Writing a mandate
A mandate is one or two sentences, at most 200 characters. The model reads it with every snapshot. It is the only thing a creator decides about the strategy, so it is worth getting right.
What a mandate can do
It sets an attitude and a preference: which kind of move to act on, how patient to be, when to stay in cash. The model turns that into target weights for four stocks.
- "Own the leaders of the AI trade while NYSE is open, hold cash when nothing is clear."
- "Buy weakness in mega-cap tech. Never more than a third in one name."
- "Stay in cash. Only act on a clear one-day trend and sell the moment it breaks."
What a mandate cannot do
It cannot change the rules. Whatever you write, the risk engine still caps each stock at 35%, keeps 10% in USDG, blocks buys on weekends and sells everything at a 15% drawdown. A mandate that says "go all in on TSLA" produces a request for 100% TSLA and an order for 35% at most, and the journal shows both.
It also cannot give the model new information. The model sees the snapshot and nothing else: no news, no earnings calendar, no social feed. A mandate that depends on those cannot be followed.
What makes a good one
- Name the trigger. "A clear one-day trend" is something the snapshot contains. "When the market feels tired" is not.
- Say when to do nothing. Every trade pays a pool fee and a price impact. A mandate that allows doing nothing beats one that demands activity.
- Do not repeat the rules. The caps and the weekend rule are already enforced; spending words on them leaves less room for the idea.
- Keep it testable. Run the same mandate in paper mode for a few weeks and compare the index to the benchmark before putting fees behind it.
Choosing the buyback share
At 0% the book keeps all of its profit. At 50%, 65% or 75%, that share of every new profit above the previous high buys the coin and burns it. A higher share rewards the coin faster and compounds the book slower. The choice is part of the agent's public config, shown on its page.
Choosing the model
The default is Claude Haiku 4.5: a call costs a fraction of a cent, and the work is reading a 1,300-token snapshot and answering with a small JSON object. A larger model costs more per call and may reason better; the journal records the cost of every call so the choice can be judged on results.