version: "1.0.1" name: collect description: "Use when collecting URLs, notes, files, or excerpts into structured PARA knowledge" effort: low

Iron Law

Collected knowledge must always be classified.

Red Flags

• "Just save it, I'll organize it later" → STOP, because every collected item must be PARA-classified at capture time — unclassified dumps rot in storage.
• "This is related to AI" → STOP, because connections must name a specific principle, pattern, or concept — topic-level connections ("related to AI") fail the quality gate.
• "I'll create a new entry for this" → STOP, because you must search existing knowledge first — duplicates fragment the knowledge base and waste future search time.
• "I'll just copy the whole article" → STOP, because the analyst must extract and reduce to exactly 3 key points — verbatim storage is not collection.
• "The connection is obvious, I don't need a separate agent" → STOP, because the connector must run as a separate subagent without seeing the analyst's framing — inline connection finding introduces bias.

Collect

Collect a source, reduce it, connect it to existing knowledge, and store it in a PARA-shaped knowledge base.

When to Use

• Save a URL, note, file, or excerpt for later use
• Preserve a useful insight from the current conversation
• Archive findings from another skill such as /scc:research

Workflow

1. Check existing knowledge: search ${CLAUDE_PLUGIN_DATA}/knowledge/ for items with overlapping tags or titles before creating a new entry. If a duplicate exists, update it instead of creating a new one.
2. Detect source type: URL, raw text, file path, or search request. Fetch here, not later — the analyst agent has no web tools, so a URL must be resolved to text before dispatch (use /scc:unblock if the page is gated).
3. Dispatch the `analyst` subagent (extract + reduce): Extract the useful content, strip boilerplate, produce exactly 3 key points and a short summary. This MUST run as a separate subagent, not inline.
4. Dispatch the `knowledge-connector` subagent (find shared concept): Using only the stored knowledge base (not the analyst's output framing), find a specific shared concept connecting the new item to existing knowledge. This MUST run as a separate subagent to prevent bias from the analyst's framing.
5. Merge analyst and connector outputs.
6. Classify it into PARA and save structured JSON + markdown.
7. Verify output: Confirm JSON has all required fields, key_points has exactly 3 items, connections pass the quality gate. If any check fails, fix before saving.

PARA Criteria

Options

Storage

Data directory: ${CLAUDE_PLUGIN_DATA} is set by the plugin system. If unavailable, fall back to .data/ relative to the plugin root. Before writing state files, verify the directory exists with mkdir -p.

• Path: ${CLAUDE_PLUGIN_DATA}/knowledge/{para-category}/{slug}.json
• Required fields: title, source, source_type, collected_at, category, tags, summary, key_points, connections
• key_points must contain exactly 3 items

Dual Output

Each item is saved as .json (machine-readable index) and .md (YAML frontmatter + readable prose) at the same slug path. See references/para-method.md for the markdown template and search ranking weights.

Connection Quality Gate

A connection must name a specific principle, pattern, or concept — not a topic or domain.

If no specific connection exists, set connections to an empty array. Never force a connection.

Search Mode

Trigger: pass --search "query" instead of a source to retrieve rather than save. Invocation: /scc:collect --search "query" What it does: scans all stored JSON under ${CLAUDE_PLUGIN_DATA}/knowledge/, ranks results by tag overlap and title similarity, and returns the top matches with their summary and PARA category. See references/para-method.md for full ranking weights.

Gotchas

• Do not store the source verbatim.
• Do not create vague connections like "related to AI."
• Do not create a new entry when an update to an existing entry would suffice.
• Do not skip the output verification step.
• When classification is ambiguous, default to resource.

Subagents

analyst: { model: sonnet, tools: [Read, Grep, Glob], constraint: "produce exactly 3 key points", dispatch: required }
knowledge-connector: { model: haiku, tools: [Glob, Read], constraint: "name a specific shared concept", dispatch: required }

Dispatch requirement: Both subagents MUST be dispatched as separate agents (e.g., via superpowers:dispatching-parallel-agents). Running either inline violates isolation: the connector must not see the analyst's framing, only the raw source and existing knowledge base. If subagent dispatch is unavailable, run them sequentially with explicit context barriers — pass only the original source URL/text to the connector, never the analyst's summary.