pvncher Profile Banner
eric provencher Profile
eric provencher

@pvncher

Followers
18K
Following
79K
Media
346
Statuses
10K

building @repoprompt

🇨🇦
Joined June 2013
Don't wanna be here? Send us removal request.
@pvncher
eric provencher
17 hours
Shipping this - I had no idea there were so many alternatives to npm lol. So many weird places to find the Claude installation!.
@pvncher
eric provencher
22 hours
Just shipped @RepoPrompt 1.3.17!.It includes a Claude Code provider, letting you use your configured Claude Code model to handle everything supported by the app. As promised @GroqInc is now natively supported with Kimi K2 as well!. Also shipped a revamped MCP search tool, and.
3
0
14
@pvncher
eric provencher
22 hours
Just shipped @RepoPrompt 1.3.17!.It includes a Claude Code provider, letting you use your configured Claude Code model to handle everything supported by the app. As promised @GroqInc is now natively supported with Kimi K2 as well!. Also shipped a revamped MCP search tool, and.
15
9
148
@pvncher
eric provencher
2 days
I really agree. But also, building ai programming tools is the peak. I can’t explain how essential @RepoPrompt is to my workflows in improving the app. Each improvement also makes me faster at improving it.
@mckaywrigley
Mckay Wrigley
2 days
I get *a lot* of inbound from 17-22yo trying to figure out what to do bc of AI. My advice:. Programming is the API for maximizing leverage in an AGI world. It’s a useful skillset in all fields that takes <20hrs to learn and compounds over time. AI + code gives you superpowers.
0
0
16
@pvncher
eric provencher
2 days
Super aligned with Kevin here.
@kregenrek
Kevin Kern
2 days
I didn't try grok 4 heavy yet, but o3-pro continues to be the most reliable model for writing accurate code and tackling specific tasks in complex codebases. Gemini is still helpful for planning, though it often makes odd or unexpected assumptions. Opus 4 works well for.
2
0
15
@pvncher
eric provencher
2 days
Myth dispelled!. My theory remains that Claude thinks it’s European and is thus lazy in the summer.
@trq212
Thariq
2 days
@iannuttall Thanks for bringing this up, wanted to address this directly!. We don't degrade model performance to better serve demand. The models do sometimes act in surprising ways though and I'd love to DM to understand if you're experiencing an ongoing performance issue. Not diagnosing.
4
1
47
@pvncher
eric provencher
2 days
Extended thinking in an agent loop is only really valuable deep in a conversation because unless you prime the agent with sufficient context upfront, it won’t know enough about what it’s doing to justify the thinking tokens. I also find asking the model to think is bad ux.
@petergyang
Peter Yang
2 days
I have started using "ultra think" and "think ultra hard" in my prompts ever since I interviewed @kieranklaassen . Maybe it's placebo effect but I think it makes AI's answers better 😅.
2
0
9
@pvncher
eric provencher
2 days
error 529. .
3
0
7
@pvncher
eric provencher
2 days
Right when I finally upgraded to the $200/mo plan.
Tweet media one
17
0
107
@pvncher
eric provencher
5 days
RT @ivanfioravanti: Don't underestimate RepoPrompt by @pvncher 👀.With all the 200$-300$ all you can eat LLM offer out there, it's extremely….
0
1
0
@pvncher
eric provencher
5 days
It’s really easy to believe this when your only experience having ai models write code is through an agent harness. Sonnet is truly exceptional and navigating a codebase and solving problems, but given the right context upfront, o3 writes far more elegant code.
@thorstenball
Thorsten Ball
5 days
@BustdownCar @pvncher Sonnet is far better at coding than o3 though.
6
2
48
@pvncher
eric provencher
5 days
I know I’ve said this before, but having o3 pro do a first pass on code changes for new features, is so incredible powerful. Nothing compares, truly.
13
2
145
@pvncher
eric provencher
6 days
Just shipped @RepoPrompt 1.1.15!.Pretty big update, revamping codemap support for Typescript and Javascript, in addition to adding php support!. Also added grok4 and plenty of optimizations.
8
4
87
@pvncher
eric provencher
7 days
I’ve noticed this too, but it’s especially bad in ChatGPT. It doesn’t affect o3 pro at all. Past a very specific threshold of like 55k tokens of context, the model just completely ignores your prompt context, while it used to not. Over api, it seems to have gotten lazier.
@benhylak
ben
7 days
@janbamjan @daniel_mac8 i think it was some sort of optimization, potentially to attention. it seems to only consider or think about the first message.
5
0
35
@pvncher
eric provencher
7 days
If this is the future of ai coding, then forking vscode is tech debt.
@ryolu_
Ryo Lu
7 days
idea:.what if @cursor_ai starts as simple as chatting with one smart Agent, and when you need more power, blooms into a full-blown IDE?. simple up front, limitless when it counts
Tweet media one
13
3
92
@pvncher
eric provencher
8 days
Was a lot of fun as always chatting with @RayFernando1337 . The ai coding meta changes so damn fast that it's hard to keep up. I use these models daily, and I really strive to codify the workflows I rely on into @RepoPrompt. Hope my perspective helps you get more out of LLMs!.
@RayFernando1337
Ray Fernando
8 days
Master Context Engineering Using Repo Prompt with Eric Provencher (Live)
7
2
48
@pvncher
eric provencher
8 days
RT @GitMaxd: Really great live stream just started with @pvncher & @RayFernando1337 . Talking about @RepoPrompt - one of the best opportuni….
0
4
0
@pvncher
eric provencher
8 days
Using seer by @getsentry a bit. So far the root causes it finds are fantastic.The coded fixes are extremely mid, and potentially a source of many regressions.
3
0
25
@pvncher
eric provencher
12 days
Seeing the cursor thing go down, it sort of validates the need in the market for tools that make more efficient use of tokens. Claude can burn tokens for now because of max pricing, but everything else is fighting gravity - being efficient matters a lot!.
22
7
152
@pvncher
eric provencher
14 days
RT @pvncher: @graemeblake @RepoPrompt Well, right now the Repo Prompt chat doesn’t support agentic tool calling for context discovery. It w….
0
2
0
@pvncher
eric provencher
14 days
Claude desktop with @RepoPrompt tools is just much more token efficient!. Claude can plan its context effectively, and then just one shot the problem instead of trying to sub task it. Too much breaking down leads to context rot.
@madebyivor
Ivor
14 days
@pvncher Meant to say: Claude Code failed at the task. It kept breaking things after following its own plan. Desktop + RP one shot it.
2
1
11