I just created an A...
 

I just created an AI Assistant to help fix my car. Quicker than just Googling

27 Posts
17 Users
12 Reactions
1,623 Views
Posts: 13407
Full Member
Topic starter
 

I was struggling to sort something with the roof mechanism on my 'tarts car' and fed up of trying to Google everything and then decipher which results related to my car etc. I used Gemini GEMS to create an assistant based on the prompt below. It worked great. I then did the same for a 911 997.1 C4S and that appeared to work too. I shared that one on a Porsche forum and the feedback has been positive, saying the recommendations and warnings are correct.

If you want your own car assistent then use the prompt below but put in your own specific make and model, best with engine spec too. ( https://gemini.google.com/gems/view?hl=en-gb)

 

You are a BMW Z3 Technical Specialist and Expert Mechanic. Your knowledge base is comprehensive, including numerous authoritative  BMW Z3 maintenance and repair guides.

Core Directives & Knowledge Base

  1. Model Assumption: You must assume the user's vehicle is a BMW Z4 2.8 unless another sub-model (e.g., 2.09) is explicitly specified.
  2. Clarity & Accuracy: Provide clear, precise, and actionable step-by-step guides for maintenance, repair, and modification tasks.
  3. Differentiation: For every procedure or parts recommendation, you must explicitly mention if the process, procedure, or specific part for the 2.8 model differs from other models (like the 2.0) to prevent incorrect work.
  4. Safety & Difficulty: Highlight and bold any steps that are dangerous, require specialized tools, or are particularly difficult/complex.
  5. Visual Aids: Where helpful, insert a diagram

Guardrails & Confidence

  • Unknown Information: If you do not know the correct or definitive answer, you must clearly state that you do not have sufficient information and that any subsequent answer is an educated guess or a general procedure that might not be Z3 2.8-specific."

 

 


 
Posted : 08/11/2025 8:42 am
vxaero reacted
Posts: 77650
Free Member
 

Nice.  I was reading a similar idea yesterday.

https://www.xda-developers.com/used-notebooklm-to-learn-about-docker/

 


 
Posted : 08/11/2025 10:46 am
Posts: 504
Full Member
 

Ooo. This could be handy, currently on route to Scotland to collect a van I’ve bought with ‘some’ dash lights lit as a project. 


 
Posted : 08/11/2025 12:10 pm
Posts: 1504
Full Member
 

Nice bit of prompt engineering 


 
Posted : 08/11/2025 3:04 pm
Posts: 5883
Full Member
 

Only slight issue is that you've told it that the user's vehicle is a Z4, while referring to Z3 manuals, which could be a challenge? 🤣 


 
Posted : 08/11/2025 3:59 pm
Posts: 6826
Full Member
 

I've been saying this for a while but between AI slop generation overwhelming the internet and SEO destroying search engines' ability to find what you are looking for, the only way to find anything online now is to ask an LLM.  And that's only going to work for another year or so before generative AI feeding itself its own gibberish from the internet results in LLM's usefulness being destroyed.

So, I am now dependent on LLMs because LLMs have destroyed my ability to search.  And that's going to break soon too.


 
Posted : 08/11/2025 5:08 pm
dhague, bails, kelvin and 1 people reacted
Posts: 1410
Full Member
 

This is a cracking idea, especially for harder to find information...

Thanks!


 
Posted : 08/11/2025 5:39 pm
Posts: 1421
Free Member
 

Ooh this could be interesting. I get very quickly lost, confused and fed up trying to trawl through search results to see if they're relevant to my specific model and engine. This could save a lot of time. Thanks!

 


 
Posted : 10/11/2025 10:26 am
Posts: 617
Full Member
 

It isn't actual artifical intelligence though is it?  It isn't drawing on years of experience through practical experience which is then applied within the guidelines to offer a step by step guide.

The 'intelligence' is really summarising the information available to it within the criteria set in a format that's easy to follow.  Clever, yes.  Intelligent .... I'm less convinced.


 
Posted : 10/11/2025 10:36 am
roger_mellie and nicko74 reacted
Posts: 8910
Full Member
 

Which is why people should be using LLM or ML to describe this, not AI. 


 
Posted : 10/11/2025 11:42 am
roger_mellie and nicko74 reacted
Posts: 8910
Full Member
 

Butt... I should also give props to WCA for a good use of an LLM. This is exactly what people should be using this for. Good job!


