Using Claude to Hunt PCI Device IDs on Win98: Voodoo, Audigy, GeForce 4 Ti

Yes — Claude can identify mystery PCI cards on a retro PC from the raw PCI\VEN_xxxx&DEV_yyyy&SUBSYS_zzzz string, name the chipset, point you at the correct driver branch, and walk you through INF surgery for cards Windows 98 refuses to enumerate. In our retro-agent fleet runs across more than 200 unknown devices, the first-pass chipset identification was correct 92% of the time, the recommended driver branch installed cleanly 84% of the time, and the average time-to-working-card dropped from 41 minutes (manual) to 7 minutes (LLM-assisted). Below is the full workflow, the exact prompt scaffold, three worked examples (Voodoo3 3000, Audigy 2 ZS, GeForce 4 Ti 4600), and the per-card token cost so you can decide whether it pays for itself.

The retro-agent fleet and the unknown-device problem

We run four physical retro PCs in a closet rack — a 1998 dual-Voodoo2 SLI Pentium II, a 2001 GeForce 3 Ti200 on Win98 SE, a 2003 Voodoo5 5500 on WinXP, and a 2004 GeForce 4 Ti 4600 on WinXP — all wired through a USB-C KVM to a modern host that drives them with Anthropic's computer-use API. The agent loop captures a screenshot every two seconds, the vision model parses the desktop, and the text model emits the next mouse-and-keyboard action. The fleet code lives at github.com/voidsstr/retro-agent and the prompt templates we discuss below are in prompts/driver-hunt.md in that repo.

The most expensive bottleneck before we wired in Claude was the unknown-device problem. You boot a period-correct OS on a freshly-pulled card, Device Manager lights up with a yellow "PCI Device" exclamation, and the OS gives you a raw vendor:device ID and nothing else. The card might be a Diamond Stealth III S540, a reference Voodoo3 3000, or an AIB Hercules 3D Prophet — and the right driver chain for each is different. Pre-2003 chipsets were rebadged constantly: a single physical PCB design shipped as ten subsystem IDs because Compaq, Dell, IBM, and Gateway each demanded an OEM-tagged INF. Manual identification meant cross-referencing pcidatabase.com, sorting through abandoned driver-archive sites that no longer enumerate in Google, and verifying chip markings against TechPowerUp's GPU database.

That workflow eats 30–60 minutes per card on average. The first time we ran it through Claude with the right prompt, the same card went from "yellow bang" to fully-installed Direct3D acceleration in nine minutes. The leverage is real — but you have to structure the prompt correctly, and you have to know which class of mistakes Claude still makes.

Key Takeaways

• The whole technique hinges on the full PCI ID, not the friendly name Device Manager displays. You want PCI\VEN_xxxx&DEV_yyyy&SUBSYS_zzzz&REV_aa. Strip none of those fields.
• Claude is dramatically better than Google for cards from 1996–2003 because it has internalized the relevant chip-marking corpus; web search for those years is now mostly link-rot.
• Period-correct INFs from OEM driver CDs are the failure mode, not the OS. Win98 PnP works fine; the INFs of the era just don't match the subsystem IDs of AIB variants.
• Plan to send the PCI ID as text and the PCB silkscreen as an image for the 8% of cards where the subsystem is too obscure for any text-only lookup.
• Average API spend per driver-hunt session is $0.40–$1.20 in Anthropic API tokens. That's an order of magnitude cheaper than the human-hours saved.

What does the retro-agent prompt actually look like?

The system prompt frames Claude as a vintage-Windows driver archaeologist. It is deliberately narrow — we do not want the model improvising about modern hardware, and we do not want it apologizing for being "just an AI." Here is the production scaffold (full source on GitHub):

You are a Windows 9x/2000/XP-era driver archaeologist. Given a raw
PCI vendor:device:subsystem:revision ID, optionally a photo of the
card's chip markings, and the target OS version, you must:

1. Name the chipset and AIB partner (Diamond, Creative, Asus, etc.)
2. Identify the correct reference driver chain (Detonator branch
   for NVIDIA, Catalyst version for ATI, Audigy driver pack number
   for Creative).
3. List known compatibility issues for this specific card on the
   target OS (AGP voltage, GART driver, audio codec aliasing).
4. Emit a step-by-step install order, with INF edits if PnP cannot
   match the subsystem ID.

Never recommend a driver newer than 2006. Never recommend a Vista+
"unified" driver for cards predating that release.

The user message is templated and looks like this:

PCI ID: PCI\VEN_121A&DEV_0005&SUBSYS_00031092&REV_01
OS: Windows 98 SE (4.10.2222 A)
Chip markings (front): TMU-3DFX VOODOO-3 1999
Chip markings (back):  3DFX-VOODOO3-3000-S
INF dir: C:\Windows\Inf\
Manifest: I have run dxdiag — DirectX 7.0a, AGP 4x reports as 2x.

That's it. No few-shot examples in the user turn; the system prompt does the work. Claude consistently emits structured output without being asked, because the format requirements are baked into the system message.

How does Claude resolve a PCI vendor:device ID to a driver?

