tooling on S Anand

Gemini CLI harness is not good enough

Sun, 08 Mar 2026 15:09:25 +0800

I’ve long felt that while the Gemini 3 Pro model is fairly good, the Gemini CLI harness isn’t. I saw an example of this today.

Me: Tell me the GitHub IDs of all students in this directory.

Gemini CLI:

SearchText 'github' within ./
Found 100 matches (limited)
Sending this message (14606686 tokens) might exceed the remaining context window limit (1037604 tokens).

Me: Only send the (small) required snippets of data. Write code as required.

Gemini CLI:

SearchText 'github' within ./
Found 100 matches (limited)
Sending this message (14606686 tokens) might exceed the remaining context window limit (1037604 tokens).

Come ON! It’s March 2026. We can’t pretend it’s October 2025 any more.

PS: I partly take back what I said. Codex had trouble, too. This problem may be harder than I thought. Still, Gemini CLI should not have gotten stuck where it did.

Using Codex to improve Codex

Sun, 01 Mar 2026 18:26:13 +0800

Instead of learning and applying new Codex features, I asked it to analyze my sessions and tell me what I’m under-using.

I'd like you to analyze my Codex sessions and help me use Codex better.

sessions/ has all my past Codex sessions.

Search online for the OpenAI Codex release notes for the latest features Codex has introduced and read them - from whatever source you find them.

Then, create a comprehensive catalog of Codex features.

Then, analyze my sessions and see which feature I could have used but didn't and make a comprehensive list.

Then summarize which features I should be using more, how, what the benefits are, and with examples from my sessions.

Document these in one or more Markdown files in this directory. Write scripts as required. Commit as you go.

It did a thorough job of listing all the new features and analyzing my gaps.

Read the full story

Here’s the summary: I’m using new models immediately, but not the new features of Codex. For example:

Parallel execution. Yesterday, I ran ~103 tool calls without the new spawn_agents_on_csv feature from last week, which would have saved a lot of time running in parallel.
Permissions. Last week, I ran a script that asked me for permissions 7 times towards the end. Instead, I could have used /permissions to set early permissions.

The best part is that it could just add a few instructions to my AGENTS.md:

Run multiple independent reads in parallel.
For 20+ tool calls, maintain update_plan throughout.
For long-running commands/tests, delegate via sub-agents and report checkpoints.
If blocked by permissions, ask me concise choices.
If sandbox/config gets in the way, use /permissions and /debug-config early.

Now, the beauty is that the tool optimized itself. I don’t even need to learn how to optimize it!

Sun, 20 Oct 2024 07:32:39 +0000

Arun Tangirala and I webinared on “AI in Education” yesterday."

“(“Webinared” is not a word. But “verbing weirds language”.)”

Mid-way, Jose Swan from the audience asked, “Can you summarise this session using an AI?

There are SEVERAL tools you can use to summarize talks. Whisper for transcription, FFMpeg for keyframe extraction, #NotebookLM for podcast generation, text-embedding-3-small for topic modelling, and of course, any regular LLM include #ChatGPT for summarization or translation.

I spent this morning applying those to the webinar video, and writing it up the process.

Read on for the details…

https://www.s-anand.net/blog/tools-to-publish-annotated-talks-from-videos/

PS: I referenced Raj Vadigepalli midway and thought of Vivekananda Vedula, bringing back hostel memories from Indian Institute of Technology, Madras.