Suhel Mangera

Job'D the unhinged, satirical, absurd take on the corporate hiring process. sure to give you a chuckle. job-d.vercel.app

example analysis of a Data Operations Manager role at Anthropic from @JobAlchemist_AI free to get started, checkout site at jobalchemist.tech

job hunting already comes with enough guesswork, I built 2 things 1/ JobAlchemist: helps you read between the lines of job ads, understand your fit, and tailor stronger applications. 2/ Job'D: satire for anyone who’s suffered through modern hiring. 1 practical. 1 slightly cooked.

Skills are among the most consequential new tools for AI, and Anthropic just released a very impressive nontechnical Cowork Skill that builds Skills, including doing interviews & providing benchmarks. I think you still need to add the human touch, but this is a big leap forward

@rohanpaul_ai not an official Google repo

🗞️ Google released gws CLI tool on Github (11.6K stars ⭐️), which lets you control services like Drive, Gmail, and Calendar all from one spot in your terminal. GWS connects to APIs for Drive, Gmail, Calendar, and other Workspace services and returns structured JSON that scripts or AI agents can parse easily. Its primarily for building AI assistants. Gives an AI a direct way to read emails or manage calendars without complex coding. e.g. a practical use case could be creating a custom smart assistant that automatically schedules your daily meetings. It automatically generates its own commands by reading live directories instead of relying on manual code updates. Normally devs write complex web requests and read dense documentation to interact with cloud services. GWS eliminates that friction by letting you type simple commands to read emails or create spreadsheets. The clever part is how it outputs everything as structured JSON, a text format easily read by AI models. It includes over 40 pre-built agent skills and a MCP server. This protocol server acts like a universal translator letting AI safely control external tools. You can grant an assistant secure access to your calendar to manage your schedule automatically. How it works - This tool uses a clever 2-step process to figure out exactly what you want it to do. - When you type a command, it first looks at the target word to identify which Google service you are trying to reach. Instead of having a hardcoded list of every possible action, it quickly fetches a live blueprint from Google called the Discovery Service. - This blueprint is basically a massive up-to-date catalog of every single feature and setting that the specific Google service supports right now. The tool then uses that live catalog to instantly build all the available commands and options right there on your local machine. - Finally, it reads the rest of your typed instructions, securely connects to Google, and runs your specific request. - The result comes back formatted as structured JSON, which is just a highly organized text layout that artificial intelligence models and automated scripts can read effortlessly. - This dynamic approach means the tool never goes out of date, because any new feature Google adds becomes available to you automatically.

@catehall @morallawwithin @grok plz explain what's going on this thread, what's implied and left unsaid but it's critical.

(h/t @morallawwithin)

@LLMJunky @dodeja ah makes sense. assume you then explicitly /your-plan-skill Vs /plan mode?

@suhelmangera @dodeja Codex is completely open source so you can actually read their current Plan Mode prompt and learn a lot from it. I used it to start as a base layer for my own custom planning. I more or less just added context7 and dependency maps and left it the same github.com/am-will/codex-…

Codex update 0.106.0 is out! This update gives us a shiny new toy. The "Ask Question" tool enabled in default coding modes! This enables custom plan modes, skills, and "interview me" style brainstorming sessions not possible without this feature. The great news is, if you don't want to use it, it's off by default. To enable it add the feature flag: default_mode_request_user_input = true Agent Memory Updates 105 and 106 are also PACKED with agent memory improvements. Memory now does diff-based forgetting so stale thread memory gets surgically pruned Memory selection is now usage-aware so active high-signal memories are prioritized. I have found memory to be working very well now. Javascript REPL Other update to speak of, Javascript REPL is now under the /experimental menu, allowing for interactive website debugging. Voice Transcription Still no voice transcription for Linux. I have not tested Windows. Please let me know in the comments if you try it. v2 Websockets Fixed responses_websockets_v2 now fully works on its own, including websocket prewarm and request routing. No more error/mismatch behavior. responses_websockets_v2 = true Very well done from the Codex team. Thank you!

