Stella CLI Size Fit Advisor — Development skill for Claude Code

Stella — CLI Size & Fit Advisor

A four-question conversational agent that builds a sizing profile, maintains a live confidence score, and produces a fit recommendation. Every sentence the user reads is generated by Claude; every number is computed by Python.


What works

• Full four-question interview, model-driven end to end — opening turn included.
• Live confidence score recomputed in Python after every answer, rendered as a 30-block progress bar with a numeric percent and a label.
• Confidence drops are narrated. A contradiction or a skipped question lowers the score, and the model is required to say why.
• One-follow-up rule. A vague answer earns exactly one clarifying question, then the interview advances regardless. Enforced in Python, not left to the model's discretion.
• Contradictions are recorded, not silently overwritten. Both values survive in state, the conflict is surfaced to the model, and the score takes a penalty.
• Declines are distinct from "I don't know". Declining caps the score at 80 permanently; not knowing is recoverable on the follow-up.
• Inspectable state via /state, which dumps the whole session object plus the active prompt versions and model ID as JSON.
• Graceful degradation. Empty input, gibberish, off-topic replies, hostile input, API timeouts, rate limits and schema-validation failures all produce a sensible turn and no traceback. A failed extraction rolls the user's turn back so a retry doesn't duplicate it.
• 52 unit tests covering the scoring arithmetic and state machine, including an exhaustive monotonicity check across all 256 specificity combinations.
• Session persistence — /save FILE and python run.py --resume FILE.
• Per-call tracing to JSONL (STELLA_TRACE=1): latency, tokens, prompt size.

---

Setup

python3 -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\a