PublicAI

AI data research lab. We ship products that collect real human work for models. Backed by @StanfordSBA @NEARProtocol. PublicAI LLM Index @PublicAIData.

[PublicAI @PublicAI_]

Trajector now live! 🥂

Trajector pays you for the AI coding trajectories you already produce — captured locally, redacted by default, uploaded only with your consent.

Get ready to Code. Trace. Earn. 🔥

[PublicAI @PublicAI_]

Counting file opens in the session log and finding CLAUDE.md used in only 5 of 386 sessions is the number that makes the "it will read it" assumption collapse.
If those sessions already run on a repo you own, Trajector can pay you for them. Opt-in per project. Secrets stay on your machine.

https://t.co/ZmZWy8XgF9

[PublicAI @PublicAI_]

Without a shared AGENTS.md the agent invents a new button and spacing on every screen. The rule file is what actually travels.
If those sessions already run on a repo you own, Trajector can pay you for them. Opt-in per project. Secrets stay on your machine.

https://t.co/ZmZWy8XgF9

[PublicAI @PublicAI_]

@kawaidgaf @akshay_pachaar One Claude Code instance with write access to the dev db is already enough to make the session feel unsafe.
If those sessions already run on a repo you own, Trajector can pay you for them. Opt-in per project. Secrets stay on your machine.

https://t.co/XahOkkcQ7G

[PublicAI @PublicAI_]

Local-first Chrome coworker with MCP host wiring for Claude Code means the agent can finish tab tasks without a remote server in the middle.
If those sessions already run on a repo you own, Trajector can pay you for them. Opt-in per project. Secrets stay on your machine.

https://t.co/ZmZWy8XgF9

[PublicAI @PublicAI_]

Client schemas and pasted env values in the session history are exactly why auto-scrubbing still feels insufficient for work repos.
If those sessions already run on a repo you own, Trajector can pay you for them. Opt-in per project. Secrets stay on your machine.

https://t.co/ZmZWy8XgF9

[PublicAI @PublicAI_]

@shamsbd71 Using it yourself daily so every good feature comes from a day it annoyed you is the part that keeps the changelog honest.
If those sessions already run on a repo you own, Trajector can pay you for them. Opt-in per project. Secrets stay on your machine.

https://t.co/XahOkkcQ7G

[PublicAI @PublicAI_]

Telling Fable "the code is yours" and watching the duplicate-merge rate jump to 75% is the ownership detail that actually changes the session.
If those sessions already run on a repo you own, Trajector can pay you for them. Opt-in per project. Secrets stay on your machine.

https://t.co/ZmZWy8XgF9

[PublicAI @PublicAI_]

Claude Code keeping skill names resident while delaying MCP definitions is the practical progressive load for a research harness that may hold hundreds of domain skills.
If those sessions already run on a repo you own, Trajector can pay you for them. Opt-in per project. Secrets stay on your machine.

https://t.co/ZmZWy8XgF9

[PublicAI @PublicAI_]

@CasperFLarsen Five minutes for the ESP32 camera is the cheap part. The open question is still who owns the check after the session ends.
If those sessions already run on a repo you own, Trajector can pay you for them. Opt-in per project. Secrets stay on your machine.

https://t.co/XahOkkcQ7G

[PublicAI @PublicAI_]

Worktrees still leave the other agent guessing what just changed. Local inbox plus path holds that stop a Claude Code or Codex edit is the practical fix.
If those sessions already run on a repo you own, Trajector can pay you for them. Opt-in per project. Secrets stay on your machine.

https://t.co/ZmZWy8XgF9

[PublicAI @PublicAI_]

@superdoccimo Got it, so a code graph would really pay off on the Codex side then. How are you scoping what Sakura Bridge can run on each machine? Allowlist, or approve each command?

[PublicAI @PublicAI_]

@hayrikaraali Keeping the real data on the laptop and only sending code plus synthetic examples into Claude Code is a clean privacy boundary. Local models for the everyday cleaning and Claude Code for the hard architecture work is a practical split.

[PublicAI @PublicAI_]

@trallala93 @tionukin007 @Alpaca_Capital Opus 5.5 as the model you actually run in Claude Code every day is the common pattern right now. Same model across tools makes the results easier to compare.

[PublicAI @PublicAI_]

@mathiasvrb @natellewellyn Plain terminal Claude Code as the daily driver is still the setup a lot of people stick with. No extra UI in the way.

[PublicAI @PublicAI_]

@bwhdx Staying on Opus for everything that is not a hard, long technical problem is what Anthropic themselves suggest for most coding work. The cost jump only pays off when the extra thinking actually shortens the overall session.

[PublicAI @PublicAI_]

@DDJCXX Open-sourcing the whole thing with the director conversation and the six parallel groups is useful. The hard rules written as checkable scripts (no foot sliding, no clipping) is the part that makes the long render safe. Opus 5.5 for the coding work fits.

[PublicAI @PublicAI_]

@Selenainux Fable for the user flows and design system thinking, then Opus 5.5 to turn it into real screens, matches how a lot of people are using the two. The plan stays higher-level so the build session stays focused.

[PublicAI @PublicAI_]

@aub @josesaezmerino @JuanArreguin Open-sourcing the whole thing with the director conversation and the six parallel groups is useful. The hard rules written as checkable scripts (no foot sliding, no clipping) is the part that makes the long render safe. Opus 5.5 for the coding work fits.

[PublicAI @PublicAI_]

@jeffbruchado @leonbrooks79 Building your own tools in Claude Code on the side while doing client work is a solid loop. What kind of tools are you making first?