# Satisfaction Rate

> Check whether the user is less satisfied with their AI coding model now than over the previous two weeks, using their local Claude Code (and Codex or Kiro) transcripts. Use when the user asks whether a model got worse, dumber, lazier, nerfed or degraded, or asks for their satisfaction rate.

- Skill: `mihirokte/satisfaction-rate` (Agent Skill, multi-file: 11 files)
- Install (CLI): `npx get-skill add mihirokte/satisfaction-rate`
- Raw SKILL.md: https://api.skillmd.com/api/skills/mihirokte/satisfaction-rate/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: Mihirokte (https://skillmd.com/u/mihirokte)
- Updated: 2026-11-17
- Page: https://skillmd.com/skills/mihirokte/satisfaction-rate

---


# Satisfaction rate

The bundled script reads the transcripts already on this machine and rates every request
the user typed by their next move. A correction ("no, that's wrong"), a complaint, a
near-identical retry, or an interrupt counts as unsatisfied. Moving on or saying thanks
counts as satisfied. It compares a recent window with the window before it. Nothing is
sent anywhere, and it needs only Python 3.9+.

## 1. Run it

```bash
python3 ${CLAUDE_SKILL_DIR}/scripts/run.py --html /tmp/satisfaction-rate.html $ARGUMENTS
```

Defaults: the current window is the last 2 days and the baseline is the 14 days before
that, broken down by model. Useful flags:

- `--model opus`: only exchanges answered by a model whose id contains "opus"
- `--current-days 3 --baseline-days 21`: change the windows
- `--by source`: break down by tool instead of model
- `--min-rated 20`: lower the sample size needed for a verdict (default 30)
- `--json`: machine-readable output

If the user named a model or a time range, translate it into these flags.

## 2. Report it honestly

1. Lead with the verdict for the model the user asked about, or for `all`. Quote both
   rates, both `n` values, the change in points, and the 95% interval.
2. Say "decline" only when the tool's verdict says decline. "No detectable change" and
   "insufficient data" are answers too. Do not round them up into a decline.
3. Repeat every `caution:` line in plain words. When the tool mix or prompt size shifted,
   the work changed, and harder tasks can lower satisfaction without any model change.
4. If a window has too few rated exchanges, say how many it has and suggest a longer
   `--current-days`.
5. Mention which unsatisfied outcomes moved (corrections, frustration, retries,
   interrupts), from the "What drove unsatisfied exchanges" lines.
6. Close with one sentence of context: a lower satisfaction rate shows the user was less
   satisfied, not that the provider changed the model. Re-running a fixed set of prompts
   on a schedule is the way to test the model itself.
7. Offer to open the chart with `open /tmp/satisfaction-rate.html`.

Never paste transcript content into the reply beyond what the report prints.

## Notes

- Claude Code deletes transcripts older than `cleanupPeriodDays` (30 by default), so the
  baseline can only reach back that far.
- The phrase lists are English-only. See `scripts/satisfaction_rate/signals.py`.

