Pradeep Singh

Tinkerer & Builder | AI Agent Tools

[Pradeep Singh @softcane]

In 40, spent most of my career building things for other people and too many nights building side projects that never found users.

So I’m giving myself 12 months to change that.

I’ll build small products, charge early, and give each one 14 days to show signs of life. If nobody cares, I move on.

The goal is €5–10k MRR. I’m doing it alone, without funding or an audience.

I’ll share what I build, what earns money, and what fails embarrassingly.

Target: €5–10k MRR
Start: €0 MRR, 0 followers
Setup: solo, no funding
Rule: launch in 14 days or kill it.

My experiments are here:
https://t.co/F0X8Rlb3uj

Let’s see where this goes.

[Pradeep Singh @softcane]

@willcb It's not consistent. For example, fable 5 through google-vertex/global works. Fable 5.1 blocks immediately.

[Pradeep Singh @softcane]

Life would be easier if we had

echo 8G > /sys/fs/cgroup/gpu/app/gpu.memory_limit

[Pradeep Singh @softcane]

@quxiaoyin https://t.co/07IoZA9QGK

[Stencil @stencil_labs]

what omp got wrong, what omp² changes. why "a while loop around a fetch" is systems software in a trenchcoat: https://t.co/BRwTnFh6k2

[Pradeep Singh @softcane]

Big labs don’t expose logits, obviously. So I was wondering, if you ever wanted to play with distillation at scale, what other signals could you get from these models?
Had some fun building this experiment: give Claude Code/Codex an enforced “workbook” and ask them to write down decisions, alternatives they considered, tradeoffs, etc., while working.
Normal model output, not hidden reasoning. Interesting to think about whether traces like these could be useful as a supervision signal.
https://t.co/GQDE8f4n2h

#AIAgents #ModelDistillation #ClaudeCode

[Pradeep Singh @softcane]

@_can1357 bitter lesson, bitter lesson
Oh please! Are we settled on what’s the minimum harness sets?

[Pradeep Singh @softcane]

@_can1357 Awesome

[Pradeep Singh @softcane]

/omfg remembers my preference even with opus 5 without spamming <system-reminders> like claude code @_can1357

[Pradeep Singh @softcane]

Why Claude code sub agents throwing raw jsnol to the main thread, I thought it supposed to return summary of the work done, no?

[Pradeep Singh @softcane]

Why Opus 5 feels bad.
System Prompt: Write code that reads like the surrounding code: match its comment density, naming, and idiom.

Vibe-coded projects using the previous model <- Garbage In
Opus 5 -> Garbage Out

[Pradeep Singh @softcane]

I guess they can’t if the user would change mid term but why a user would change mid term.
Many decisions in Claude code, I just don’t understand.
For example, /advisor instructions goes in system prompts but output preferences are injected every turn. So that tells it’s bad in specific instructions following for sure.

[Pradeep Singh @softcane]

Agents Workbook, watch Claude Code, and Codex write down their working notes
A proxy for Claude Code and Codex that gives the agent a “workbook.”

Before ans, it writes a lonnnnnng note about what it thinks the problem is, what options it considered, what it rejected, and why. You can watch those notes live while the agent works.

The interesting part for me: does what the agent says it plans to do actually match what it does next?

This is normal model output through a tool call, not hidden COT extraction.

Repo: https://t.co/GQDE8f4n2h

[Pradeep Singh @softcane]

@_can1357 I had some fun as well https://t.co/2onaW1PGRe

[Pradeep Singh @softcane]

I use Grill-Me daily; however, many times I feel overwhelmed. I try to write a modified version which works with my brain https://t.co/md1Rv9vSv7

@mattpocockuk, would you consider Grill-Me Lite for an ADHD mind like mine?

[Pradeep Singh @softcane]

@_lyraaaa_ Oh, nice, in my experiment, it was called a workbook, though the same idea

[Pradeep Singh @softcane]

@thsottiaux Because Codex tends to keep working until it gets stressed when I'm out of credit.
It's like, "Oh, I'm running out of credit; let me stress, perhaps cheat on the task, and say I'm done"

[Pradeep Singh @softcane]

@_can1357 Claude Code also uses 2 breakpoints, I believe. I started observing Claude Code cache usage on a grafana plot

[Pradeep Singh @softcane]

@SpotifyEng @ClaudeDevs @GeminiApp @OpenAI @Spotify Banking on the backstage success?