Two Nails, Two Assistants

Confession first: my tire pressure light had been on for two weeks. I kept telling myself I’d check it tomorrow. Then on Monday I finally looked, and found two nails in one tire and a bubble on the sidewall of another.

A bubble means the tire’s internal cords are broken and it can blow out at any moment. I had been driving on it for two weeks. Nothing happened, and I was lucky. If your tire light is on, go look at your tires today.

So I did what I do now with every problem: I asked AI. Twice, actually.

In one corner was Claude, with web search, maps, email access and full vision. In the other was H_AI, the personal AI engine I’m building on my Mac Studio, running Qwen 3.8 27B. It’s free and private, and it’s very much a work in progress. I asked both the same eight questions, typos included (“contenintal pro contact”).

Round 1

“I got 2 nails in my tire, what do I do?”

A tie. Both said the same sensible things: leave the nails in because they plug the holes, drive gently to a shop, and ask for a proper patch from the inside. On general knowledge, a model running on my desk is just as good as the cloud.

Round 2

The Photo

This was the round that mattered, and Claude got it right. My local model had a small identity crisis. It said it couldn’t see images, then looked at them and read the exact tire model off the sidewall, which is pretty impressive. It even used the word “shoulder.” It just didn’t conclude that a shoulder puncture can’t be fixed.

Read more: why the shoulder matters

Shops patch punctures in the main tread, but not in the shoulder (where the tread curves down toward the sidewall) or in the sidewall itself. The rubber there flexes too much for a patch to hold. On a low-profile tire like mine, the shoulder is very close to where most nails land.

A turn later H_AI did add a caveat: “if either nail is in the outer shoulder, some shops will decline the repair.” By then, though, its headline was still “repair, ~$20–50.” A good shop would have caught it, so no harm done, but I’d have walked in expecting a $30 bill instead of buying a tire.

The “I don’t have vision” moment is a plumbing issue. H_AI’s system prompt lists an image tool, but the photo arrived as a pasted file path, and the model answered from habit before the tool ran. Making image handling automatic is an easy fix.

Round 3

“235/40 R18 91W”

This round went to H_AI. It produced a neat table decoding every character of the tire size (235 mm wide, a 40% sidewall, an 18-inch rim, rated for 168 mph) and explained why a short sidewall leaves less room for a repair. Claude was more practical: it found Continental’s free-replacement plan and noticed the fine print that it doesn’t cover the tires the car came with.

Read more: the size decode and a calculator mishap

H_AI’s decode: 235 mm tread width; 40 means the sidewall height is 40% of the width (about 94 mm); R18 is a radial for an 18-inch rim; 91 is a load index of 615 kg per tire; W is a speed rating of 168 mph.

Claude added the shopping details: two Continental models are sold in this size, they cost about $123–225 online, and I should match the other tires and have the shop check tread depth on all four.

Behind the scenes, H_AI saw “235/40”, decided it was a math problem and called the calculator tool, which promptly failed with invalid syntax. The model shrugged and carried on.

Round 4

“Search the price and best place to fix”

This was my favorite moment of the day. Claude found the tire shops on my street and noted that one closed at 6 and the one next door was open until 7. It was 5:48 PM, so that was exactly what I needed.

H_AI’s search returned… Priceline hotel deals and the Merriam-Webster definition of the word “price.”

Read more: what H_AI actually searched for

The query builder took my question, kept everything after the word “on”, and dropped everything before it, including the part that gave context. The word “tire” never made it into the search:

[web_search: on the price and best place to fix]
- Priceline.com – Deals on Hotels, Flights and Rental Cars
- TOP 10 BEST Auto Repair in San Diego – Yelp
- Price.com: Cash Back & Coupons
- PRICE Definition & Meaning – Merriam-Webster

The model handled it well. It said “I’ll be honest about the limits of the search,” gave sensible price ranges and compared types of shops. Then it threw out the Yelp results as “San Diego, not your location” and asked for my zip code. I was in San Diego. It didn’t know where I was, so it rejected the one useful result.

Round 5

“What if one tire has a bubble?”

Both were great here, and H_AI was arguably the more careful one. Both said: it can’t be repaired, it can blow out, put the spare on. H_AI added two things Claude didn’t say: check all four tires, because whatever caused one bubble may have damaged another, and it asked whether the bubble was on the same tire as the nails. That’s the advice you’d want from a friend who knows cars.

Round 6

“Does my insurance cover this?”

Claude got me the answer, which wasn’t the one I was hoping for. H_AI didn’t get an answer, but I like how it handled it. Its PDF reader handed it 27 empty pages, and instead of making something up, it said so and suggested three other ways to find out. When an AI doesn’t have the information, that’s exactly how I want it to behave.

Read more: the policy details and the privacy angle

Claude also searched my email first, but found only messages with links to documents. From the PDF it found that tires damaged on their own are excluded, nails count as wear and tear, and Emergency Road Service (if it’s on my policy) covers a spare change or a 15-mile tow, but not the tire.

One stumble from H_AI before I sent the PDF: it suggested I might have a “road hazard” add-on that covers the tire minus a deductible. That was a reasonable general answer, but not true for my policy.

The blank pages were scanned images, and H_AI’s PDF reader only extracts text. Sending those pages to the vision model would likely have solved it.

The privacy side matters too. In the H_AI chat, my policy never left the Mac Studio. In the Claude chat, my photos, my policy and a search of my email all went to the cloud. Claude’s answer was right and I chose to share that data, but it’s a real trade-off.

The Scoreboard

Why I’m Not Disappointed

Look at H_AI’s misses again. A search query that dropped half the question. A PDF reader that can’t read scans. No idea what city I’m in. Photos that didn’t go to the image tool automatically. Those are all things I built, and I can fix all of them. The model was mostly let down by the tools around it.

Only one miss was really the model’s: noticing “shoulder” and still saying “repair.” That’s the gap between a 27B model on a desk and a frontier model in a datacenter, and it’s smaller than I expected. Meanwhile H_AI won the tire-size round, gave the more careful bubble advice, refused to guess about my insurance, and did all of it for free without anything leaving my house.

Read more: H_AI’s to-do list and a fairness note

Fixes coming to H_AI: keep the whole question when building search queries; send scanned PDF pages to the vision model; give the model a rough location; send attached photos to the image tool automatically; stop sending tire sizes to the calculator. Then I’ll run the same eight questions again and post the rematch.

A fairness note: the Claude side here comes from a summary Claude wrote at the end of the chat, while the H_AI side is the full transcript, mistakes and all. And the “right answers” I judged against came partly from the Claude chat itself. So the comparison is honest, but not perfectly neutral.

How It Ended

Both tires get replaced, insurance pays nothing, and I’m buying two Continentals. The real lesson has nothing to do with AI: when the tire light comes on, go look at your tires. I had two weeks of luck, and I don’t plan to count on that again.