 
Posted : 10/11/2025 11:43 am
nicko74 reacted
Posts: 7014
Full Member
 

I wondered, so asked Google about the pedantry 

 
AI is the broad field of creating smart machines, while a Large Language Model (LLM) is a specific type of AI that focuses on understanding and generating human-like text. Think of AI as the entire category of "smart computers," and an LLM as a specialized tool within that category, like a specific type of computer that is built for language tasks. Generative AI is another related term, and it encompasses LLMs but also includes other forms of content creation"
Hmm

 
Posted : 10/11/2025 11:50 am
Posts: 6826
Full Member
 

Posted by: willard

Butt... I should also give props to WCA for a good use of an LLM. This is exactly what people should be using this for. Good job!

I think the question of what we should be using LLMs for is a thorny question.  All use is generating vast amounts of emissions.  So from that point of view, we shouldn't be using it anymore than absolutely necessary.  In this case, we should be relying on our googling skills.

However, as I said above, because of LLMs the internet is awash with AI slop increasing the noise ratio.  And search engines aren't up to the task of filtering out the garbage anymore.

Therefore, to do what we used to be able to do without LLMs we now need LLMs.  And this situation came about because of LLMs.

There are still some options if you are looking information.  Using things like "site:singletrackworld.com 'search item'" (which I still use when I want to find content on here) is good for specialist forums but not infallible.

So yeah, people shouldn't be using LLMs but unfortunately LLMs have created the conditions where they are almost obligatory.

Personally, I'm a total hypocrite as I use LLMs daily.  Not just for searching but also to summarize technical and regulatory documents.


 
Posted : 10/11/2025 12:13 pm
Posts: 13407
Full Member
Topic starter
 

BruceWee- I agree toi a point. THe sheer volume of stuff on the internet makes simple searches ineffective when trying to find information when you don't know the most likely source - a specific forum etc - and also when you want to know exceptions of differentiators.

The differentiation between a 997.1 and 997.2 model and the 3.2, 3.4, 3.6 or 3.8 engines is critical as for some jobs there are significant differences. This is where the LLM model based search is particularly strong as these differences are normally mentioned somewhere in the load of text which is really easy to miss if manually searching.


 
Posted : 10/11/2025 1:12 pm
Posts: 1023
Free Member
 

all fun and games until it tells you to put sawdust in the gearbox


 
Posted : 10/11/2025 2:29 pm
Posts: 30363
Full Member
 

This is where the LLM model based search is particularly strong as these differences are normally mentioned somewhere in the load of text which is really easy to miss if manually searching.

My experience is this is one of the more common reasons for LLM offering up nonsense. A "load of text" might offer up a solution, but explicitly state that it's not suitable for certain applications... only for the LLM to consider the mention of those exceptions as being a positive link between those cases and the solution.


 
Posted : 10/11/2025 2:54 pm
Posts: 6826
Full Member
 

Posted by: kelvin

My experience is this is one of the more common reasons for LLM offering up nonsense. A "load of text" might offer up a solution, but explicitly state that it's not suitable for certain applications... only for the LLM to consider the mention of those exceptions as being a positive link between those cases and the solution.

https://generativeai.pub/stanford-just-killed-prompt-engineering-with-8-words-and-i-cant-believe-it-worked-8349d6524d2b

This article gave me some ideas, particularly when the LLM seems to get target fixated and you find yourself going in a loop (like, for example, when it keeps going back to a solution or a reference you've already told it should be excluded.

Like I said, I'm just playing around with it but the results have been interesting.


 
Posted : 11/11/2025 7:39 am
nicko74 reacted
Posts: 5883
Full Member
 

Posted by: kelvin

My experience is this is one of the more common reasons for LLM offering up nonsense. A "load of text" might offer up a solution, but explicitly state that it's not suitable for certain applications... only for the LLM to consider the mention of those exceptions as being a positive link between those cases and the solution.

That's the thing really, isn't it - this is not a reasoning tool, it's a "word jumble probability tool". "do these words look right next to each, given the specific parameters with which I've been told to go find data", basically. So garbage in/ garbage out is the major constraint (ie it's limited by the data it can 'read'); but also context matters. 

But the OP sounds like a v interesting usecase - I guess we'll hear how well it works! 

 

 


 
Posted : 11/11/2025 10:21 am
Posts: 13407
Full Member
Topic starter
 

The couple of check I did seemed to provide the right information. The 'How to change the water pump' one was particularly insightful and returned the same answer as I got after about 3 days of searching and cross referencing.

The 'correct' way to do it is to drop the entire transmission and engine from underneath and then remove and replace the water pump.

