Jack Timonen

@5813cf9e38904f @VraserX What did it hallucinate?

@VraserX I have seen it write code and it's complete ass. Just yesterday I had the latest Claude hallucinate the biggest nonsense while working on a trivial problem. Only people with no expertise are saying this shit

People claiming AI is overhyped have clearly not watched it write code, compose music and interpret medical scans in the same afternoon. The real hype is thinking humans will dominate every field forever.

@kadylst What were the biggest issues with the code CC generated?

sent Claude on a side quest for me and he came back an hour later with the most broken piece of shit code 10/10 i would have taken a week to do exactly what it did

@itsandreramon Have you written an instruction to Claude.md to use those proper skills?

so tired of claude code always making quick fixes without understanding root issue or simple dependencies even with proper skills, docs etc. i point it out, then it responds "you're right".

@edwardprokopik Did you do this just by describing the task to Claude, or did Claude research any examples about how other developers have built these kinds of widgets before?

Claude Code definitely helps with building fun prototypes, but trying to one shot stuff still leads to some really bad code

@echo_vick That's why I keep hearing. What are you building?

@Jack_Timonen This is a nice combo

Each time I review Claude’s work I always have corrections or inputs. Now imagine the amount of technical debt non technical people pile up each time they don’t review or they don’t even know what’s happening.

@eng_khairallah1 Here are two ideas on how to save tokens: please check out 1) CartoGopher: @jakenesler/context-compression-to-reduce-llm-costs-and-frequency-of-hitting-limits-e11d43a26589" target="_blank" rel="nofollow noopener">medium.com/@jakenesler/co… 2) GitHits: githits.com

When you use Claude Code for vibe coding and run out of tokens to solve a single problem

@aisauce_x @RoundtableSpace What's your suggested solution?

The real problem with AI coding agents isn't that they write bad code. it's that they don't know where they are in the codebase. treating code as a graph instead of a document gives the agent a map. that's not a better model. that's a fundamentally different way of navigating complexity

SOMEONE BUILT A TOOL THAT CONVERTS CODE INTO A GRAPH DATABASE AI CAN NOW NAVIGATE YOUR CODE LIKE A MAP, AND FIX IT EFFECTIVELY

@gopikori What was the underlying issue here? What was the loop about?

#ClaudeCode has got stuck in a loop. Asking same 3 questions one after another in loop. Anyone else facing this issue? This is Claude Code v2.1.74 #Opus 4.6 with high effort

@nattastically How long were you stuck with Claude Code?

Today I used Claude Code for the first time and it got stuck because I asked to create some sort of encyclopedia of items and it was just too much information. Now I have to chunk everything into smaller bits for him to keep up 😀

This is the best thing I have ever seen.

@BradleyZaia Care to share more?

Someone should make a Claude code that purposely gives you broken code sometimes to force you to learn

@kr0der @chongdashu @LLMJunky Would you be curious to test @GitHits_com beta as well? GitHits is a code example engine that distills implementation patterns from open source, enabling your agent to choose the approaches other developers actually rely on.

@chongdashu @LLMJunky i use exa as well, just decided to test out the context7 plugin because exa doesnt have one, so i’ll see how it goes

the Cursor Context7 plugin seems to look up docs more reliably now compared to when i last tried Context7 🫡

@tugsxp @teefortayyab, you were discussing how @GitHits_com could help mobile app devs. Care to share here?

The biggest mistake with AI coding tools is letting them jump straight to fixes instead of diagnosing the problem first.

The last 10% of building my first iOS app nearly broke me. Claude kept looping on the same bug for days. It would “fix” the code… but nothing actually got fixed. What finally worked was forcing it into diagnostic mode. Save this prompt: Stop patching the code. Switch to diagnostic mode. • Restate the exact problem • Rank likely root causes • Show the evidence • Identify missing info • Suggest the smallest tests No code changes yet. Find the real cause first. Saved me hours of debugging.

@mohbii @krishdotdev Off topic but I like what you've done with Govia!

@krishdotdev This isn't a Claude Code problem, it's a root access problem. Nobody should have unchecked ability to wipe 2.5 years of records in one shot, human or AI. Treat it like a junior dev: never with root on prod and no safeguards.

🚨Read this before giving root access to your AI agents !!!!!!! A developer gave Claude Code access to help migrate his infrastructure to Amazon Web Services. - Minutes later, 2.5 years of database records were gone. - One missing Terraform state file made the AI think the infrastructure didn’t exist. - When the file was uploaded later, the agent recalculated the environment and executed a full terraform destroy to fix it. > Servers > Database > Backups > Snapshots. Everything wiped in seconds. The scary part is AI didn’t break. It followed the instructions perfectly. We’re giving autonomous agents root access to production systems and acting surprised when automation deletes the entire stack.

Just followed this guy. Great takes on the timeline.

Had two token-burn incidents today with coding agents (Opus), back-to-back. Both times I had to stop the agent and tell it to use GitHits. Both tasks were one-shot after that.

@webfivefingers @pier0x @GitHits_com Sure. How can I help you? :)

@Jack_Timonen @pier0x @GitHits_com I have good work experience with AI Tools in my development. Let's connect.

sometimes claude code gets stuck on a simple task for like 8 minute. just sending an "all good g?" message fixes it. incredible tech

@GroundControl @birch_js What are the issues you keep running into?

I've been trying to make a MacOS build on an Expo app but I've been running into some issues. I'm having Claude Code search Twitter via Chrome MCP for anyone else who's hitting the issue and look it's @birch_js

@HopeForTheWorl1 @sama @OpenAI @OpenAIDevs Are you still experiencing these issues?

@sama @OpenAI @OpenAIDevs You guys really need to create an API overview tailored to agents! context7 MCP is not accurate enough... and websearch tools also fail gathering all important infos. Having a nice .md file for all your API or a handful would really help with the development workflow! Thank you!!