Experiment 2: thematic analysis with 1M context

Part of claude-code-whats-new-for-the-garden.

Feature under test: 1M context window (Opus 4.6, standard pricing) Question: Can Claude hold an entire garden in context and discover themes that the gardener hasn’t explicitly structured?

Design

Two passes over the same garden, compared:

Approach A: fresh thematic analysis

Load all ~200 content files as raw markdown. No prior structure, no embeddings. Code themes purely from what the text says. Tests whether 1M context can find patterns a human would miss.

Approach B: graph-augmented thematic analysis

Start from existing embeddings and key phrases, cluster by similarity, then read representative files per cluster to name themes. Tests whether existing infrastructure accelerates or constrains discovery.

Risk noted before starting: B might just confirm what the existing tag/relation structure already captures — locking us into the known rather than surfacing the unknown.

Findings

Approach A

See thematic-analysis-approach-a-fresh. Seven themes and one emerging pattern discovered from a fresh reading of all 244 files.

Approach B

See thematic-analysis-approach-b-graph-augmented. Ten embedding clusters compared against the fresh themes. The embeddings confirm topical structure but miss cross-cutting themes about truth, authenticity, non-linearity, and bilingual critique.

Observations

• The 1M context window held all 244 files comfortably. Loading was the bottleneck (tool output limits required chunked reads), not context capacity.
• Fresh reading (Approach A) found themes that embeddings (Approach B) structurally cannot: themes defined by stance, autobiography, and argument structure rather than topic.
• Embeddings confirmed the known format contamination problem: videos cluster together regardless of topic.
• The combination is stronger than either alone: embeddings map the landscape, full-context reading finds the threads that cross it.