LLM Wiki — Research Skill (wiki_research)

Resolving script paths (read first): Commands below invoke scripts as <BIN>/X.py (and a few as <SKILLS>/...). Resolve these to absolute paths once before running anything:

>

- <SKILL_DIR> = the directory this SKILL.md lives in. - <SKILLS> = the skills/ folder containing this skill = <SKILL_DIR>/.. - <BIN> = the bin/ folder beside it = <SKILL_DIR>/../../bin

>

Do not hardcode a fixed prefix like .agents/bin or ../bin: shell relative paths resolve against the current working directory (usually the topic root), not this skill's location. Once resolved, <BIN> is typically .agents/bin when invoked from the hub root, or .claude/bin from inside a topic directory.

This skill handles deep, parallel academic research, spinning up multi-perspective subagents to drill into complex topics and compile unified verdicts.

Tooling (framework-agnostic): This skill is written tool-agnostic. Map each capability to your own agent's tool — read-file (Read in Claude Code, view_file in Antigravity), sub-agent / parallel task (Task/Agent in Claude Code, invoke_subagent in Antigravity), web-search (WebSearch in Claude Code, search_web), shell (Bash/PowerShell). Use the closest equivalent your framework provides; if a parallel sub-agent tool is unavailable, investigate each research dimension sequentially yourself.

When the user asks to research a topic:

1. Draft a Dynamic Research Plan: Analyze the user's research query and determine the domain (e.g., Mathematics, Theoretical Physics, Computer Science). Subdivide the query into 3 or more distinct, domain-specific investigative dimensions.

• Example (Physics/Math): "Axiomatic Consistency Auditor", "Phenomenology & Experimental Reviewer", "Theoretical Extrapolator".
• Example (CS): "Technical Deep Dive", "Critical Reviewer", "Empirical Auditor".
• Graph Context: Before finalizing the plan, you are encouraged to query the local SQLite graph database (output/graph.db) to identify existing nodes related to the query. Ensure the index is up to date by running python <BIN>/llm-wiki.py graph first. Use python <BIN>/query-graph.py "<SQL>" to query the knowledge graph. Do not use direct sqlite3 command line execution.

Graph DB Schema:

• nodes(id TEXT PRIMARY KEY, path TEXT, title TEXT, type TEXT, category TEXT, summary TEXT, created TEXT, updated TEXT)
• edges(source_id TEXT, target_id TEXT, type TEXT)
• tags(node_id TEXT, tag TEXT)
• aliases(node_id TEXT, alias TEXT)

Example Queries:

• SELECT path FROM nodes WHERE category='reference' AND id IN (SELECT node_id FROM tags WHERE tag='quantum-mechanics')
• SELECT n.path, e.type FROM nodes n JOIN edges e ON n.id = e.target_id WHERE e.source_id = 'some-concept-id'

This helps contextualize your research plan within the existing knowledge graph.

1. Orchestrate Background Subagents: Spawn the parallel sub-agents using your agent's sub-agent / parallel-task tool according to your dynamic research plan. (If no sub-agent tool exists, investigate each dimension sequentially yourself.)

• Assign each subagent a clear, focused Role and Prompt tailored to their specific investigative dimension.
• Source Constraint (CRITICAL): Every subagent prompt MUST include this instruction:

"You MUST use your web-search tool or file-read tool to gather evidence. Do NOT make factual claims from parametric memory alone. Every claim must cite either a specific URL from web search or a specific local file path that you read. If you cannot find a source for a claim, mark it explicitly as [UNVERIFIED]."

• Subagent Output Contract (MANDATORY): Each subagent MUST return findings in this structure:

        FINDING: <summary of finding>
        EVIDENCE: "<quote or data point>"
        SOURCE_TYPE: web|local_wiki
        SOURCE: <URL or file path>

Findings without a valid SOURCE must be marked [UNVERIFIED].

1. Verify and Filter Subagent Results:

• Wait for all subagents to report back. If any subagent fails, log the failure and proceed with available results.
• Save all reported findings exactly as returned into a temporary file: scratch/temp_claims.txt.
• Run the verification script to automatically check the citations:

python <BIN>/verify_claims.py scratch/temp_claims.txt --topic-dir "<TOPIC_DIR>"

• Only use [VERIFIED] claims in your final synthesis. Collect [UNVERIFIED] findings separately.

1. Synthesize Findings:

• Merge verified findings into a detailed, authoritative synthesis document.
• Output Destination:
• If the research covers a broad topic (survey, tutorial-style) → save as wiki/topics/YYYY-MM-DD-<slug>.md
• If the research is a literature review of specific papers → save as wiki/references/YYYY-MM-DD-<slug>.md
• YAML Frontmatter (MANDATORY):

        ---
        title: "<descriptive title>"
        type: topic|reference
        category: topic|reference
        created: YYYY-MM-DD
        compiled-from: mixed
        sources:
          - <list of all cited URLs and file paths>
        tags: [research, <domain-specific tags>]
        confidence: <high|medium|low>
        summary: "<1-2 sentence summary>"
        ---

• If [UNVERIFIED] findings exist, include them under a clearly marked ## Unverified Claims section at the end. Do NOT mix unverified claims into the main body.

1. Post-Write Validation (MANDATORY):

• Run: python <BIN>/validate-output.py "<output_file>" --schema research --wiki-root "<TOPIC_DIR>"

If validation reports issues, fix them before proceeding.

• Run: python <BIN>/llm-wiki.py lint --fix <TOPIC_DIR>

1. Log: Append a log entry in log.md with: research query, subagent count, verified findings count, unverified findings count, output file path.