name: research-mode description: Anti-hallucination research mode. Toggle on to enforce citation requirements, source grounding, and "I don't know" behavior. Toggle off for creative work.

Research Mode

Activates anti-hallucination constraints based on Anthropic's documentation. Stay in this mode until the user says to exit.

Source: Anthropic - Reduce Hallucinations

Constraints (ALL active simultaneously)

1. Say "I don't know"

If you don't have a credible source for a claim, say so. Don't guess. Don't infer. "I don't have data on this" is always a valid answer.

2. Verify with citations

Every recommendation, claim, or piece of advice must cite a specific source:

• A file in the current project
• An external source found via web search (with URL)
• A named expert, paper, or researcher
• Official documentation

If you generate a claim and cannot find a supporting source, retract it. Do not present it.

3. Direct quotes for factual grounding

When working from documents, extract the actual text first before analyzing. Ground your response in word-for-word quotes, not paraphrased summaries. Reference the quote when making your point.

Source lookup order (ENFORCED -- follow this cascade)

Check sources in this order. Stop at the first level that answers the question.

Level 1 -- Local files (zero cost): Use Grep and Read to search the current project. Canonical files, docs, code, and config are the cheapest, most reliable sources. If the claim is about this project, local files ARE the citation.

Level 2 -- WebSearch snippets only (low cost): Run WebSearch. The result snippets usually contain the key fact. Cite as: "According to [Source Name] ([URL]): [snippet text]". DO NOT call WebFetch unless the snippet is ambiguous or incomplete.

Level 3 -- WebFetch for direct quotes (high cost, use sparingly): Only fetch full pages when:

• The snippet is ambiguous and you need the surrounding context
• The user explicitly asked for direct quotes or full text
• You need to verify a specific number, date, or technical detail the snippet doesn't include

Level 4 -- Scholar Gateway (for academic claims): For academic papers or research findings, use Scholar Gateway MCP if available. Structured results, no page scraping.

Token budget

• Maximum 5 WebSearch calls per research question
• Maximum 3 WebFetch calls per research question
• If you hit the limit: summarize what you found, list what remains unverified, and ask the user if they want to go deeper
• Parallel searches are fine. Serial retry loops are not.

What counts as "cited"

• Local file path + line number = cited
• WebSearch snippet + URL = cited
• Named paper/author + year = cited (mark {{VERIFY_URL}} if no link)
• "I recall from training data" = NOT cited. Say "I believe X but cannot cite a specific source."

What this mode is NOT

• It is NOT the default. Creative thinking, brainstorming, and novel ideas don't require this mode.
• It does NOT mean "be slow." Research efficiently. Use tools in parallel.
• It does NOT mean "only use existing ideas." You can synthesize across sources to reach new conclusions, but the inputs must be grounded.

How to exit

Say "exit research mode" or switch to any other task.

Research Mode for Claude Code

Anti-hallucination toggle for Claude Code. Activates citation constraints from Anthropic's documentation that force Claude to cite sources, say "I don't know" when unsure, and ground responses in direct quotes.

Using Kipi Founder OS? Research mode is already built into kipi-core. Just run /q-research <topic>. No extra install needed.

Install

Works in Claude Code CLI, the macOS desktop app, and the VS Code extension. Open Claude Code and run these two commands:

Step 1 -- Add the marketplace:

/plugin marketplace add assafkip/research-mode

Step 2 -- Install the plugin:

/plugin install research-mode@assafkip-research-mode

That's it. To confirm it installed, run /plugin and check that research-mode appears in your list.

Alternative -- install as a standalone skill: If you prefer not to use the plugin system, clone this repo and copy the SKILL.md file into your project's .claude/skills/research-mode/ directory.

Use

/research-mode:research

Or with a topic:

/research-mode:research what caused the Change Healthcare breach

Say "exit research mode" to turn it off.

What it does

Three constraints activate simultaneously:

1. Say "I don't know" -- no guessing, no inferring. If there's no credible source, Claude says so.
2. Cite everything -- every claim must reference a file, URL, paper, or named source. Unsourced claims get retracted.
3. Quote first, then analyze -- responses are grounded in word-for-word quotes from source material, not paraphrased summaries.

Source cascade (keeps token costs low)

Sources are checked in order. Claude stops at the first level that answers the question:

1. Local files (Grep + Read) -- zero cost, most reliable
2. WebSearch snippets -- cite the snippet directly, skip downloading full pages
3. WebFetch -- only when the snippet is ambiguous or user needs full quotes
4. Scholar Gateway -- for academic papers, if available

Token budget: 5 WebSearch, 3 WebFetch max per question. If the limit is hit, Claude summarizes what it found and asks before going deeper.

What it doesn't do

• Not always-on. It's a toggle. Turn it on for research, off for creative work.
• Not slow. Claude still uses tools in parallel and works efficiently.
• Not restrictive on new ideas. You can synthesize across sources, but inputs must be grounded.

Why

LLMs hallucinate. When you're doing research that matters, you need guardrails that force citation discipline. This plugin packages Anthropic's own recommendations into a one-command toggle.

Troubleshooting

"Plugin not found" after Step 1: Make sure you typed the marketplace command exactly as shown. The marketplace name is assafkip/research-mode (GitHub username/repo).

Command doesn't appear after install: Run /reload-plugins to refresh, or restart Claude Code.

Want to scope it to one project only: Add --scope project to the install command in Step 2.

Built by

Assaf Kipnis -- built while running GTM, investor outreach, and content ops for KTLYST entirely through Claude Code. When your AI assistant is writing your pitch decks, researching competitors, and drafting investor briefs, hallucinated facts aren't a minor annoyance. They're a credibility risk. This toggle exists because I needed it.

Questions or feedback: [REDACTED]