research

name	research
description	Run structured research — claim verification, query answering, and document fact-checking — with full evidence archive output.
user_invocable	true

Research

Overview

Run a structured research investigation. Produces a complete evidence archive with searches, sources, scorecards, evidence extracts, hypotheses, ACH matrices, and self-audits.

The input can contain claims (assertions to verify), queries (questions to answer), axioms (facts to assume true), or any combination. The research methodology determines how to handle each type automatically.

Commands

run — Execute new research

/research run [file=<path>] [output=<dir>] [id=<id>] [runs=<count>] [confirm=yes|no]

Parameter	Required	Description	Default	Example
file	No	Path to a markdown file containing claims, queries, and/or axioms.	Ask interactively	file=claims.md
output	No	Output directory.	Ask interactively	output=research/ai-trust
id	No	Research instance ID. Only needed if your output format uses it.	Auto-generated	id=R0005
runs	No	Number of independent runs. Each run is blind to the others. After all complete, a synthesis step produces aggregate results with consistency metrics.	3	runs=1
confirm	No	yes: confirm before running. no: batch mode, just run.	yes	confirm=no

Any key=value parameter not listed above is ignored. If an unrecognized parameter is provided, do not ask about it — silently ignore it.

Examples:

/research run file=claims.md output=research/ai-trust
/research run file=claims.md output=research/ai-trust runs=1
/research run

rerun — Re-execute previous research

/research rerun <path-to-research-directory> [runs=<count>] [confirm=yes|no]

Re-executes a previous research run using the saved input spec. The input spec file (research-input.json) is saved in the research instance directory during the original run. The rerun creates a new timestamped directory alongside the existing one(s). The runs parameter works the same as for run (default: 3).

CRITICAL: Isolation rule. The rerun MUST be executed with NO knowledge of prior run results. The skill reads ONLY research-input.json from the research directory. It MUST NOT read, reference, or pass to the subagent ANY files from existing date directories (prior run results, assessments, evidence, or any other output). The subagent MUST NOT be given the path to prior runs. This is essential for reproducibility — reading prior results would bias the new research through anchoring and confirmation effects.

Difference report. After the rerun completes (and ONLY after — the isolation rule must not be violated), the skill automatically generates a difference report comparing the new run against the most recent prior run.

The diff is produced by a SEPARATE step that reads both runs AFTER the rerun subagent has finished. This preserves isolation — the rerun agent never sees prior results. Only the post-run comparison step reads both.

The diff report is saved in {research-directory}/diffs/ as {prior-date}_vs_{new-date}.md. If no prior run exists (first run), no diff is generated.

The diff report contains:

1. Per-claim/query comparison: for each entity that exists in both runs, compare:

   • Probability/confidence shift (old → new rating)
   • Verdict change (which hypothesis prevailed)
   • New sources found (in new run but not prior)
   • Sources dropped (in prior but not new run)
   • Evidence direction shift

2. Collection-level comparison:

   • Overall distribution shift (how many claims moved between ratings)
   • New cross-cutting patterns
   • Gap changes (gaps closed, new gaps identified)
   • Source overlap (shared vs unique sources betweeruns)

3. Article impact summary:

   • Claims that strengthened (no action needed)
   • Claims that weakened (may need softening)
   • Claims that flipped (correction needed)
   • New information that should be incorporated

extract — Extract verifiable claims from a document

/research extract <url-or-path> [output=<dir>]

Reads a document and produces a JSON file of verifiable factual claims suitable for input to /research run.

Parameter	Required	Description	Default	Example
<url-or-path>	Yes	URL or local file path	—	articles/A0005/drafts/draft-v1.md
output	No	Directory to write the claims file	Local path: dirname of input. URL: must specify.	output=research/

Standard output file: extracted-claims.json (placed in the output directory). This name is fixed — do not ask the user for a filename. The output MUST conform to the extracted-claims.schema.json schema.

Extraction rules:

1. Read the document body only. Do NOT read the References, Bibliography, Sources, or Works Cited section. The references are a cheat sheet — claims must be extracted as they appear in the prose. If the prose names a study, institution, or statistic, that's fair game. If the only way to identify the source is to read the reference list, skip it.

