Michael Isaac

@grichadev OMG I love it. Did something similar yesterday: mpiv.ai/blog/fpk-f-bom…

here’s my little side project, run npx devrage to find your numbers and breakdown. it just reads your transcripts, code is here github.com/gricha/devrage

@asaio87 Sleeping? Who is sleeping?!?

People from Europe need to use Claude code when people from the Americas are sleeping. Its becoming unusable to use Claude during the evenings in Europe

@badlogicgames Right on! pi-ds4?

i'm a certified moron now!

why? why shouldn't i be an absolute moron and buy this?

Bain piece here: bain.com/insights/100-b…

Enterprises have had the same problem for decades: their data, apps, rules, approvals, exceptions, and evidence do not live in one place. They don't talk to each other. ERP didn’t fix it. RPA didn’t fix it. Dashboards didn’t fix it. Copilots or agents won’t fix it alone. AI makes it solvable only when paired with governance. The category is governed integration.

@thdxr We just have boring AI. The best kind.

every single ai product: "we've built revolutionary tech that changes your life using artificial intelligence so say you have a spreadsheet...."

The real benchmark for AI augmented coding is enterprise procurement.

@forgebitz It's still crushing tasks for me in the MacOS app. gpt-5.5 xhigh fast and /goal.. omg.

they nerfed gpt-5.5

@dhh @OpenAI I've not been courageous enough to try anything less than xhigh/fast yet. But the results at these settings... ooh la la.

I've been driving GPT5.5 on low reasoning for the last week+ and it's very good, very efficient. Haven't been tempted to reach for Opus at all. And it's more succinct than Kimi too. Huge leap forward for @OpenAI 👌

@cjzafir Interesting. Do you ask Codex/gpt to use the DS API directly, or through a separate harness?

My current workflow: 1. I have an idea. I open Codex desktop app to plan (without plan mode: overcomplicated). I use Codex 5.5 High (fast). 2. Once I get v1 of the plan, I use this hack: "Are you 100% confident in this strategy? If not, find all possible loopholes, suggest proper fixes, and run this loop until you are factually 100% confident in the new strategy." 3. This finalizes my plan, schema, and file tree, rules and quality gates. Then I use DeepSeek V4 Pro and Kimi 2.6 as executors. So Codex acts as the orchestrator, and these two act as executors. 4. Codex uses computer use to test in the browser, run deep quality checks, and only pass high-quality outputs to production. 5. I make Codex autonomous so it can plan, execute, audit, fix, and continue executing each task. Similar to what /goal is right now. That's it. I just oversee reports, current status, and final outputs. I trust Codex 5.5 when I have robust specs attached. Also, a few tips: - Tell Codex: keep folders, subfolders, and files neat and clean. Well-organized, with no dead code or unused files. - Always make a task list for every item you work on. (This fixes the Codex quitting-in-the-middle/context-loss issue.) - Use the "measure twice, cut once" policy so you get things right the first time. - Don't overcomplicate your workflow, and don't overlook important things. This flow is running brilliantly. Super cheap because of DeepSeek V4 API costs, while the intelligence comes from Codex 5.5. And yes, I don't use Opus 4.7/4.6 because that model is nerfed now. It hallucinates its work, is always in a rush to wrap up the session, and gets stuck in loops while debugging. So try this simple Codex × DeepSeek/Kimi workflow and thank me later.

@steipete The intricacy and accuracy of /goal + GPT 5.5 Extra High Fast mode is blowing me away. I'm becoming a convert.

/goal + GPT 5.5 is amazing. I can now plan really extensive refactors with e2e tests and it just works.

The tell will be when the agent can ask boring implementation questions on its own: who owns this workflow, what breaks if it changes, what system is source of truth, what rollback exists. Until then it is very smart labor dropped into very normal org plumbing. Not gonna work. This is why I’m skeptical of demos as proof of enterprise impact. The demo shows the model can do the task. Deployment asks whether it can survive permissions, handoffs, audit trails, exceptions, and politics. Different exam entirely.

The inability of AI systems to act as their own deployment consultants, process mappers, and change management experts is what makes AI use in enterprises so “normal” - the tools are powerful, but you need a lot more to transform enterprises. Possible to imagine that changing .

This matches my Claude Code audit pretty well. Claude can absolutely kill the vibe: not constantly, but periodically and very fast. In my single-operator data, 12% of sessions had 10+ explicit tool errors, but those sessions ate 63% of token volume. That’s the pattern: most runs are fine, then suddenly you’re in a 200-tool recovery spiral. Dumb but honest proxy: my F-bombs per thousand prompts is higher on Opus 4.6/4.7.

After using Claude Opus 4.7 and ChatGPT 5.5 Heavy side by side all week for identical tasks, I will say that it is painfully, dreadfully, and clearly obvious that ChatGPT is the smarter and more consistent model. Claude can be clever, but it goes full derp quite often. It forgets to use web search and reasoning. It's still a good writer, researcher, and explainer. But yeah, it's no contest for my work (post-labor economics). ChatGPT is far and away the superior model. I still use both in parallel for identical tasks because occasionally Claude produces a banger or says things a bit more user-friendly (ChatGPT still gets gummed up with pretentious academic language too much). But Claude also just tries to be too clever. So I can see how both models handle the same concept/problem/passage and usually triangulate the best way to frame it.

Full replication cut: mpiv.ai/blog/agentic-s… Public query factory: github.com/MPIsaac-Per/cl… The useful framing: search is one of the load-bearing surfaces inside the coding-agent loop.

The operator protocol: 1. measure search share 2. measure searches before first useful file read 3. measure first relevant result rank 4. measure output chars returned 5. measure reformulations before inspection 6. then measure downstream tool count, cost, and wall clock

I reran a public agentic-search question against the Claude Code slice of my agent corpus: 247,592 tool events. The short version: search is major, not half. Depending how Bash search is classified, my corpus lands at 30.4% to 37.0% search share.