Warmcoffee Agent — Development skill for Claude Code

warmcoffee-agent

A coffee dialing assistant that learns your palate — not the SCA average.

Built for daily use with [Claude](https://claude.ai). Drop `CLAUDE.md` into a Claude project and it becomes your dialing partner.

---

What it does

Every morning it asks three things: temperature, humidity, how you're feeling. It tells you exactly how many steps to adjust your grinder, what flow time to expect on the first shot, and how confident it is — based on your data, your machine, your water.

After 50 shots, the model stops using industry defaults and switches to coefficients fitted from your actual sessions. Nobody else can use your model, because your model contains you.

---

The WARM system

The operator is the most expensive and least stable sensor in the dialing chain.

If you're sick, sleep-deprived, anxious, or have already had four coffees today — your palate is unreliable. WARM mode detects this with a single question at session start. Under WARM:

• Flow time data is logged at full weight
• Taste feedback is logged at 0.2 weight
• Grind adjustment is based on flow deviation only
• Output: "Trust the numbers today, not the taste."

Accumulated WARM records can reveal personal patterns — frequent triggers during allergy season, for example — and automatically reduce sensory weight during those periods.

---

Architecture

**LLM does not do arithmetic. LLM writes the code that does arithmetic.**

User input (natural language)
        ↓
   Claude (CLAUDE.md)
   · parse variables
   · detect missing data
   · map taste vocabulary
   · classify WARM state
   · route to correct script
        ↓
   Python scripts
   · freshness decay (two-slope model)
   · temp / humidity offset
   · grind offset calculation
   · confidence score
   · quarterly personal formula
        ↓
   JSON data layer
   · sessions / grinders / beans / anomalies / vocab
        ↓
   Formatted output
   · recommendation + confidence + driver explanation

Every number you