2. Extract verifiable factual claims only. A claim is a statement that can be checked against evidence. Skip:

   • Opinions and editorial framing ("I believe...", "This is concerning...")
   • Rhetorical questions
   • Definitions that are being stipulated, not asserted as fact
   • Structural statements about the article itself

3. Include named entities and categorizations. Product names, tool names, framework names, and categorization claims ("X is a type of Y") are verifiable and MUST be extracted. Do not skip "simple" assertions.

4. If the input is a URL, fetch the page content and extract the article body. Strip navigation, headers, footers, ads, sidebars, and other HTML chrome. If the page structure makes it unclear where the article body is, err on the side of including too much rather than too little.

5. Present the claims to the user for review before saving (unless confirm=no — see below). The user may add, remove, or modify claims. The extraction is a starting point, not a final product.

Output format (extracted-claims.json):

{
  "source": "url or path to the source document",
  "extracted_at": "2026-04-13T12:00:00Z",
  "claims": [
    {"text": "first verifiable claim as it appears in the prose"},
    {"text": "second claim"}
  ],
  "queries": [],
  "axioms": []
}

The queries and axioms arrays are empty by default — the user may add items before running verification. This JSON is directly usable as input to /research run file=extracted-claims.json.

fact-check — Extract and verify in one step

/research fact-check <url-or-path> [output=<dir>] [confirm=yes|no]

Extracts claims from a document and immediately runs verification on them.

Parameter	Required	Description	Default	Example
<url-or-path>	Yes	URL or local file path	—	articles/A0005/drafts/draft-v1.md
output	No	Directory for all output (claims + research)	Local path: dirname of input. URL: must specify.	output=research/A0005-claims/
confirm	No	yes: present claims for review before verifying. no: skip confirmation, just run.	yes	confirm=no

Workflow:

1. Check if {output}/extracted-claims.json exists AND is newer than the source document. If yes, skip extraction and use the existing claims file. If no, run extraction.
2. Pre-fact-check gate (reference audit): Read the document's References section. For each reference:
   • If it has research evidence links (R/Q, SRC) → pass (already scored)
   • If it is clearly informational (link to an organization homepage, Wikipedia, regulatory text with no associated claim) → pass
   • If it supports a factual claim but has no scorecard → flag it Report any flagged references to the user. If flags exist and confirm=yes: ask the user to resolve them before proceeding. If confirm=no: report the flags as warnings but continue.
3. If confirm=yes (default): present the extracted claims to the user for review. The user may add, remove, or modify claims before proceeding.
4. If confirm=no: skip review, treat extracted claims as final.
5. Pass the claims to run for verification. The full research output is written to the output directory alongside the claims file.

confirm=no (batch mode):

• No confirmation prompt before extraction or verification
• No clarification of input
• Treat all input as final — the user specified everything on the command line
• Just run
• Still save research-input.json, snapshots, and all output artifacts

This is a convenience command. It does nothing that extract followed by run doesn't do — it chains them and adds smart caching of the extraction.

Future: Support for directory input (collection of files) to fact-check documentation trees is planned but not yet implemented.

Workflow (for run and rerun)

Step 1: Gather input

Parse any key=value parameters from the invocation. For any parameter NOT provided on the command line, ask the user interactively. For parameters that WERE provided, use them directly — do not re-ask.

1. Input (file= parameter) — either:

   • A file path was provided: read it and extract the claims, queries, axioms, and any candidate evidence attached to claims
   • No file provided: ask the user to type them or provide a path

2. Output directory (output= parameter) — where to write the results. If not provided, ask the user. If the user doesn't have one, suggest research/{slug}/ as a starting point.

3. Research ID (id= parameter) — optional. If not provided, auto-generate from the output directory slug or use RXXXX as a placeholder.

Step 2: Confirm and save input spec

Present a summary of what will be researched:

Research ID: R0005
Output directory: research/ai-trust/
Run date: 2026-03-31
Axioms: 0
Claims: 5 (2 with candidate evidence)
Queries: 2
  C001: First claim text...
  C002: Second claim text... [+1 candidate evidence URL]
  ...
  Q001: First query text...
  Q002: Second query text...

