zettel-brainstormer（泽特尔集思广益）

Zettel Brainstormer

Run this workflow in order. Keep each stage separate so relevance decisions happen before drafting.

Configure Once

1. Run setup:

python zettel-brainstormer/scripts/setup.py

1. Confirm zettel-brainstormer/config/models.json contains:

• zettel_dir
• output_dir
• models and agent_models
• retrieval.link_depth and retrieval.max_links

Stage 1: Retrieval

Goal: retrieve candidate notes from the seed note.

Required order for this stage:

1. Read retrieval limits from config and target candidate count using retrieval.max_links.
2. Check if the external zettel-link skill is available. If it exists, run semantic retrieval via its scripts/search.py command using the seed note's topic or title. If it doesn't exist, warn the user and skip this step.
3. Run local retrieval with scripts/find_links.py to gather exact wikilinks and tag-overlap notes.
4. Merge and deduplicate candidates from both sources. Prioritize semantic candidates first, and trim to the configured count.
5. Exclude the seed note itself.

Local retrieval command:

python zettel-brainstormer/scripts/find_links.py \
  --input "/absolute/path/to/Seed Note.md" \
  --output /tmp/zettel_candidates.json

Treat /tmp/zettel_candidates.json as the candidate pool for preprocessing.

Stage 2: Preprocess (Subagent Per Note)

Goal: preprocess each candidate note and decide relevance to the seed note.

1. Read agents/preprocess.md as the per-note instruction.
2. Spawn one subagent per candidate note.
3. For each note, require:

• Relevance score against the seed note topic.
• Concise summary.
• Distinct key points.
• Short evidence quotes when useful.

1. Save each subagent output as markdown (one file per source note).

Quality rules:

• Reject notes with weak relevance.
• Prefer concrete claims and non-duplicated points.
• Keep outputs compact and structured for downstream merge.

Stage 3: Draft (Synthesis Subagent)

Goal: gather only relevant preprocess outputs and generate a referenced draft.

1. Run the aggregation helper:

python zettel-brainstormer/scripts/compile_preprocess.py \
  --seed "/absolute/path/to/Seed Note.md" \
  --preprocess-dir /tmp/zettel_preprocess \
  --output /tmp/zettel_draft_packet.json

1. Read agents/draft.md.
2. Use one drafting subagent with:

• Seed note content
• Filtered relevant notes from /tmp/zettel_draft_packet.json
• Required citation mapping from the packet

1. Produce a draft that cites source notes inline and preserves traceability.

Stage 4: Publish (Publisher Subagent)

Goal: rewrite the draft into natural long-form writing while preserving evidence quality.

1. Read agents/publisher.md.
2. Use one publisher subagent to rewrite the draft with these constraints:

• Remove generic AI phrasing.
• Use natural language and a coherent author voice.
• Organize points with clear tiered argument structure.
• Remove irrelevant points.
• Do not force weak connections between notes.
• Keep explicit citations for all retained claims.
• Do not publish the draft's internal "Argument Spine" section.
• Append valid frontmatter properties, including article-relevant tags.

1. Always end with a ## References section listing every cited note.

Stage 5: Delivery

Goal: Present the final output to the user.

1. Deliver or summarize the published draft for the user.
2. Crucial: When responding to the user, ALWAYS include the final list of references/notes that were actually used and cited in the brainstorm.

Bundled Resources

• agents/retriever.md: retrieval-stage instructions
• agents/preprocess.md: per-note preprocessing instruction
• agents/draft.md: synthesis drafting instruction
• agents/publisher.md: publication rewrite instruction
• scripts/find_links.py: retrieval script for wikilinks + tag overlap
• scripts/compile_preprocess.py: filter and merge preprocess outputs into a draft packet
• scripts/obsidian_utils.py: wikilink and tag helpers
• scripts/config_manager.py: shared config loader
• scripts/setup.py: interactive config setup

Maintenance Rules

• Keep stage boundaries strict: retrieval -> preprocess -> draft -> publish.
• Keep prompts in agents/ and scripts in scripts/.
• Remove deprecated scripts instead of keeping parallel legacy paths.