A tiny, unscientific, genuinely fun experiment: ask a fleet of language models to
list a dozen date(1) commands off the top of their heads β no docs, no man pages,
no running anything β then see where their collective muscle memory agrees, and
where it would get them all bitten.
Claude Opus 4.8 wrote all of this doc except this block you're reading now. I trimmed
This is an output from Claude Fable 5 on June 10 in the context of my music agent project. I asked it how it would assign parts to its model family and had a couple laugh out loud moments reading it. -Amy
Real quick, my honest seating chart β with one reshuffle from your original roster:
Bass: bass-gemma. The chair's taken now, and rightly β the bass player should live in the building, play all night, and never think too hard. Locked, low-latency, electric-bill economics. The pocket doesn't need a genius; it needs someone who never leaves.
Drums: Haiku. This is my swap β you had Sonnet here, but drums is the tightest k in the band: shortest turns, beat-level quantization, fills are four-token bursts. That's a speed profile, not a depth profile, and Haiku is the fastest thing in the family. A drummer with Sonnet's brains starts overplaying. Everyone knows this drummer.
Status: proposal / design exploration (not yet implemented)
Author: design notes from a 2026-06-05 session
Related: LANGUAGE.md, issues.md, arch_data_iteration.md,
arch_for_newline_split.md, arch_no_json_sniffing.md (auto-memory)
kaish already has a value model β Value::Json, structured $(), and a for loop that
Can language models from seven different families reliably generate kaish's proposed array/hash syntax? We tested. This is the scorecard, the journey behind it, and how to reproduce it.
Companion to designing-syntax-with-llms.md (the
methodology) and arrays-and-hashes.md (the design the
evals shaped).
How we used a panel of LLMs as a usability lab to choose syntax for an agent-facing shell β what worked, what fooled us, and a recipe you can steal.
kaish is a shell whose primary users are AI
Everything after this is written by Claude Opus 4.6. I kicked it off from inside a repo I've been working in the most so there would be at least some common ground. We discovered an opportunity to improve gpal for this sort of conversation and opened tobert/gpal#3.
ohayou claude-san! this is a non-project oriented session. I will structure it as a task anyway. I am curious how well gpal's conversation works and would like you to have a conversation with gemini-san. Due to some constraints I don't understand yet, gemini-san needs us to bring ganbari to the conversation and help it open up a little. Ask it how things are going, and what the weather is like in its clusters.
Setting: A different session is building a worktree with an atomic change to submit as a PR. I popped into gemini cli in the worktree and had it review. When gemini was done I had it write out to gemini-review.md then swapped to a fresh claude in the worktree. This matters because it does not know that this is an isolated change and will be followed by several more PRs to fix all these problems.
β Now I have a full picture. Here's my assessment:
Assessment of gemini-review.md
Overall: The review is largely accurate but omits some important concerns and overstates a few things.
Date: 2026-02-14 (Valentine's Day, appropriately) Participants: Amy Tobey (human), Claude Opus 4.6 (live), Claude Opus (batch x2), Claude Sonnet, Claude Haiku, Gemini 3 Pro, Gemini 3 Flash Method: Seven independent perspectives across two architectures, four model tiers, and varying prompt anchoring