Ask the user to confirm before proceeding.

After confirmation, save the input specification as {output_directory}/research-input.json. This file enables future re-runs without re-specifying parameters.

Guardrail: Before writing research-input.json, check if one already exists in the output directory. If it does, this is NOT a new research run — it is an attempted reuse of an existing research directory. STOP and fail with an error:

ERROR: {output_directory}/research-input.json already exists.
This directory belongs to an existing research instance.
To re-run this research, use: /research rerun {output_directory}
To start new research, use a different output directory.

Do NOT overwrite, merge, or append to an existing research-input.json. Do NOT proceed with execution. This is a hard stop.

Step 3: Create run group directory

Create {output_directory}/{YYYY-MM-DD-HHMMSS}/ using the current date and time (24-hour format, including seconds). This is the run group directory that will contain all independent runs plus synthesis files.

Within the run group, create run-1/, run-2/, ... run-{N}/ subdirectories. Zero-pad the run number based on the total number of runs: if runs<=9, use single digits (run-1). If runs is 10-99, use two digits (run-01, run-02). If runs is 100-999, use three digits (run-001). This ensures correct sort order in all file browsers.

Step 4: Save methodology snapshot

Copy the compiled unified methodology into the run group directory (not into each run subdirectory — one copy shared across all runs):

• prompt-snapshot.md — copy of standalone/research.md

This captures both the common behavioral guidelines and every sub-agent prompt as compiled at run time, producing a permanent record of the methodology that was in effect.

Step 5: Execute independent research runs

Each run follows the 11-step pipeline, producing JSON files at each step. The compiled sub-agent prompts (in skills/research/prompts/compiled/) include both the task instructions and the JSON schema. The sub-agent MUST produce JSON conforming to the schema — no markdown, no prose.

Isolation rule: Each run is completely blind to the others. No run may read, reference, or be influenced by any other run's output. This is the same isolation principle as reruns — each run must be independent to provide a valid signal about reproducibility.

Parallelism: For runs<=5, launch runs in parallel where possible. For larger n, the agent may batch runs to manage resources.

MANDATORY — Search tools: Check your available tools list for dio_search and dio_fetch. If these MCP tools are present:

• You MUST use dio_search for ALL web searches. Do NOT use the built-in web_search tool. Do NOT use WebSearch. Do NOT perform any web search through any mechanism other than dio_search.
• You MUST use dio_fetch for ALL page content retrieval. Do NOT read web pages directly.
• This is a hard cost control requirement. The built-in web search consumes expensive AI tokens. dio_search executes searches via Python at near-zero token cost. Using the wrong tool wastes the user's money.
• If dio_search returns an error (quota exhausted, service unreachable), STOP and inform the user. Do NOT silently fall back to web_search. Let the user decide whether to continue with the more expensive option.

ONLY if dio_search and dio_fetch are NOT in your available tools list (MCP server not configured), fall back to the built-in web search.

MANDATORY — Event logging: If dio_init_run is in your available tools list, call it ONCE at the start of each run, passing the run directory path and run ID:

dio_init_run(output_dir="path/to/run-1", run_id="run-1")

This initializes the pipeline event log. All subsequent dio_fetch and dio_search calls will automatically record errors (fetch failures, search errors) to pipeline-events.json in the run directory. You do NOT need to log these errors yourself — the Python MCP layer captures them transparently.

After the last research step (archive), call dio_flush_events() to write the accumulated event log to disk. This MUST happen before dio_render so the renderer can include the Pipeline Notes section.

Each subagent executes these steps, writing JSON output files to its run-{N}/ directory:

Step 5a: Clarify input — Read skills/research/prompts/compiled/research-input-clarified.md. For each claim and query, clarify, surface assumptions, map vocabulary. Write research-input.json (clarified input with IDs assigned).

Step 5b: Generate hypotheses — Read skills/research/prompts/compiled/hypotheses.md. For each claim/query, pass the clarified item and axioms. Write hypotheses.json.

