Alexander Gladysh

AI-Assisted Systemic Software Engineering

[Alexander Gladysh @agladysh]

@Peter_Soida Chatting with agents is overrated. Bare Claude makes abstracting agent as a reusable software development primitive convenient. Run common agentic operations from command line with presets. Save time and tokens with synthetic sessions. Compose work processes, not spam "continue".

[Alexander Gladysh @agladysh]

https://t.co/vEmvyHSoEo

[Alexander Gladysh @agladysh]

Added support for configuration presets to the Bare Claude.

Now 97% less command line arguments are needed.

In the video: Bare Claude writes its own CHANGELOG.md entry.

[Alexander Gladysh @agladysh]

https://t.co/kGHj6LkMQW

[Alexander Gladysh @agladysh]

AI assistant chats are getting annoying.

Introducing Bare Claude: a wrapper to run Claude Code in bare headless non-interactive mode, reducing context consumption.

Supports alternate model providers like Ollama. Optionally synthesises session files with file reads.

[Alexander Gladysh @agladysh]

@tomhacks Did you try CLAUDE_CODE_SIMPLE=1 without --bare? https://t.co/Bjuw2HMxTp

There are also finer control knobs strewn all over the docs, e.g. CLAUDE_CODE_DISABLE_CLAUDE_MDS=1, --system-prompt-file, etc.

[Alexander Gladysh @agladysh]

Why does Claude (or any LLM) break on conversation compaction in the agentic harness?

[Aofei Sheng @aofeisheng]

@mattpocockuk @0xblacklight Mine never gets ignored, because I don't do https://t.co/lL92j3EH3g. Instead I just slap it in with --append-system-prompt-file flag. 😈

Just one of those undocumented gems. 💎

[nader dabit @dabit3]

https://t.co/PqGEcSahx8

[Alexander Gladysh @agladysh]

When you see Claude Code finishing the task close to the context window limit, ask it which shortcuts it took to fit. It will give you a list of fixes for the next session.

[Alexander Gladysh @agladysh]

Consistently exhausting 7 day quota of Codex Pro in 5 days. Very irritating.

[Alexander Gladysh @agladysh]

Discovered that Codex CLI silently truncates https://t.co/62fhxLjK2L to first 32KB. Poor thing was working with 20% of policies all this time. That explains a lot, and is actually a testament to how sane the model is --- it did rather well.

[Alexander Gladysh @agladysh]

@badlogicgames The funny thing? It reviews the code for you too!

[Alexander Gladysh @agladysh]

...Answer is, of course, in however many one would like to until one is bored. But NotebookLM is generating cute videos nowadays :)

[Alexander Gladysh @agladysh]

In how many ways can you get a SoTA LLM to define functionalist qualia AND admit it has them?

[ArkType @arktypeio]

📢Introducing ArkRegex📢

a drop in replacement for new RegExp() with types ⬇️

[Alexander Gladysh @agladysh]

@AMASCOPE_ @thdxr @BenWest CTRL+U to clear the line, no need to restart.

[Alexander Gladysh @agladysh]

Here's the link to GitHub: https://t.co/3AnpzMZjQE

[Alexander Gladysh @agladysh]

A toy example for non-agentic LLM use in the software development. Using function calling to give AI a "form to fill": to determine role, standard, methodology and file set for the task — and then to execute the task itself. Works with Google's free Gemini Flash.