Claude Mental Models — Development skill for Claude Code

Claude Code Mental Models

Analysis of how Claude Code's changelog and Anthropic's blogs reveal the evolving mental models of Claude Code use — 4,919 changelog entries across 380 releases (v0.2.21 – v2.1.251, April 2025 – August 2026) and 45 blog posts. The 13 themes were derived on the corpus through 2026-05-23 and extended by an incremental refresh (`cmm.refresh`, see `docs/methodology.md` §5). See the design spec and plan in `docs/superpowers/`, and findings in `docs/findings.md`.

Run the pipeline

uv sync
# LLM calls use the local `claude` CLI (subscription auth) — no API key needed.
uv run python -m cmm.collect_changelog
uv run python -m cmm.collect_blogs
uv run python -m cmm.tag_and_join
uv run python -m cmm.embed_cluster
uv run python -m cmm.thematic_coding
uv run marimo run notebooks/analysis.py

To extend the corpus to the latest upstream release **without** re-deriving the themes (incremental refresh — themes and clusters stay fixed, only new entries are assigned):

uv run python -m cmm.collect_changelog   # checks out origin/main, re-parses
uv run python -m cmm.collect_blogs       # carries previous URLs forward
uv run python -m cmm.tag_and_join
uv run python -m cmm.refresh             # writes refresh_report.json; stops if a trigger fires

All network and LLM calls are cached in `data/cache/`.

Deliverable

`notebooks/analysis.py` — a marimo notebook: feature-volume / churn / blog-coverage / mental-model-emergence charts, an interactive cluster explorer, a themes reference table, and a RAG chat panel grounded in the corpus.

Licensing

• Code (src/, notebooks/, tests/): MIT, © 2026 Vishal Sachdev.
• Derived data (data/processed/ theme, assignment, join, persona and coherence tables, and the JSON summaries): CC BY 4.0 — see the data card for the per-file table, provenance and a suggested citation.
• Upstream text: the