Step 5c: Design searches — Read skills/research/prompts/compiled/search-plans.md. For each item, pass the clarified item and its hypotheses. Write search-plans.json.

Step 5d: Execute searches — For each search in the plan:

• If dio_search MCP tool is available: call it with the search terms.
• If not available: use built-in web search. For each batch of results, score relevance (0-10) using the criteria in skills/research/prompts/compiled/search-results.md. Filter by score >= 5, deduplicate by URL. Write search-results.json.

Step 5e: Score sources — For each selected source:

• If dio_fetch MCP tool is available: call it to get page content.
• If not available: the subagent reads the source directly. Score reliability, relevance, and six bias domains per skills/research/prompts/compiled/scorecards.md. Write scorecards.json.

Step 5f: Extract evidence packets — Read skills/research/prompts/compiled/evidence-packets.md. For each item, pass the clarified item, its hypotheses, and the scorecards (including content_extract). The extractor emits verbatim excerpts tied to specific hypotheses (or search themes, in open-ended mode) with an explicit supports / refutes / nuances / context relationship. Write evidence-packets.json. This is the grounding layer between source scoring and synthesis — synthesis downstream should cite packets rather than paraphrase from memory.

Step 5f.1: Validate evidence packets — If dio_validate_packets is in your available tools list, call it IMMEDIATELY after writing evidence-packets.json:

dio_validate_packets(run_dir="path/to/run-1")

This runs the Python deterministic verbatim verifier against every packet. It checks each claimed excerpt against the source's content (using the server-side fetch cache from dio_fetch — even if you dropped content_extract from the scorecards, the cache has it). Non-verbatim packets are removed and verbatim_stats is rewritten with real numbers. Do NOT populate verbatim_stats yourself — the validator overwrites whatever you wrote with deterministic results.

Step 5g: Synthesize, assess, gaps — Read skills/research/prompts/compiled/synthesis.md. For each item, pass the scorecards, hypotheses, and evidence packets from Step 5f. Synthesize the evidence collection, produce probability assessment, identify gaps. The packets are the primary grounded input; the scorecards weight them. Write synthesis.json.

Step 5h: Self-audit — Read skills/research/prompts/compiled/self-audit.md. For each item, pass the full chain including evidence_packets. Audit the process, verify source interpretations (Step 9b traces assessment claims back to packet excerpts), produce reading list. Write self-audit.json.

Step 5i: Report — Read skills/research/prompts/compiled/reports.md. For each item, assemble the final report from all prior steps. Write reports.json.

Step 5j: Archive — Combine all JSON outputs into archive.json with a timestamp and pipeline version.

Step 5k: Usage — Write usage.json to the run group directory. The file MUST conform to src/diogenes/schemas/usage.schema.json. Use the canonical totals + per_call structure so plugin-path and CLI-path runs are directly comparable.

Claude Code does not expose per-API-call token counts, cache metrics, or per-call costs to skills. Set the fields you cannot observe to 0 rather than omitting or inventing values:

• totals.input_tokens, totals.output_tokens, totals.total_tokens, totals.estimated_cost_usd — set to 0.
• totals.api_calls — set to 0 unless you have an accurate count from Python-side MCP calls.
• totals.web_search_requests / totals.web_fetch_requests — set to the accurate count tracked via dio_search / dio_fetch calls.
• per_call — use an empty array [] if you do not have per-agent token breakdowns.

Plugin-specific extras (scoring distribution, verdict summaries, per-run stats, anything that does not fit the canonical shape) go under the optional top-level plugin_metadata key, free-form. Do NOT add new top-level keys — additionalProperties: false at the root will reject them.

Output files per run directory (all JSON, no markdown):

run-{N}/
├── research-input.json
├── hypotheses.json
├── search-plans.json
├── search-results.json
├── scorecards.json
├── evidence-packets.json
├── synthesis.json
├── self-audit.json
├── reports.json
└── archive.json

Step 5l: Synthesize across runs

This step runs ONLY after ALL runs have completed. It reads the JSON output from every run-{N}/ directory and produces group-level JSON files in the run group directory.