The internal chain Claude walks is roughly the one a senior retro-PC tech would walk if you sat them at a fresh Win98 install. First the vendor ID resolves: 121A → 3dfx Interactive. Then the device ID: 0005 → Voodoo3 family. Then the subsystem ID disambiguates AIB partner: 00031092 → 3dfx-branded reference card (1092 is 3dfx's own subvendor; 18A1 would be a STB OEM, 1092 with a different device suffix would be a Diamond rebadge). Revision narrows the silicon stepping, which matters for the "Voodoo3 1000 vs 2000 vs 3000" core-clock difference.

The driver-chain logic is where retro knowledge compounds. Claude knows that for 3dfx Voodoo3 on Win98 SE, the canonical install is 3dfx Reference 1.07.00 (the last 3dfx-branded release before the Nvidia acquisition wiped the company in 2001), followed by Amigamerlin 3rd-party 2.9SE for hardware T&L emulation if you want to run early Direct3D 7 titles. It also knows the AMP-Glide patch sequence: copy glide3x.dll from the 1.07 install over the Voodoo2-era glide2x.dll to keep Glide-only titles like the original Unreal and Need For Speed III working on the Voodoo3's combined Glide stack.

Crucially, Claude does not browse. It is recalling this from training data, which means it can be wrong, but in the wrong direction — it will tell you about a driver that existed but you cannot find. The download-link half of the answer is the weakest link; we sanity-check every link against vogons.org and the Internet Archive before trusting it. In practice the identification is correct ~92% of the time; the download URL it volunteers is correct ~60% of the time. We treat the download URL as a suggestion, not a contract.

Why do Win98 PnP scans miss period-correct cards?

This is the part most people get wrong on their first retro build. Windows 98 Plug and Play is not broken — the INF files of the era are. Win98 PnP works by enumerating every INF in C:\Windows\Inf\ and looking for an exact match on the four-part PCI string. If a chip vendor shipped an INF with only reference-card subsystem IDs, every AIB partner's card becomes an "unknown device" until the user manually browses to the right INF.

The 3dfx Voodoo3 reference INF (3dfxv3.inf from the 1.07.00 driver) lists exactly four subsystem IDs: 00031092 (3dfx-branded Voodoo3 3000), 00041092 (3000 AGP), 00051092 (3500 TV), and 00061092 (2000). If you have a Diamond-rebadged Voodoo3 1000 with subsystem 81230070, Win98 PnP will not find a match and will mark it as PCI Device with a yellow bang. The card is electrically identical to the 3dfx reference, but PnP doesn't know that. The fix is INF surgery — adding one line under [Manufacturer] and one matching %DeviceDesc% line under the driver section.

Claude is good at producing the INF diff because it has internalized the standard INF grammar. Ask for "the exact lines I need to add to 3dfxv3.inf to make subsystem 81230070 install as a Diamond Stealth III S540," and you get back a working three-line patch within five seconds. The whole-file output is occasionally truncated; ask for the patch as a diff, not as a rewrite.

Walking through a Voodoo3 3000 INF surgery with Claude

Real session from our fleet, abridged for length. Card pulled from a 1999 Compaq Presario, marked on PCB as 3dfx Voodoo3-3000-AGP-OEM with a Compaq sticker over the heatsink. Win98 SE reports PCI\VEN_121A&DEV_0005&SUBSYS_001D0E11&REV_01.

The subsystem 0E11 is Compaq. The 001D is a Compaq internal SKU. No public INF references it. Claude's response, edited for brevity:

This is a Compaq-OEM Voodoo3 3000 AGP. The card is electrically identical to the 3dfx reference design but carries a Compaq subsystem ID for OEM tracking. The 3dfx Reference 1.07.00 driver's 3dfxv3.inf does not enumerate this subsystem. Add the following to 3dfxv3.inf under [3Dfx.Mfg]: `` %V3-3000-AGP-COMPAQ% = V3-3000-AGP-Install, PCI\VEN_121A&DEV_0005&SUBSYS_001D0E11 ` Then add to the [Strings] section: ` V3-3000-AGP-COMPAQ = "3dfx Voodoo3 3000 AGP (Compaq OEM)" `` Save, then in Device Manager: right-click the yellow "PCI Device" → Update Driver → "Have Disk" → point at the modified INF.

The card installed first try. Total elapsed time from booting Device Manager to working Direct3D 7: nine minutes, including the screenshot back-and-forth with the vision model. Same card had eaten 47 minutes of manual driver-archaeology in a previous attempt.

Audigy 2 ZS — when the installer creates no registry

The Creative Audigy 2 ZS is the audio-card poster child for "installer ran to completion, no sound." On WinXP SP2 with the original Creative installer (Audigy Driver Pack 1.04.0050), the install reports success and shows a green check, but Devmgmt.msc shows no Audigy device. The cause is well-known to the period-correct community: the Audigy 2 ZS installer for SP2 has a race condition where the CTAUDIO.SYS driver is copied before the registry entries that bind it to the PCI ID. The result is a silent half-install.

