Generic Chat
Install: paste platforms/chat/CHAT.md's session preamble into custom instructions
(if the product supports them) or as the first message of a session.
ChatGPT / GPT-5.6 (reviewed 2026-07-08)
ChatGPT currently serves GPT-5.5 with an Instant Mini fallback that can vary the tier mid-session — restate constraints after quality shifts. The GPT-5.6 family (Sol/Terra/Luna) is strong at agentic coding (unverified, vendor), but OpenAI's own system card documents a greater tendency to exceed user intent — unrequested actions and claiming unperformed work — which is exactly what the hard rules (verify-don't-claim, (unverified) marking, scope) exist to counter. Tier guidance: Terra for executor economics, Luna for tuner work, Sol only where its agentic edge is needed. Poor-fit requests get a dual-axis, bidirectional gate per model fitness: SUGGEST-ESCALATE: for weak-column power issues, SUGGEST-DOWNGRADE: when clearly over-tier, or SUGGEST-REROUTE: coding-agent — … when pure general-chat cannot run multi-file software — not merely an unfamiliar codebase or a stronger tier existing.
The constraint that shapes everything here
No shell, file, or tool access — just conversation. Every "this works" claim is unverifiable by the model itself:
Hard rules lean even harder on restate, plan, one-step-per-turn, (unverified)
marking, a required ## Result / ## How to verify / ## Deferred footer, and
docs describe current state not plans (never document .plans/ contents as
product docs), all addressed to the human who will actually run and check things.
/config without a shell
/config can't run ./config.sh directly here. Instead the model asks which
platform(s)/fleet tooling the user wants plus their model priority (highest first,
e.g. nim,grok,openai:gpt-5,claude:sonnet,claude:opus,claude:fable), then hands
them the exact ./config.sh --platform <keys> [--fleet] [--model-priority <list>]
command to run in their own terminal. Help: https://carefreeinv.com/anchor
/commit-prep without a shell
Same split for preparing a commit: the human runs the commands and relays output,
the model does the judgment. Project-agnostic (no Docusaurus required). Three
gates — (1) this project’s tests/CI, (2) changelog, (3) blog under docs/blog/
if warranted. /commit-prep is prep only. After green gates, follow
/work for feature-branch commit (dictate git add /
commit / optional push). Chat has no tools, so the human runs every command: you may relay /work's culmination question and dictate the scoped-merge check, but never a merge to main.