| Average AI cost | $4.19 | $4.86 |
What the AI did with PartsPlan
It followed the per-pin capacitor placement, the analog supply filter and the crystal placement conditions, citing the datasheet pages.
What the AI without PartsPlan missed
It followed page 36's "4.7 µF on the VDD3 pin" in only one of three runs.
Conditions: Same AI (Claude Opus 5.5) · three runs per side · the side without PartsPlan also had a shell and KiCad · automatic scoring of overlaps and datasheet placement requirements · October 4–5, 2026
Footprints — 2.5× faster, and accurate even when the maker's site was down
The AI drew the manufacturer-recommended land patterns for six parts. The dimensions are in each datasheet's package drawing.
10.1 vs 24.8 min · 59.3 vs 58.3 points
| Same tools + PartsPlan | Without PartsPlanweb, shell, KiCad | |
|---|---|---|
| Average score (of 60) | 59.3 | 58.3 |
| Average time | 10.1 min | 24.8 min |
What the AI did with PartsPlan
It opened the archived package-drawing pages directly and copied the dimensions. For the part whose manufacturer site was down, it got the dimensions right in all three runs from the archived drawing.
What the AI without PartsPlan missed
In the run where the manufacturer's site did not respond, it filled in dimensions from memory and marked them "not verified" itself.
Conditions: Same AI (Claude Opus 5.5) · three runs per side · the side without PartsPlan also had a shell and KiCad · October 5, 2026
Datasheet question tests
All 30 graph-only values right in six of six repeats; all 12 hard design questions right in four of four runs; all six hidden rating violations in a 30-part BOM found in 1.9 minutes.
- In every case PartsPlan does not judge values for the AI. It finds the pages, tables, graphs, drawings and condition sentences the AI needs in the manufacturer's datasheet and puts them in front of the AI, which designs from that evidence.
- These are small tests (one to three runs per side). They do not guarantee the same results for other parts or tasks.