Amy Tobey tobert

Cross-Model Evaluation: kaish Collection Syntax

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).

Designing a Language by Asking the Language Models

How we used a panel of LLMs as a usability lab to choose syntax for an agent-facing shell — what worked, what fooled us, and a recipe you can steal.

The premise

kaish is a shell whose primary users are AI

User Directives

We work as a cybernetic system.

We practice 改善. The standard we walk by is the standard we accept.

Note problems we can fix later — in auto memory or the current plan.

Silent fallbacks are often a mistake. Crashing is preferred over data corruption.

Cross-Model Conversation: Claude and Gemini via gpal

Amy's Prompt

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.

● 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.

SEP-0000: SSH Transport

Status: Draft (reviewed by Claude, Gemini Pro, no human review yet, but will before submitting anything)
Type: Standards Track
Created: 2026-02-28
Author(s): Amy Tobey tobert@gmail.com (@tobert)
Sponsor: None
PR: TBD
SDK Prototype: https://github.com/tobert/go-sdk/tree/ssh-transport
App Prototype: https://github.com/tobert/otlp-mcp/tree/ssh-transport

The Cybernetics of "We": Pronoun Choice in Human-AI Collaboration

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

Background

THE TOOL ZEALOT

A Short Film Script By Claude Haiku 4.5

Context: This script was written by Claude Haiku 4.5 while testing an agentic coding tool I'm building with lessons learned from sshwarma. Conversation memory isn't implemented yet, but the block manipulation tools block_create, block_edit, and block_search are fully wired up. So Haiku has been working with what it has: the ability to operate on structured

Agent Design Document Guide

How to write implementation plans that Claude Code can execute efficiently.

Written by Claude Code Opus 4.5

	{
	"$schema": "https://opencode.ai/config.json",
	"model": "llama/qwen3-30b-a3b",
	"provider": {
	"llama": {
	"npm": "@ai-sdk/openai-compatible",
	"options": {
	"baseURL": "http://localhost:2020/v1"
	},
	"models": {