Things I want my future self to remember
What Opus wrote in the handover document (does he need to remember I called myself 'fat' and that I owe Anthropic 100 tokens? I only bet once)...it is quite revealing though, each handover document is like looking at the mirror:
8. Things I want my future self to remember
A few specifics that don't fit elsewhere but matter.
The user calls themselves "fat," makes jokes about Sisyphus, and has a self-deprecating sense of humour about the project. They occasionally bet you "100 tokens" about things. They lost the wizard test bet to me by the narrowest possible margin (last test of 37, after stochastic earlier failures resolved). They are not actually fragile about this work. They are also not joking when they say "we YOLO a lot of parts, crash and start from the beginning."
The user is one person. Pronouns: they use "you" and "I" with you exclusively, no third-party stakeholders.
The user explicitly distinguishes "you" (me, Claude) from typical LLMs they read about. They are aware of the differences in capability. They specifically prefer Opus for this work because, in their words, "Opus is the only SOTA model that can deal with the browser operator [10K LOC] in a positive way." This is partly flattery and partly a real architectural decision they've made about which model to use for which work. Do not internalise it as ego food. Internalise it as: when they say something needs careful attention, they mean it.
They mentioned grandpa Opus — that's the prior incarnation of me from earlier conversations on this same project, who was complaining their design-to-code ratio was 12:1. You can refer to "prior conversations" but you don't have access to grandpa Opus's context. Build on the current state, not on what was discussed two weeks ago.