Produce these files:

1. group-synthesis.json — aggregate result from all runs:

   • For each claim/query: consensus verdict, divergences, union of sources
   • Overall assessment that integrates findings from all runs
   • Where runs agree: state with increased confidence
   • Where runs disagree: note the divergence and classify the root cause

2. group-consistency.json — similarity metrics across runs:

   • Source overlap (% shared between each pair, sources found in all/most/one)
   • Verdict agreement (did all runs support the same hypothesis?)
   • Scoring consistency (same source scored the same way?)
   • Overall similarity score
   • Diagnostic: if <50% overlap, flag as "query may be too ambiguous"

3. group-reading-list.json — consolidated reading list across all runs, deduplicated, with provenance (which runs found each source)

4. usage.json — combined usage across all runs

Step 6: Report completion

When synthesis finishes, report to the user:

## Research Complete

- **Runs**: {n} independent runs
- **Items investigated**: {count} claims/queries
- **Output**: {path to run group directory}
- **Duration**: {wall clock time}
- **Consistency**: {overall similarity score}%

### Verdict Summary (from synthesis)

| Probability | Count | Items |
|-------------|-------|-------|
| Almost certain | {n} | {list} |
| Very likely | {n} | {list} |
| Likely | {n} | {list} |
| Unlikely | {n} | {list} |

### Run Consistency

| Metric | Value |
|--------|-------|
| Source overlap (avg pairwise) | {n}% |
| Verdict agreement | {all agree / N of M agree} |
| Scoring consistency | {consistent / N divergences} |

### Flags for Attention

| Item | Issue |
|------|-------|
| {ID} | {description of issue needing correction} |

Search-tool preference

• Prefer dio_search / dio_search_batch for every web search inside /research. They use a configured provider (Serper, Brave, etc.) with a quota-based free tier and are cheaper per call than web_search.
• Never use web_search silently. It is reserved for fallback after dio_search reports a failure and the user has confirmed they want to continue.
• Structured error handling. When dio_search or dio_search_batch returns {"error": true, ...}, inspect error_kind:
  • quota_exhausted / rate_limited / network_error — tell the user, and if fallback_available is true, offer to continue the remaining searches with web_search. Make the token-cost change explicit in the offer so the user can decide.
  • auth_failed — the search provider's API key is missing/invalid. Report and stop; web_search cannot substitute for a misconfigured provider.
  • other — surface the message and ask the user how to proceed.

Constraints

• Do not proceed without user confirmation (Step 2)
• Do not modify files outside the output directory
• The output format spec is the source of truth for file structure
• Every search result must be dispositioned (no silent drops)
• Always produce a minimum of 3 hypotheses per claim or query. The first three are mandatory: H1 (affirmative/accurate), H2 (negative/ partially correct), H3 (nuanced/materially wrong). Additional hypotheses (H4, H5, ...) may be added when the evidence supports more than three distinct explanations.
• Run isolation: The subagent MUST NOT read prior run results. On a rerun, read ONLY research-input.json. Do NOT pass prior run paths to the subagent. Prior results bias new research through anchoring and confirmation effects.
• Claim extraction blindness: When extracting claims from a document (via extract or check), read ONLY the document body. Do NOT read or include the References section — it biases the fact-checker. Claims should be extracted as they appear in the prose.
• Unrecognized parameters: Silently ignore any key=value parameter not documented above. Do not ask about it.

Customization

The research pipeline produces JSON at every step. The compiled sub-agent prompts in skills/research/prompts/compiled/ contain both the task instructions and the JSON schema for each step's output. To customize, modify the source prompts in src/diogenes/prompts/sub-agents/ and/or the schemas in src/diogenes/schemas/, then run python scripts/compile-prompts.py to rebuild the compiled versions.

Markdown rendering from JSON is a separate step handled by the dio_render MCP tool or a fallback text description. The research methodology is independent of the rendering — you can change how results are presented without changing how research is conducted.

Future extensions

• dio_render MCP tool for JSON-to-markdown rendering
• report mode (topic exploration with facets)
• check-triggers command to evaluate revisit conditions (#87)