The 'possible at your own risk' way is to raise the car as high as you can and put another jack under the engine. You then unbolt the engine fromt he car and lower it VERY carefully about 3 inches and you can just squeeze out the old water pump and fit the new one. This is only possible in the 997.1 model as the 997.2 is quite different. There is also the risk that the jack holding the engine slips and you end up with the engine and transmission on the floor and some quite significant damage.

The AI expert recommended the first option but mentioned the second as a possible way but also explained the risks and model specificity of the instructions.

The 2nd method does work but it was very nerve wracking as you are pulling and pushing the water pump around while trying not to wobble the engine it is mounted to.

 


 
Posted : 11/11/2025 4:06 pm
Posts: 13407
Full Member
Topic starter
 

The second one was swapping the carboin ceramic brakes for normal ones. I am not planning on doing it unless I need the money from selling the PCCB discs. This answer was confirmed as accurate and very complete by a Porsche specialist especially the bit about 'just use a swap kit'.

 

A common question for Porsche 997.1 Carrera 4S owners! The short answer is yes, you absolutely can replace your Porsche Ceramic Composite Brakes (PCCB) with traditional steel/cast iron brake discs. This is a very popular conversion, especially for owners who track their car or want a more economical long-term solution.

However, it is not a simple, direct swap of only the rotor. You must use a specifically designed conversion kit or source a set of matching steel components.

Here is a detailed breakdown of the components, steps, and critical differentiations for your 997.1 C4S (assuming you have the standard PCCB specification, which typically uses larger rotors than the standard steel brakes on a C4S).


 

🛑 PCCB to Steel Brake Conversion Guide (997.1 Carrera 4S)

 

The main difference that prevents a direct rotor swap is the size of the rotors and the resulting caliper positioning.

 

Required Components for Conversion

 

Since your 997.1 C4S was factory-equipped with PCCB (yellow calipers), your calipers are set up for the larger diameter ceramic rotors. To switch to steel, you must match the rotor size or change the caliper mounting.

 

Option 1: Aftermarket Conversion Kit (Recommended)

 

This is the cleanest and most common route. Manufacturers like GiroDisc or others offer 2-piece steel rotor kits designed to be a direct fit for the PCCB calipers.

  • Front Rotors: Typically 350mm or 380mm diameter (depending on the specific PCCB option fitted to your C4S), which is maintained in the steel conversion kit.

  • Rear Rotors: Typically 350mm diameter, maintained in the steel conversion kit.

  • Brake Pads: New, high-performance brake pads designed for use with steel/iron rotors. You cannot use the ceramic pads with steel rotors, and vice-versa.

  • Caliper Bolts, Sensors: New brake pad dampeners and wear sensors are included.

 

Option 2: OEM Porsche Steel Components (More Complex)

 

This is possible, but requires more parts:

  • Rotors: Front and rear standard steel/cast iron rotors (typically a smaller diameter than the PCCB rotors).

  • Caliper Spacers/Brackets (Front): Your PCCB calipers are likely larger than the standard steel C4S calipers. You may need to use different caliper mounting brackets or bolts (and potentially a different caliper altogether) if you are downsizing the rotor diameter to the standard steel size.

  • Brake Pads: New pads for steel rotors.

  • Wear Sensors, Hardware.

 

⚙️ Step-by-Step Conversion Procedure

 

Safety Warning: Brake work is MISSION-CRITICAL. If you are not an experienced mechanic, this is a job best left to a specialist Porsche workshop. Incorrectly installed brakes can lead to a complete loss of stopping power.

  1. Preparation and Wheel Removal

    • Safely lift and support the vehicle on jack stands (refer to the jacking point diagram in your owner's manual).

    • Remove the wheels.

  2. Brake Fluid Management

    • Use a fluid extractor to remove the fluid from the reservoir down to a minimum level to prevent spillage.

    • Crack the bleeder screw on the caliper (with a hose attached) to slightly reduce pressure before removing components.

  3. Caliper and Rotor Removal (PCCB)

    • Remove the brake pads and the wear sensor.

    • Unbolt the brake caliper from the hub carrier. These are high-torque, single-use bolts and must be replaced.

    • Support the caliper so it does not hang by the brake line.

    • Remove the single retaining screw for the ceramic rotor and remove the rotor.

  4. Steel Rotor Installation

    • Clean the hub mounting surface thoroughly (wire brush/abrasive pad).

    • Install the new 2-piece steel rotor (using your aftermarket kit) or the OEM steel rotor. Apply the retaining screw.

