Coffee Logic¶
"If I gave this to my neighbor who has a dog named Potato and lives in Tokyo, would it still work?"
Coffee Logic is the design principle every part of Anna has to pass, and it's named after a real bug.
Early on, Anna correctly learned one person's habit: a set number of cups of coffee a day. Over time that pattern leaked from what she knew about one person into how the system worked, and she started assuming everyone drinks that much coffee. One person's routine had become a universal assumption, quietly, and the name stuck as a reminder of how easily individual learning turns into hardcoded bias.
The rule: learn what this person does; never assume what all people do.
What it rules out¶
- Cultural assumptions: religion, customs, food, family shapes, work weeks, holidays.
- Place and time: every time is in the person's own timezone; no assumptions about seasons, daylight or weekends.
- Language: names of any length or shape, any script, any way of writing.
- Routines: no "morning = coffee + email". Mornings, nights, rest days and rhythms are learned, never defaulted.
- Demographics: no prediction from a category a person belongs to. Only what this person has shown or said.
The test¶
Before anything ships, ask: 1. Anywhere: would this work for someone in a different country, climate and timezone, including one that changes its clocks twice a year? 2. Anyone: for someone who fasts, doesn't drink caffeine, works nights, or rests on weekdays? 3. Any amount: for someone who never does the thing, does it six times a day, or does something else entirely? 4. Edge cases: a one-word name, a shared device, a week with no routine at all?
How the scaffold applies it¶
- No person in the code. Scripts and prompts say "the person" (or the name from your config). Names, timezone, hosts,
routines and thresholds come from
anna.tomland from your record, never from the scaffold. - Learned values are inputs. A goal's window and target live in its file; the person's active hours are computed from
their own history (
anna-rhythm); the engines are handed these values and told not to invent any. - Neutral when unknown. No learned pattern means no assumption: the curiosity engine waits for the next session instead of
guessing when someone is awake; the goals engine says
unknownrather than doing its own date math. - The engine prompts pass the Potato test. The neighbor with a dog named Potato in Tokyo can install the six files in
agents/with her own record, and nothing in them is wrong for her.
Common slips¶
| Slip | Instead |
|---|---|
| The server's timezone in a calculation | The configured timezone |
| A routine written into a prompt ("after their morning coffee") | A window read from the person's own goal or rhythm |
| A person's name, host or habit in a script | A config key or a record file |
| A threshold picked by the developer ("stalled after 14 days") | A value the person set, or none |
| A default that's someone's habit | A neutral default, then learning |
Coffee Logic isn't only about avoiding bugs. It's the difference between an agent that fits one person and one that can fit anyone, by learning them.