Tim Gachot

Building AI systems
with natural language.
Construction + AI

[Tim Gachot @TimGachot]

Over 200 documents came out of my brother’s recent hospitalization following a motorcycling accident.

Progress notes, consults, discharge summaries, medication schedules, wound care — the list goes on

Across these documents, buried in medical jargon, exists a routine that will maximize his chances of walking again.

The care team?
My parents & myself
3 of us — rotating in/out

The game plan?
Simple.

Digitize everything.
Record everything.
Drop in one shared folder.
The rest takes care of itself.

@Adobe scan to PDF
@NotebookLM is our RAG
@claudeai chron keeps it up to date + transcribes audio.

Now anyone in the family can query this entire database @NotebookLM.

To create medications schedules,
Ask about symptoms,
provide details to a specialist,

And turn hundreds of documents into a recovery playbook for my immigrant parents.

Feeling optimistic about his recovery,
Now we can focus on being there for him

[Tim Gachot @TimGachot]

@nbaschez Because it’s not just about fixing the system now, it’s also about making sure you see what you need to see to know it’s a quality output.

Also — assumption debt / cascading errors. Ensuring quality in inputs, outputs, and process at every step of the process.

[Tim Gachot @TimGachot]

@nbaschez Yes! Massive erosion of trust

I’ve found that by using telegram + CLI — I’m able to preserve the thread of how I’m having iterate, verify and my frustration.

Then I’m able to feed that log back to an agent to understand how to optimize our interactions

[Tim Gachot @TimGachot]

@omarsar0 Also quality of process as well — verifying assumptions, inputs, and outputs at each step of journey to the final deliverable.

That assumption debt if left unchecked can cascade undetected into the final “verified” deliverable. Learned the hard way…

[Tim Gachot @TimGachot]

@svpino Where does one draw the line between needing to build a custom model vs. needing to build a better harness for the existing model?

[Tim Gachot @TimGachot]

@measure_plan Brilliant

What inspired the idea?

[Tim Gachot @TimGachot]

For non-developers, I’ve been seeing the telegram + Claude CLI connection be a great way to create that ongoing history log.

Or configuring your telegram-to-CLI use to always paste into an ongoing local interactions log as a first step.

It’s also been a life saver to automate context management and reset so I have a seamless experience via telegram.

[Tim Gachot @TimGachot]

Set up a systems level agent on a /chron to run a deep analysis of human-ai interactions to identify other opportunities to revise the system at its core and across specific agents

Also — have it recommend how you as the user must leverage the system moving forward given the system improvements applied.

Because there’s no point in fixing the verification process if you’re still going to specify the detailed verification process whenever you engage with the agent.

Behavior needs to change on the system AND user side

[Tim Gachot @TimGachot]

Terminal sessions disappear. Chat windows compress. Memory files summarize.

Leveraging CC CLI + Telegram groups preserve everything — every correction I made, every time the AI got it wrong, every "no, not that" that shaped the system.

I accidentally built a complete archive of how human-AI partnership actually evolves.

The use cases are pretty cool

[Tim Gachot @TimGachot]

And don’t just update the system.

Instruct your agent to always test and iterate on the implementation approach until it is confident that the implementation will result in the intended effect.

Configure your agent to leverage historical human-ai interactions to simulate my usage patterns in its sandbox testing.

Have it prove viability to itself before it even presents it to you.

[Ronak Malde @rronak_]

I have long felt that agent harnesses - even claude code - are too restrictive, because they are still designed by humans.

New paper for Tinsghua and Shenzhen says, what if AI itself runs the harness, rather than defining it in code? Given a natural language SOP of how an agent should orchestrate subagents, memory, compaction, etc., we can just have an LLM execute that logic! (And AI could design that SOP dynamically and depending on the task too)

It's a bit mind-warping to think about, but genius once it clicks.

Makes you wonder how else we should be designing AI systems as we can start consuming more and more tokens

[Yoonho Lee @yoonholeee]

How can we autonomously improve LLM harnesses on problems humans are actively working on?

Doing so requires solving a hard, long-horizon credit-assignment problem over all prior code, traces, and scores.

Announcing Meta-Harness: a method for optimizing harnesses end-to-end

[Tim Gachot @TimGachot]

This is how I experience Claude now.

If Claude forgets, I just tell it to review the recent telegram group history

An infinite log of the evolution of my human-ai interactions.

A data source for optimization

[Tim Gachot @TimGachot]

I’ve offloaded Claude context management using tmux + telegram

1 - Context hits activation threshold
2 - Agent(s) complete task at hand
3 - Current state documented locally
4 - System agent updates me & reset terminal
5 - Current State + Recent telegram thread reviewed
6 - Agent(s) continue where they left off

Approaching a near seamless experience with CC

[Tim Gachot @TimGachot]

Finally have freed myself from being at the computer to use Claude.

As a non-developer, I started in the browser experience leveraging Claude projects. 2024.

Reached my breaking point with manual context & memory management

2026. Claude code is allowing me to manage 9 different terminals, powering their respective agents, via tmux, all from my phone. With a system agent here to troubleshoot all terminals.

Managing context and memory from the systems level.

I cannot believe this

[Tim Gachot @TimGachot]

@rokhladnik Okay but shit happens — @klaviyo responded pretty quickly and didn’t sugarcoat anything. What are we going to use instead? Mailmonkey? Ooo oo oo

[Tim Gachot @TimGachot]

@ShipBob That would be awesome!

[Tim Gachot @TimGachot]

@NickaMalinowski @ShipBob Hey @NickaMalinowski thank you for reaching out! I’m hitting you up via dm’s would love to hear more 🔥🔥