Michael Ramos

Cofounder, AI @EQTYLab / prev dc - complex systems / veteran / For fun: @plannotator

[Michael Ramos @backnotprop]

Anthropic code review this, clanker review that ... why don't you shut up and review+annotate your own code.... (yes im a loser who still manually reviews code)

Originally inspired by a bunch of feature requests and then seeing @dillon_mulroy tweet a similar cool ux.

@plannotator for reviewing plans (primary focus) and code, fully oss.

OpenCode, @badlogicgames 's https://t.co/tc56yk4Uom, and Claude Code and other clankers

[Michael Ramos @backnotprop]

@Cloudflare The link is broken, better link to use?

[Michael Ramos @backnotprop]

Have been working with Nvidia on hardware root of trust & verified agentic network traffic.

https://t.co/GZTIypoFpZ

[Michael Ramos @backnotprop]

@badlogicgames an industry standard/schema for what is persisted would be interesting

[Michael Ramos @backnotprop]

@AOJDevStudio @mattpocockuk @IndyDevDan Link/describe the evals.

[Michael Ramos @backnotprop]

@brandon_galang Facts

https://t.co/PEzygYWNIL

[Michael Ramos @backnotprop]

The initial prototype was fun. Took all boolean query results and reranked them - parallel fan out style up to 50,000 results, using flatbuffers (no json), near-perfect cpu utilization with golang (limited by resource pinning on k8s).

[Michael Ramos @backnotprop]

One early enterprise ai app I developed was for the USPTO. I partnered with google and we built AI search at the agency from 2018-2023 - augmenting their onprem lucene. ~250M longdoc embeddings (scann retrieval and transformer rerank), model architecture optimized for ~100 preemptible high-cpu k8s cluster - we beat GPU inference performance/cost, private data (all patent applications, 10,000 examiners)

A primary UX ended up being simple: upload the application, select/highlight the text you wanted ai to focus on, get the results. Integrated into their legacy app

[Michael Ramos @backnotprop]

@Steve_Yegge https://t.co/SayYp8uEQN

[Michael Ramos @backnotprop]

@thsottiaux @jxnlco no !bash in skills

[Michael Ramos @backnotprop]

@SIGKITTEN @OpenAI @ajambrosino congrats!

[Michael Ramos @backnotprop]

will probably refine a 1000 more things, but we're close to shipping a big @plannotator update.

a review surface that lives along side your agents. There when you need it, agents stay integrated & connected, organized by project workspaces.

(legacy/ephemeral mode can be switched back on).

[Michael Ramos @backnotprop]

@th4lweg No, I I've been trying it out in CC & so i can evaluate the workflow.js files it creates.

I was thinking it would be a fun weekend project to set it up with Pi.

[Michael Ramos @backnotprop]

this replaces /goal and I'm very bullish on this.

It automates much of my own manual orchestration.

I expect all harnesses will have a core UX that looks like this 6mo from now. Again, it relieves a lot of manual overhead in terms of how I think we are all already orchestrating agents. Frontier models are being trained to do more on their own, gather own context, verify own work - replacing the same prompts/skills i run over and over.

[ClaudeDevs @ClaudeDevs]

New in Claude Code (research preview): dynamic workflows.

Claude writes an orchestration script on the fly, then spins up a large fleet of coordinated subagents in parallel to take on your most complex tasks.

Use the word "workflow" in a prompt to get started.

[Michael Ramos @backnotprop]

@nbbaier @kieranklaassen /workflows

[Michael Ramos @backnotprop]

@zeeg try it with @plannotator

[Michael Ramos @backnotprop]

@muhashmii @RhysSullivan run claude code dynamic workflow and then save it. basically it

[Michael Ramos @backnotprop]

@0xSero @plannotator Look ma.

Preesh!

[Michael Ramos @backnotprop]

@ryannystrom https://t.co/Jo3RD9efi4

[Michael Ramos @backnotprop]

@poteto's "interrogate" review is the best code review skill i've come across - better than codex/claude defaults. It's how I've been manually orchestrating a bunch of concurrent reviews+combined output inside of Plannotator, but nicer to run as a single skill.

note: It creates an ensemble of agents with varying models you need access to (you can customize to make provider specific).

https://t.co/np6RERkJpO

[Michael Ramos @backnotprop]

@poteto's "interrogate" review is the best code review skill i've come across - better than codex/claude defaults. It's how I've been manually orchestrating a bunch of concurrent reviews+combined output inside of Plannotator, but nicer to run as a single skill.

note: It creates an ensemble of agents with varying models you need access to (you can customize to make provider specific).

https://t.co/np6RERkJpO