Vineeth

Announcing, 50% sale of my app : AuraTracker:Habits & Goals, thru the month of June. AuraTracker is one productivity app that replaces the usual stack of habit tracker, reminders, goal planner, task manager, focus timer, and journal with one coherent workflow that helps you set your goals, plan your day, do the work, and reflect. No ads, no subscriptions, account setup needed. Purchase once and own it forever. More about the app : gitgudapps.com/auratracker iOS store link: apps.apple.com/us/app/habits…

Now that my weekly limits are reset, trying a split in my Claude workflow this week: 1. A cheaper, faster model for routine edits, boilerplate and brainstorming (ex: sonnet) 2. Opus 5.5 only for design calls and complex features or really hard to detect bugs. 3. Use antigravity for all other work (ex: product website modifications) Will report back next week, if it helped.

Some safe practices when using coding agents to do autonomous tasks: 1. If your coding agent can open a browser, sandbox it. Run agent browser and MCP tools under a separate user in your OS or in a container. 2. Don't give it terminal accessibility or full-disk permissions. 3. If you use Playwright/Puppeteer for automation, keep the browsers updated, because they cache old Chromium builds.

Some guilt-free weekend indulgence, taking a break from work. Found this gem hiding in my steam backlog since long. thought of taking it out for a spin. The visuals/gfx is so good (despite the dystopian setting), most of the time I feel like stopping and take photos like this instead of progressing through the quests. Game : Metro: Exodus

Here's a system design hack if you are using AI API calls in the backend : Not all the AI features need to be in the request path. If the answer doesn't depend on the previous context (for ex: what the user typed n seconds ago), generate it ahead of time in a batch job and cache it. Cheaper, faster, and nothing breaks when the model API has a bad day.

Small thing that saves hours when you are building an iOS app: add a "do not touch" list in CLAUDE.md. Entitlements, Info.plist, signing settings, the data model. Claude still suggests changes there, but now it asks first instead of "helpfully" breaking the build.

Hot take: your first RAG app probably doesn't need a vector DB. Start with keyword - elastic search (apache lucene) or a hybrid cloud service like azure cognitive search plus some 20 test questions. Add embeddings only when the evals tell you to, don't follow tutorials blindly.

Not sure how many of you know this, but find your airpods on your iphone now features a "near signal/far signal" kind of hot/hotter/cold way of searching for it. it even allows you to remotely beep your airpods. I had misplaced it today and was able to find it in a jiffy, ngl, it felt like a video game 😂

I think the biggest challenge vibe coders / indie devs going to have in the coming years is the affordability, with increasing RAM prices and inflation everywhere, running infra is going to eat at your margins. Was reading a post about the Google head honcho saying that they are sacrificing cloud compute for the short term to make space for AI. This is not a good trend, what will the AI apps that you build would run on ?

"Human in the loop" isn't the safety net people think. In a recent Anthropic's study, expert analysts blindly adopt AI recommendations over 80% of the time, without even realizing it. What this means in plain terms is, when you're reviewing AI output, you're already anchored (also called subconscious bias). Interesting times ahead :) link to the study in the first comment below.

If you are using claude code to debug your infra issues, here's a tip (for azure specifically, but might apply similarly to AWS & GCP too) : When you want Claude Code to debug Azure issues for you, give it a Reader role and nothing else. It can run az commands, read configs and logs, and tell you what's wrong and even get suggestions for the fix, but be patient and apply the fix yourself. An agent with Contributor access on your subscription is one hallucination away from completely messing up your resources.

Quick question for indie devs: How much % of your time do you allocate to building vs marketing on a normal work day ?

If you are someone who has a LLM API running on the backend of your mobile app, this architecture/design choice might help you switch models easily : 1. Add one API gateway in front of every model call. Your app talks to only this endpoint. 2. Behind this endpoint, you can swap models, add caching, log tokens, and set per-user limits without touching app code. Advantage: Model choices can change now indepedent of your app code.

The amount of bookmarks I store vs I actually read is so disproportionate 😂. who else are in the same boat. i save them like every single one of them would change my life, and then forget what i saved in the next one hour !

For those who are blindly building with claude (or any other coding agent for that matter), here's a learning : Yesterday, after i implemented a new feature, Claude Code said the feature was done and all tests passed (like it does usually). While the tests indeed passed, what it didn't say was that it had mocked the exact functions, the test was supposed to check, so the test run was mostly useless. Now I always add a suffix: "show me which tests touch real code paths and which are mocked." Don't blindly trust your coding agent, verify the process it follows once in a while. And btw: using opus 5.5 is an overkill for test scenarios, sonnet 5 is more than enough.

For those who are blindly building with claude (or any other coding agent for that matter), here's a learning : Yesterday, after i implemented a new feature, Claude Code said the feature was done and all tests passed (like it does usually). While the tests indeed passed, what it didn't say was that it had mocked the exact functions, the test was supposed to check, so the test run was mostly useless. Now I always add a suffix: "show me which tests touch real code paths and which are mocked." Don't blindly trust your coding agent, verify the process it follows once in a while. And btw: using opus 5.5 is an overkill for test scenarios, sonnet 5 is more than enough.

Quick tip for people who are using Google Flow, it's more efficient to use the agent mode and specify the duration in the prompt itself. The agent would understand and charge tokens accordingly (ex: 15 for 10 secs, 10 for 6 secs etc). If you do not use agent mode, you will have to choose this manually and sometimes if you forget, you waste a lot. In agent mode, you actually get a confirmation button to review and approve, in chat mode you don't.

Here're some quick claude code habits that has helped me: 1. No code on the first pass. My follow up to a feature dev prompt is always "Write a plan to PLAN.md. Don't edit anything" (you can also switch to plan mode here). Review the plan and have claude execute it step by step. 2. Before any feature, ask it to list 3 ways the feature could break in production, and only then write code. It might add some delay, but the first version comes out with lot of gotchas like retries and input checks that claude would miss normally (which you are forced to add after a bug report.)

A Claude Code habit worth trying: before any feature, ask claude to list 3 ways the feature could break in production, and only then start implementing. This used to be called TDD (Test Driven Development) much before the AI fad. It might add some delay to your shipping dates, but the first version comes out has stuff like retries and input checks, it would normally add (or you are forced to add as part of the bug fix later on) !

Completed shadow of the tomb raider, one of the best games I’ve played. Purely in terms of platforming and puzzle solving this is even better than all the uc games. Can’t believe I got this for free on egs ! The gfx has aged really well and looks awesome, the visuals are spectacular and all the tombs are really well thought through. The one on EGS was the full version, so I got even the DLC tombs! Overall a 9.5/10 game, would’ve been perfect if the non/stealth combat was better !