TLDR: with a little bit of effort/engineering, Gemini's Gem feature gives you a debug assistant personalized to your company's processes.
DISCLAIMER: I am not affiliated with Google or Gemini in any way. I do like Demis' bald head + Sundar's natural BoredApe look.
I'm a dev at a not so small robotics company. Our project has many, many, many docs split across Github, Google Drive, & the truly accursed Atlassian suite. Different teams struggle to parse each other's lingo and finding the right resources because of this fragmentation. Also Jira search is thoroughly dog shit, so we very often make duplicate issues for known bugs.
To solve this I wanted to make a debug dumpster: one source w/ everything that's ongoing + links to all existing docs. This way you don't have to wade through a vast maze to find things, and you can short-circuit the debugging process.
Key idea: forget about organizing info; just use an LLM as the interface.
Looked around for a hacky solution, but turned out Gemini's already shipped a super simple solution: Gems. It's a feature that lets you create a SHAREABLE modded Gemini tailored for a specific task (essentially just a saved system prompt). Saw it when it first came out, but thought it looked stupid b/c the examples are e.g. a women's cricket watch buddy (real example - no hate on women's cricket btw).
The real value is the "Knowledge Store" which has files from Google Drive. Unfortunately there's a hard limit of 10 files. Here's the workaround:
- You split up the assistant into 2 modes
- a direct question + answer engine like AI Overviews
- a retrieval engine like Google Search
- You allocate 9 to docs that are most common + extensive: these are for the Q & A mode w/ Gemini directly providing info from higher quality docs.
- The 10th doc is a "dumpster" - which should be shared across the team. Here you can either throw quick + dirty notes in e.g. for tips regarding ongoing, transitory issues OR just link to more polished existing docs.
- Now Gemini can surface a link for you even if smth wasn't impt enough to make it into the 9 workhorse resources. Would be great if Gemini could also open the links & find your desired info on its own, but it doesn't have the tools to do that.
- Instead you can take the link and (as long as its in Drive) add that doc to Gemini's temporary chat context to keep asking it questions re the source.
- Most of the dumpster looks like a key:value store e.g. "if the user wants to do X, point them to resource Y [insert full link]".
- Also have to tell the Gem to not use web search & instead rely on the internal docs: otherwise it will start pulling up irrelevant shit/hallucinating when it can't figure something out. This also means that you still want to use normal Gemini for questions that do require external resources.
- It's also great b/c if the Gem doesn't have smth, then you can just find it on your own once, & then throw it into the dumpster for your future self and your teammates.
There's a couple killer features here:
1. Google model so it's obv tuned for fast processing of lots of data + search.
2. Google model so it's obv great at image processing and even OCR natively.
3. Google model so it's obv not well tuned for quick responses + takes way too long on even simple queries. That's suboptimal, but it's still better than taking forever to wade through everything on your own - or even worse, j not finding it.
Figured out Gemini can do this yesterday & polished it up this morning. So far it's been a big hit on my team.
P.S. Thanks to my entrepreneur friend (who actually did YC Spring '25) for telling me: a business doesn't actually have a pain point if they haven't tried fixing it themselves. I still think he's underestimating how much ppl just suffer through shit, and I prolly wouldn't have j tried doing this if he hadn't said that.
garryssaunaboy•2mo ago
DISCLAIMER: I am not affiliated with Google or Gemini in any way. I do like Demis' bald head + Sundar's natural BoredApe look.
I'm a dev at a not so small robotics company. Our project has many, many, many docs split across Github, Google Drive, & the truly accursed Atlassian suite. Different teams struggle to parse each other's lingo and finding the right resources because of this fragmentation. Also Jira search is thoroughly dog shit, so we very often make duplicate issues for known bugs.
To solve this I wanted to make a debug dumpster: one source w/ everything that's ongoing + links to all existing docs. This way you don't have to wade through a vast maze to find things, and you can short-circuit the debugging process.
Key idea: forget about organizing info; just use an LLM as the interface.
Looked around for a hacky solution, but turned out Gemini's already shipped a super simple solution: Gems. It's a feature that lets you create a SHAREABLE modded Gemini tailored for a specific task (essentially just a saved system prompt). Saw it when it first came out, but thought it looked stupid b/c the examples are e.g. a women's cricket watch buddy (real example - no hate on women's cricket btw).
The real value is the "Knowledge Store" which has files from Google Drive. Unfortunately there's a hard limit of 10 files. Here's the workaround:
- You split up the assistant into 2 modes
- a direct question + answer engine like AI Overviews
- a retrieval engine like Google Search
- You allocate 9 to docs that are most common + extensive: these are for the Q & A mode w/ Gemini directly providing info from higher quality docs. - The 10th doc is a "dumpster" - which should be shared across the team. Here you can either throw quick + dirty notes in e.g. for tips regarding ongoing, transitory issues OR just link to more polished existing docs.
- Now Gemini can surface a link for you even if smth wasn't impt enough to make it into the 9 workhorse resources. Would be great if Gemini could also open the links & find your desired info on its own, but it doesn't have the tools to do that.
- Instead you can take the link and (as long as its in Drive) add that doc to Gemini's temporary chat context to keep asking it questions re the source. - Most of the dumpster looks like a key:value store e.g. "if the user wants to do X, point them to resource Y [insert full link]".
- Also have to tell the Gem to not use web search & instead rely on the internal docs: otherwise it will start pulling up irrelevant shit/hallucinating when it can't figure something out. This also means that you still want to use normal Gemini for questions that do require external resources.
- It's also great b/c if the Gem doesn't have smth, then you can just find it on your own once, & then throw it into the dumpster for your future self and your teammates.
There's a couple killer features here:
1. Google model so it's obv tuned for fast processing of lots of data + search.
2. Google model so it's obv great at image processing and even OCR natively.
3. Google model so it's obv not well tuned for quick responses + takes way too long on even simple queries. That's suboptimal, but it's still better than taking forever to wade through everything on your own - or even worse, j not finding it.
Figured out Gemini can do this yesterday & polished it up this morning. So far it's been a big hit on my team.
P.S. Thanks to my entrepreneur friend (who actually did YC Spring '25) for telling me: a business doesn't actually have a pain point if they haven't tried fixing it themselves. I still think he's underestimating how much ppl just suffer through shit, and I prolly wouldn't have j tried doing this if he hadn't said that.