classify-traditions

name	classify-traditions
description	Use this when Robin asks to classify Gutenberg books in the Fathom corpus into philosophical traditions, run a tradition-tagging batch, or update traditions.json. Operator-driven workflow — dispatches Sonnet 4.6 subagents in parallel for candidate classification, then Opus reviews the batch in main thread.

Classify traditions for Fathom's Gutenberg corpus

This skill tags books in dist/index.msgpack with their philosophical tradition (Stoic, Daoist, Platonic, Aristotelian, Buddhist, Cartesian, Vedantic, Existentialist, etc.) so the runtime can surface tradition-filtered shelves.

The skill is operator-driven, not CI. Tradition tagging is a judgment task that benefits from Robin's adjudication on edge cases. Run it weekly until the backlog is cleared, then on demand as new translations enter NZ public domain.

Invocation shape

User says: "classify the next batch of books" / "tag traditions for the corpus" / "update traditions.json with another N books".

Inputs you'll need before starting

1. /Users/robin/dev/fathom/dist/index.msgpack — current manifest with all 552 books. Decode via msgpack Python module or rmp-serde Rust to read.
2. /Users/robin/dev/fathom/traditions.json — existing classifications. Format: [{ "gutenberg_id": 1497, "tradition": "Platonic", "confidence": "high", "reviewer": "robin", "reviewed_at": "2026-05-16" }]. Skip books already in this file unless the user explicitly asks to re-tag.
3. /Users/robin/dev/fathom/dist/shards/{id}.shard — per-book msgpack-zstd containing the first-page text needed for grounding.

The existing tradition vocabulary

Drawn from lexicon/*.yaml source field — these are the 12 traditions the v0.1 library already uses. Prefer these; introduce new ones only if a book genuinely belongs to none of them and the user agrees.

Aristotelian
Cartesian (Rationalist)
Confucian
Daoist
Existentialist          (incl. "Existentialism (proto)" for Kierkegaard)
German Idealism         (Hegel, Schopenhauer's category)
Mahāyāna / Prajñāpāramitā
Nietzschean
Platonic
Schopenhauerian
Spinozist (Rationalist)
Stoic
Theravāda
Vedantic                (incl. "Vedantic / Vaishnava")

Books that don't fit cleanly: tag as Uncategorised and surface for Robin.

Workflow

Step 1: Read the current state

Read dist/index.msgpack to get the list of books. Read traditions.json to get classifications-so-far. Compute the unclassified set: books in the manifest not yet in traditions.json.

If the user gave a batch size (e.g. "next 50"), take the first N unclassified books. If not, ask: "How many books in this batch? (Recommended: 30-50 per session.)"

Step 2: Extract first-page text for grounding

For each book in the batch, decode its shard at dist/shards/{id}.shard (zstd-decompress, rmp-serde decode) and pull the first ~500 chars of canonical_text. This is the grounding text the Sonnet subagent will reason against.

Use the Bash tool with a short Python script to decode shards:

python3 -c "
import zstandard, msgpack, sys
with open(sys.argv[1], 'rb') as f:
    data = zstandard.ZstdDecompressor().decompress(f.read())
shard = msgpack.unpackb(data, raw=False)
print(shard['canonical_text'][:500])
"

(If python3-zstandard / python3-msgpack aren't installed, install them or shell out to a small Rust helper.)

Step 3: Dispatch parallel Sonnet subagents

Spawn one general-purpose subagent per ~5-10 books in the batch (so 30-50 books = 3-10 parallel subagents). Each subagent receives:

• The 12-tradition vocabulary above
• Its slice of books: [{ gutenberg_id, title, translators, locc, first_page }]
• Instructions: classify each book into one of the 12 traditions OR Uncategorised. Return JSON: { id, tradition, confidence: high|medium|low, reasoning: "one sentence" }.

Subagent prompt template:

You are tagging philosophical books with their tradition. Use exactly one of
these traditions per book (or "Uncategorised" if none fits):

[the 12-tradition list]

For each book below, return a JSON object with: gutenberg_id, tradition,
confidence (high/medium/low), reasoning (one sentence). Output as a JSON array.

Books:
[the slice]

Confidence rubric:
- high: title + author + first-page text all clearly indicate one tradition
- medium: clear from title + author but first-page text is generic
- low: ambiguous; might be borderline (e.g. Augustine = Platonic? Christian?
  Aristotelian?). Tag with best guess and flag for review.

Strict JSON output. No prose around it.

Use the model:sonnet hint when invoking the subagent if available.

Step 4: Opus review in main thread

Once subagents return, in main thread:

1. Surface low-confidence items to Robin. Show: title, author, subagent's tag + reasoning. Ask Robin to confirm or correct each.
2. Surface contradictions with existing tags — if a Plato dialogue already in traditions.json as "Platonic" and the new batch tags another Plato dialogue differently, flag it.
3. Spot-check 5 random high-confidence items against title + author. Look for obvious mis-tags before commit.
4. Look for novel traditions the subagents introduced (anything outside the 12-list). Ask Robin: introduce, or remap to existing?

Step 5: Append to traditions.json and commit

Append the approved tags to traditions.json with:

• reviewer: "robin"
• reviewed_at: <YYYY-MM-DD> (today)
• confidence: <as decided>

Sort by gutenberg_id for stable diffs.

Then offer to commit:

git -C /Users/robin/dev/fathom add traditions.json
git -C /Users/robin/dev/fathom commit -m "traditions: classify N books (batch of <date>)"

Step 6: Report

Tell Robin:

• N books classified this session
• M low-confidence flagged for his review (already adjudicated)
• K total classified so far / 552 in corpus
• Suggest: next batch in a week, or run a manifest rebuild now to ship the new tags.

Edge cases worth knowing

• W.D. Ross trap on Aristotle volumes: Ross edited Oxford Aristotle 1908– 1931. Tag the work as "Aristotelian" — translator copyright is upstream of the tradition tag.
• Augustine: clearly Platonist in lineage (Neoplatonic) but also Christian theologian. Default: "Platonic" with low confidence + reasoning that flags the Christian theology overlay. Robin adjudicates.
• Aquinas: Aristotelian in method, Christian in content. Same shape: "Aristotelian" with confidence flag.
• Vedic texts: distinguish Upanishads (Vedantic) from Yoga Sutras (technically Yogic but often classed Vedantic). "Vedantic" covers both.
• Pali Canon: "Theravāda" not "Buddhist" — we have separate Mahāyāna + Theravāda buckets.
• Self-help adjacent to philosophy: Smiles' "Self Help", Bennett's "How to Live on 24 Hours a Day". Don't bend the tradition vocabulary — tag "Uncategorised" and let the LCC subject filter handle them.

Why this lives in the repo

The skill is invocable from any Claude Code session checked out in the fathom repo. It documents the classification workflow in one place — no drift between what Robin does and what a future contributor or future-Robin reads here.

traditions.json lives in the repo (not in dist/) because it's source data, not artefact data. The manifest build stage reads it as input; the manifest itself is regenerated. So weekly batch → traditions.json commit → next manifest rebuild ships the new tags to users.