Tao

Trusted SOC 2 Compliance | Turn loom video to living Guide

[Tao @Tao_DeepGuide]

@conductor_build in the latest version on MacOS, the input was compressed to a single line with a big margin between the tools and bottom. This severely impacted entering prompts. Would you kindly fix it?

[Tao @Tao_DeepGuide]

open-attest v0.5.0 is out!

Agent now reports 20 signed device-posture checks across mac/win/linux, including:

- a daily installed-app inventory,
- SSH exposure (daemon + key count),
- full local user/admin roster,
- the MDM-managed lock policy signal.

Open-source endpoint attestation, no MDM required.

https://t.co/y2iZKleh8r

[Tao @Tao_DeepGuide]

A little fun with kanban board in Screenata

[Tao @Tao_DeepGuide]

@llama_index is your email sign up broken? Never received verification code after creating an account.

[Tao @Tao_DeepGuide]

@maybepratikk This is so close to what I'm doing with @screenatadotcom in compliance space.

Chat agents + proactive briefings is the way to go. Love to exchange notes on designing in this direction.

[Tao @Tao_DeepGuide]

Today we're shipping Kanban for SOC 2, powered by AI agents.

Compliance is no longer a dashboard with 200 empty checkboxes or a spreadsheet of “evidence.”

It's a task Kanban of real work. Easy to digest in a glimpse. Powerful to search with filters.

The best thing is half of the work is being done by AI agents, half landing in my queue for a human decision.

The traditional path is a vCISO + a GRC dashboard + a CTO doing screenshots at 11pm. The new path is an AI Compliance Officer that actually owns the work, plus a human for the decisions that legally must be human.

This is where Screenata is heading to.

[Tao @Tao_DeepGuide]

GPT 5.5 is more capable than Opus 4.7 even in frontend development now.

Just had a sophisticated double-rendering bug in react component. First tried Opus 4.7 High as usual, got in an endless loop of debugging, couldn't solve it.

Switched to GPT, fixed in one shot.

[Tao @Tao_DeepGuide]

GRC tools make you click through 14 dashboards to upload one PDF.

We made it a CLI.

SOC 2, HIPAA, ISO 27001 — from the terminal. Where engineers actually live.

No portal. No clicking. Audit-ready.

[Tao @Tao_DeepGuide]

@willkramer @parkerconrad Agreed. Curious to see how the claim of 80% converage is concluded.

[Tao @Tao_DeepGuide]

@bytaesu @better_auth fantastic. Thanks for the great work! It was indeed one area that people often overlooked and left curious why.

[Tao @Tao_DeepGuide]

We built our authentication layer with @better_auth - the best TS library for auth, but it nearly killed our first customer deal.

Invitation sent. User reports login failure. Checked logs. Backend returned "User not found" error. But "User" table contained the user's email.

It turns out the user's email address is mixed case. Our invitation table isn't normalizing the email address.

It's a subtle bug, but perhaps should have dug deeper (and tested more thoroughly) when using a library with AI agent.

[Tao @Tao_DeepGuide]

@bytaesu @better_auth "better-auth": "1.4.9", "better-call": "1.1.7". But the bug was due to we directly inserting the data into DB instead of using better-auth api. Apologies for the misunderstanding.

[Tao @Tao_DeepGuide]

I lowered my Vercel bill by 98%.

There have been overage bills on my $20 Vercel Pro plan in recent months for a landing website that is pretty much SSR. I was wondering why and took some time to investigate. The usage decreased by 98% after making a single adjustment.

It turns out most of the cost comes from "Image Optimization Transformations" and "Image Optimization Cache Writes". Vercel default value for image optimization cache is really small, so it keeps transforming and writing cache for <Image/> tags. For a landing site where image doesn't change often, I changed the settings at `next.config.mjs`:

```js
// next.config.mjs
const nextConfig = {
images: {
minimumCacheTTL: 31536000, // seconds — how long Vercel caches transformed images
formats: ['image/avif', 'image/webp'],
deviceSizes: [640, 750, 828, 1080, 1200, 1920, 2048, 3840],
imageSizes: [16, 32, 48, 64, 96, 128, 256, 384],
remotePatterns: [/* ... */],
},
};
```

I also reduced deviceSizes or ImageSizes because the bandwidth cost isn't much. Redeploy and I immediately see the difference.

This is definitely a required step to set up your Vercel project.

[Tao @Tao_DeepGuide]

Open Chronicle gives your AI coding agent photographic memory.

captures your screen locally,
OCRs it on-device,
summarizes into 1-minute memory windows,
serves those memories to Claude Code and Codex CLI

so when you ask "why is this failing?" or "continue where I left off,"
the agent actually knows.

👉 GitHub: https://t.co/ZBCdCs74qN
👉 MIT licensed, macOS-first, one-command install

[Tao @Tao_DeepGuide]

I was thinking of the same thing today! Interesting feature to add to open-chronicle. Working on it.

https://t.co/YMfdvHWow1

[Anthony Kroeger @kr0der]

i just found an amazing use case for the Chronicle feature in Codex.

it can help you optimize your day and become more productive because it has context of what's going on via your screen all day.

use this prompt - i share my own example at the bottom:

-----

Analyze all of my Chronicle memories. Please help me <CONTEXT ABOUT YOUR GOAL/ISSUE>

-----

my issue is that i stay up really late working and get <6 hrs sleep everyday. so i used that prompt and Codex found these issues through my Chronicle memories:

- validation/tooling errors are causing delays, e.g. failed linting/typechecks in worktrees
- checking Slack/Linear almost instantly after getting notifications, which steals attention from tasks and slows me down
- too many active WIP tasks (heavy context switching is bad)
- picking up tasks/debugging late at night which feels productive but sometimes these tasks blow up and delay sleep by a lot

[Tao @Tao_DeepGuide]

Zed releases Threads, orchestrating multiple worktrees and sessions in one place.

All coding agents and IDEs are converging to the same interface. The switching cost is close to zero. What are you using lately?

https://t.co/SDFI4OxIsn

[Tao @Tao_DeepGuide]

Foundation models are eating the upper levels. Agent harness layer is already dead, now general agent layer is dead.

Startups today need to have at least two of three below:

- Domain Data & Knowledge & Workflow
- Distribution in a vertical that requires more than wrapper
- Better and specialized UX

It's tough but exciting time.

[OpenAI @OpenAI]

Introducing workspace agents in ChatGPT—shared agents that can handle complex tasks and long-running workflows across tools and teams.

[Tao @Tao_DeepGuide]

@OpenAI https://t.co/YMfdvHWow1

[Tao @Tao_DeepGuide]

Open Chronicle now runs 100% offline.

Screen capture → Apple Vision OCR (on-device) → Ollama / LM Studio summarization (on-device) → SQLite (on-device) → MCP to Claude Code and Codex.

Zero cloud calls. Zero network. Works on a plane.