@LLMJunky @dodeja is codex team planning 😉 on open sourcing the recommended skill to customize plan mode?

Via skills, and you would just not switch to plan mode. You would just describe how you want Codex to plan step by step in a skill, and launch it that way. Also you'd probably want to turn on the request user input tool in default mode for this, and also describe how you want it to use that.

You know … just a thought. If AI was making me a LOT of money (or giving me a LOT of time back to make the same amount of money) I … wouldn’t be posting about it. And why should I? Folks would then copy me and arbitrage it all away. Stay smart y’all. Profit in the dark.

@thsottiaux @corbin_braun I wish there was chat mode w/ codex. it's difficult to ask questions for info gathering or understanding. 5.3 codex just wants to code. I love the plan mode but I wish it saved the plan somewhere perm automatically like Cursor.

@corbin_braun Codex reacts: unpopular opinion By the time you reach an unpopular opinion post, it’s definitely already mainstream

unpopular opinion GPT-5.3 Codex High > Opus 4.6

Clone a functional Billion $$ SAAS Part 2 - Figma. Now that Claude Opus 4.6 and Codex 5.3 are available, I decided to try this again. The result - CanvasKit Tools used - @windsurf Cascade & @OpenAI Codex CLI First agent run - 30 minutes Total time spent w/ extra prompting - 3 hours I think this could be a new metric for model evaluations - TTBS (time to billion $$ SAAS 😂) Code and details below:

@alistairmcleay @levelsio @AnthropicAI @OpenAI

@levelsio I used Opus 4.6 for the architecture / planning and GPT-5.3 in Codex for the actual implementation With Opus passing context to Codex I love both @AnthropicAI @OpenAI

Just shipped v2 of my flight sim inspired by @levelsio! You can now fly anywhere on Earth: - NY - Hong Kong - Rio - Your hometown - Literally anywhere Speedran this in ~1hr using Opus 4.6 and GPT-5.3 since they both shipped today! fly.alistairmcleay.com

@lennysan @steipete

Email from my 80 year old uncle 🦞

@clairevo @acdlite yeah same. we keep all spec docs in the repo, and auto archive to keep in repo. all handled via spec skill in codex. magic ✨

We do /docs/[feature] in the repo It has - plan - TODO (can hand off between ppl) - docs (sometimes multiple files per feature depending on how beefy On merge these get double checked, added into a changelog, and a version published to our public docs if necessary Specs! As! Code!

Is there any reason I shouldn't put the entire planning document into the commit message...? I'm already incentivized to keep it good so the context survives agent sessions. Might as well preserve it in the git history forever, right?

@alistairmcleay Claude Code is the way to go! 2026 is going to be wild. such an interesting application of CC. goes to show that we grow / limited by own creativity. if you ask the question: "i wonder if Claude can help me with this?" you're miles ahead of others already nice one @alistairmcleay

@Dimillian true. I have copied the suggestions into an existing chat thread and asked to check. and expected review was incorrect. we still have to be awake at the wheels I guess. it's not full FSD yet 😆

Codex review is really good but you have to be absolutely careful not to blindly ask it to fix any issues it finds. It (not very often) think you introduce a regression when it’s actually a new intended behavior.

@Dimillian i respond with continue and it keeps going. tho I have also seen it do a complete review then respond with a list as well.

@suhelmangera Yes they’re amazing but I need them to send much more than just the first issue it stumble upon

Next, let’s kill human review. I don’t want to read code I’m not familiar with. I want to prompt my way into review mode. I want an overview of the code implementation and logical feedback around the diff. I want to ask questions and refine the change but not through the code.

@alistairmcleay curious, if you didn't have cowork what would you have used?

I've been using Claude Cowork for the first time today and I have to say this thing is way more powerful than I expected. I have long-running tasks developing a plan for a very complex project - and it's absolutely killing it.

@alistairmcleay its like playing a game now. one prompt at a time

I really should go outside, but I can't stop coding (orchestrating agents). We're right in the middle of most insane timeline. There is not a moment to waste.