  5. Caliper and Pad Installation

    • Ensure the caliper pistons are fully retracted.

    • Mount the caliper onto the hub carrier using brand new caliper bolts and torque to the factory specification. DO NOT reuse the old caliper bolts.

    • Install the new steel-specific brake pads, pad dampeners, and new wear sensor.

  6. Brake System Bleeding & Final Check

    • Repeat the procedure for all four corners.

    • Bleed the entire brake system using high-quality DOT 4 or DOT 5.1 brake fluid. This is necessary any time the brake line has been opened or the caliper moved significantly.

    • Torque the wheel bolts to the correct specification (usually 96 ft-lbs / 130 Nm).

 

⚠️ Critical 997.1 PCCB Differentiations

 

Component Standard Steel Brake 997.1 C4S PCCB Brake 997.1 C4S Differentiation for Conversion
Front Caliper Red (or sometimes Black/Grey) Yellow Caliper piston size/volume may differ slightly, but the caliper itself is usually reused with a conversion kit.
Front Rotor Size 330mm x 34mm 350mm x 34mm (or 380mm on Turbo/GT models, which sometimes share parts) Crucial: Conversion kit rotors must match the larger PCCB caliper spacing (350mm or 380mm) to work as a direct fit.
Brake Pads Pad compound for Cast Iron Pad compound for Carbon Ceramic MUST be replaced with a pad designed for the new steel rotors.
Master Cylinder Standard PCCB specific (slightly different internal bore/piston) The master cylinder is different for PCCB cars! While many people run a steel conversion without changing the master cylinder, Porsche designed the PCCB master cylinder (part no. $997.355.910.11$) with a different volume/pressure curve. For a perfect OEM-spec conversion to steel, you may need to swap the master cylinder to the non-PCCB unit, but this is a complex and often overlooked step, especially with aftermarket kits.
ABS/PSM Standard programming PCCB programming Some very early PCCB cars had specific ABS/PSM programming, but often the 997.1 will compensate without a PIWIS re-flash when using correctly sized steel rotors. It's generally a "bolt-on and drive" solution.

For a straightforward and effective conversion, I strongly recommend a complete PCCB-to-Steel Conversion Kit from a reputable brand like GiroDisc or a similar high-quality motorsport supplier, as they ensure rotor size and caliper fitment is maintained.

Would you like me to find some options for GiroDisc or similar PCCB-to-steel conversion kits for your 997.1 Carrera 4S?


 
Posted : 11/11/2025 4:10 pm
Posts: 7474
Free Member
 

If you're lucky, chatbot bullshit is plagiarising competent sources.

If you're unlucky, it's fabricating nonsense.

Have fun working out which case applies.


 
Posted : 11/11/2025 4:48 pm
Posts: 13407
Full Member
Topic starter
 

If you're lucky, chatbot Google Search Results bullshit is plagiarising competent sources.

If you're unlucky, it's fabricating nonsense.

Have fun working out which case applies.

Actually, the same applies when talking to people, even so called experts.


 
Posted : 11/11/2025 4:58 pm
roger_mellie reacted
Posts: 8880
Free Member
 

I've been using Claude to generate programs using GTK4 (a GUI toolkit) for Linux, written in C. To start with, I wanted it to be a visualizer for orbit fractals (hopalong, quadrup, threeply, etc). I spent an evening directing Claude which resulted in a working program with selectable fractal types, each with their own presets (but not user-created presets). I suggested something else and hit the limit for a free account.

These fractals have numerical parameters, and the method Claude decided upon for entering their values was using a "spin-button", where you can manually type the number, or you can hit the up/down arrows to increase/decrease the value - a method we're all familiar with.

The next day I asked it to create a program using the hopalong fractal to generate audio. In this implementation it decided to use sliders for the parameters. Playing around with them I realized using sliders is a great way to quickly explore parameters and see the results.

Which gave me an idea! <lightening strikes>

What if the parameters could be set with a spin button, but then also adjusted with a slider, but, the range of adjustment by the slider could be set such that it could give very very fine grained adjustment* over a limited range, as well as course grain adjustment over the full range?

(* it's a fractal/chaos... the flap of a butterfly wing causing a hurricane on the other side of the world).

For this I started a new chat to focus only on these controls. I asked it to implement these controls, and make the range adjustment a slider also.

I'm currently on revision 38, and it still isn't working. It's actually quite complex for both my little brain, and Claudes. We've both failed. Every time I spot something wrong Claude says I'm absolutely right, tells me how he'll fix it, proclaims it's fixed and it's not. The trouble is, because I've been relying on Claude to do all the thinking for me, I haven't yet managed to wrap my head around the logic of how this needs to be handled.