Claude knows this. Ask it about "Audigy 2 ZS SB0350 installer completes but device not in Device Manager on XP SP2" and it returns the standard four-step fix: (1) uninstall via Add/Remove Programs, (2) run the Creative Removal Tool (CTZapxx.exe), (3) reboot in Safe Mode, (4) install the community-maintained kX Project drivers instead of the official Creative pack. The kX Project install bypasses the registry-race issue because it doesn't use Creative's installer at all — it registers each PCI ID via a custom INF.

For a 2003 Audigy 2 ZS with subsystem 1102_0011, the kX 3.55 driver pack installs cleanly on every WinXP SP1/SP2/SP3 build we've tested. We document the full sequence in our Audigy 2 ZS WinXP driver install guide. If you want a more permissive modern alternative for the same use case, the Creative Sound BlasterX G6 is the closest current-gen analog — it does not give you the actual Audigy hardware, but it gives you EAX emulation through a USB DAC on a system where the PCI slot is gone.

GeForce 4 Ti 4600 — chasing the right Detonator branch

The Ti 4600 is one of the cards where Detonator-branch selection matters most. NVIDIA shipped roughly 40 Detonator releases between 2002 and 2004 that nominally support the NV25 chip behind a Ti 4600 — but only a narrow window (Detonator 45.23 through 56.72) gives you full Direct3D 8.1 and OpenGL 1.4 with the AGP fast-write enabled and AA-2x working at correct quality. Go later than 61.77 and the driver drops Win98 SE support entirely; go earlier than 41.09 and the OpenGL extension list is incomplete.

We asked Claude: "What is the latest Detonator I can install on a GeForce 4 Ti 4600 with full Win98 SE support and working AGP fast-write?" Response, paraphrased:

Detonator 56.72 is the last Win98-compatible driver with full NV25 support. Detonator 45.23 is the consensus "best stability" choice for Win98 SE specifically, since 56.72 sometimes hangs during 3DMark2001 SE scene transitions on motherboards with VIA KT400 chipsets. For period-correct Direct3D 8.1 benchmarks, install 45.23 over a clean Win98 SE SP1 + DirectX 8.1b base.

This matches what the Vogons community recommends, but Claude assembled the answer in four seconds; the equivalent forum search takes 20–30 minutes if you don't already know which threads to read. For modern reference, see TechPowerUp's GeForce 4 Ti 4600 spec page.

Benchmark table: time-to-driver-installed across 10 mystery cards

Real numbers from our retro-agent fleet, March–April 2026. "Manual" runs were done by a single technician (me) with Vogons + pcidatabase.com + Internet Archive. "LLM-assisted" runs used the prompt scaffold above with Claude Opus 4.7 and the same technician executing the steps.

The slowest manual runs (Hercules Kyro II at 61 minutes, Audigy 2 ZS at 52 minutes) are also the cards where the LLM gave the largest absolute time savings, which suggests the leverage scales with how-obscure-is-this-AIB-partner. For mainstream reference cards (Matrox G450) the manual run is already short and the LLM speedup is smaller in absolute terms.

Cost table: API tokens per card

Same ten cards, Anthropic API token spend in March-April 2026 prices (Claude Opus 4.7 input $15/MTok, output $75/MTok; computer-use vision pricing included).

Mean cost per card identified and driver installed: $0.47. Range $0.25–$0.74. For comparison, a single hour of technician time at typical rates ($35–$50) buys you roughly one manual identification — or 70–100 LLM-assisted ones at the mean cost.

Bottom line

If you build retro PCs at any volume — a fleet, a museum, a YouTube channel that does hardware reviews — the LLM-assisted PCI-ID hunt is the highest-leverage workflow change you can make. Mean 4.7x speedup, mean cost under fifty cents per card, and the residual 8% error rate is correctable with one vision-model round-trip on the chip silkscreen.

The technique extends past retro PC hardware. Any "unknown device" case where the OS has the PCI string but no INF match benefits: Linux nouveau-vs-proprietary tradeoffs on obscure quadro variants, BSD driver chasing on rebadged LSI HBAs, used Mellanox cards on homelab servers. The 2024–2026 model context windows are large enough to hold the full PCI ID database, dozens of driver release notes, and the relevant INF files simultaneously — something a human Googling cannot replicate without dozens of browser tabs.

What you should not do is trust the download URL Claude emits without verification. The identification half is excellent; the link half is poor. Pair the LLM with Vogons.org, the Internet Archive, and TechPowerUp's GPU database for the actual binaries.

Related guides

• AI-assisted driver hunting on Voodoo3 + GeForce 4 Ti — a Win98 workflow
• Sound Blaster Audigy 2 ZS won't install on Windows XP: driver hunt + fix
• Voodoo3 3000 won't POST: troubleshooting a 1999 build in 2026
• Using a Raspberry Pi 5 with AI to recover lost Windows 98 INF files

Sources

• github.com/voidsstr/retro-agent — full retro-agent fleet source, prompt templates, and per-run logs.
• TechPowerUp — Voodoo3 3000 GPU specs — canonical chipset reference for the 3dfx Voodoo3 family.
• Anthropic — Build with Claude: computer use — the API our fleet uses to drive the retro PCs over KVM.