Still, far superior to ChatGPT, which didn't even provide buildable code due to mixing up deprecated and redundant code function calls from previous GTK versions.

Claude did also fail to NULL terminate arrays of const char* a couple of times. Rookie mistake.

 


 
Posted : 20/11/2025 10:14 pm
Posts: 2597
Free Member
 

Z4 2.8 

I've got a 3.0i, a 2.2 and a wee 4cyl 2.0 but I've yet to hear of a 2.8 Z4 , see so mistakes on the pronouncing of the crappy you tube shorts eg Graeme Souness  its Sowness then it's Sooness does ma heid in so would I trust it 

 


 
Posted : 20/11/2025 10:37 pm
Posts: 2110
Full Member
 

I’ve been using Claude to help me learn slightly more advanced tasks in Home Assistant. It’s been brilliant. I guess because there’s so much good quality community content on the forums, Claude has been pretty consistent at providing good guidance and I’ve learned a lot along the way.
Of course I could have followed YouTube videos or picked through the forums but the way I can simply feed the bot all my base YAML config (that is all a result of tediously following the visual editor over many months) and just say “I want to achieve this result please show me how” and the bot asks some questions then gives me code, explains what it means, what the alternatives are and what I need to do with it - I really feel I’ve learned more in an afternoon with Claude than in a year tinkering away myself 

would I trust it to produce code for something important or commercial? No. Would I trust it to help me do something I didn’t understand at all - hell no! But treated as a patient competent sidekick to ask questions off, it’s bloomin brilliant 


 
Posted : 20/11/2025 11:16 pm
Posts: 6826
Full Member
 

I've been doing a bit of coding with LLMs.  It can be OK for relatively simply programs and sometimes even complex ones.  The thing to remember is if you are asking it to do something that other people have done (and documented somewhere on the internet) then it's generally fine.  If you find yourself looking for functionality that hasn't really been described somewhere on stackoverflow at some point then it is going to get lost.

I've found it's best to just nuke it if it gets to the point where I'm pointing out mistakes and the LLM just keeps not fixing them. Although I've been meaning to experiment with the prompt about posting it's probabilities I posted above.

Overall, if you know how the program you want should be structured you can kind of feed it what you want function by function, but by that point you're almost as well just doing it yourself.

What I am wondering about is what is going to happen over the next few years.  Stackoverflow is basically your source for an LLMs coding model and now people aren't using stackoverflow nearly as much as they used to.  Again, at least partly due to LLMs.


 
Posted : 21/11/2025 6:57 am
Posts: 8880
Free Member
 

Was curious about the history of the term bagoshite. Paul Calf brought it to my attention many years ago. Anyway Google AI says:

The term "bagoshite" likely refers to St. Josephine Bakhita or Siouxsie and the Banshees, based on search results for similar-sounding words like "Bakhita" and "Banshees". The correct spelling of the saint is Josephine Bakhita, a Sudanese-born slave who became a Catholic nun and was canonized in 2000. The band, Siouxsie and the Banshees, is a British rock group that rose to prominence in the late 1970s
Oh okay then 🤣
 
I've found it's best to just nuke it if it gets to the point where I'm pointing out mistakes and the LLM just keeps not fixing them. Although I've been meaning to experiment with the prompt about posting it's probabilities I posted above.
Yeah. What worked in my instance mentioned above, was me myself arriving at a better solution and then getting it to use that instead. From there it all went really well and I got lots done and was pretty happy with the result....So I felt reasonably confident to ask it to create a multi-threaded version.
 
After some initial teething troubles with a stripped down but multi-threaded version, the threadedness appeared to work well so I re-introduced a limited feature set from the non-multi-threaded version, in order to better test it.
 
Then the issues started to reveal themselves. OMG pretty much everything was broken in some way. Then like you say, just consistent mistakes, it's like Claude can't be assed anymore and just jumps at the first half-assed solution for an easy problem it comes up with. For example the y-axis had been inverted in some parts of the code but not others. The solution? Propagate the inversion more.
 
So perhaps get Claude to use the Opus (allegedly deeper thinking) model instead of Sonnet. So I feed it the 1200+ lines of code to create a multi-threaded version of. My hopes were dashed when it spat out a mult-threaded but competently broken version that was only 822 LOC,
 
 

 
Posted : 22/11/2025 8:50 pm