Research

---

name: research description: Deep research powered by GREP — investigate APIs, libraries, companies, markets, and any topic with sourced citations. Build interactive HTML apps, slidedecks, and spreadsheets backed by research. Route to one of 27 domain experts (legal, medical, patent, financial, real estate, supply chain, maritime, etc.). Use when you need to research something, fact-check a claim, plan an implementation with real docs, create a new skill, build a deliverable, install Grep as an MCP server, check GREP account status, or upgrade your plan. Trigger on: "research", "look up", "investigate", "fact check", "what's the best way to", "how does X work", "find out about", "build me a", "make a slidedeck/spreadsheet", "use the X expert", "follow up on", "research using these files", "install grep MCP", "GREP", or any task where knowing current facts would improve the outcome. Use proactively before writing code against unfamiliar APIs.

GREP Research

One skill, multiple research modes. Routes to the right workflow based on what the user needs.

Script path

Scripts are bundled directly in this skill directory:

SCRIPTS_DIR="${CLAUDE_SKILL_DIR}/scripts"

Auto-update check

On first use per session, silently check for updates (throttled to once per hour):

node "${CLAUDE_SKILL_DIR}/scripts/update-check.js" 2>/dev/null &

Run in background — must never delay the user.

Auth gate

Before any operation, verify the user is authenticated:

node "${CLAUDE_SKILL_DIR}/scripts/auth.js" status

If "authenticated": false, do not tell the user to log in — instead, read ${CLAUDE_SKILL_DIR}/references/login.md and execute that workflow automatically. Resume the original request once authenticated.

Intent routing

Determine what the user wants, then read the matching reference file for the detailed workflow. Read exactly one reference file per invocation — do not load them all.

Route 1: Deep research (DEFAULT)

Read: ${CLAUDE_SKILL_DIR}/references/deep.md

When: This is the default. Use for any research request that doesn't clearly match another route. "Research X", "look up X", "investigate X", "what's the deal with X", "how does X work". Roughly 5 minutes.

Route 2: Quick fact check

Read: ${CLAUDE_SKILL_DIR}/references/quick.md

When: The user wants a fast, simple answer — not a deep dive. Cues: "quick", "fast", "just check", "what version is", "is X still supported", simple factual questions. About 25 seconds.

Route 3: Exhaustive investigation

Read: ${CLAUDE_SKILL_DIR}/references/ultra.md

When: The user explicitly wants exhaustive coverage and is willing to wait up to an hour. Cues: "exhaustive", "thorough", "everything about", "security audit", "legal research", "full survey", "leave no stone unturned". Use sparingly.

Route 4: Research-informed planning

Read: ${CLAUDE_SKILL_DIR}/references/plan.md

When: The user wants to research before building — they need best practices, architecture patterns, or API docs to inform an implementation plan. Cues: "plan", "before I build", "best approach to", "how should I implement", "what's the right way to".

Route 5: Create a new skill

Read: ${CLAUDE_SKILL_DIR}/references/skill-creator.md

When: The user wants to build a new AI agent skill. Cues: "create a skill", "build a skill", "make a skill for X", "skill that does Y", "SKILL.md".

Route 6: Account status / job check

Read: ${CLAUDE_SKILL_DIR}/references/status.md

When: The user asks about their GREP account, recent jobs, or wants to check on a running research job. Cues: "status", "my account", "my jobs", "check on job", "recent research".

Route 7: Upgrade / billing

Read: ${CLAUDE_SKILL_DIR}/references/upgrade.md

When: The user wants to change their subscription, check pricing, or hit a quota limit. Cues: "upgrade", "plan", "pricing", "subscribe", "credits", "quota".

Route 8: Login / auth

Read: ${CLAUDE_SKILL_DIR}/references/login.md

When: The user explicitly asks to log in or authenticate. Also auto-triggered by the auth gate above on 401. Cues: "login", "authenticate", "connect my account", "API key".

Route 9: Route to a domain expert

Read: ${CLAUDE_SKILL_DIR}/references/domain-expert.md

When: The user wants a specific domain specialist — legal, medical, patent, financial, real estate, supply chain, maritime, KYC/KYB, vehicle/VIN, government policy, etc. Cues: "use the legal expert", "kyc/kyb on", "due diligence on", "patent landscape", "clinical trials for", "decode VIN", "vessel tracking". Routes to one of 27 public experts via expert_id.

Route 10: Build an interactive HTML app

Read: ${CLAUDE_SKILL_DIR}/references/build-app.md

When: The user wants a runnable HTML/JS deliverable — dashboard, calculator, data explorer, interactive comparison tool. Cues: "build me a tool", "make an interactive", "create a dashboard for", "I want to be able to click/sort/filter". Runs the app-builder expert at effort=build (10-15 min).

Route 11: Build a slidedeck

Read: ${CLAUDE_SKILL_DIR}/references/build-slidedeck.md

When: The user wants a presentation. Cues: "make a slidedeck", "build a deck", "presentation about", "slides on". Produces HTML deck with arrow-key nav + PDF export.

Route 12: Build a spreadsheet

Read: ${CLAUDE_SKILL_DIR}/references/build-spreadsheet.md

When: The user wants tabular data. Cues: "make a spreadsheet", "build a table of", "list of X as a sheet", "comparison table with columns". Produces sortable HTML table with CSV export.

Route 13: Multi-step research workflow

Read: ${CLAUDE_SKILL_DIR}/references/research-workflow.md

When: The user wants research chained into a deliverable, or multiple coordinated jobs. Cues: "investigate X and then make a deck", "research X and build a tool", "do deep research on X then summarize as a spreadsheet". Orchestrates orient → deep dive → optional build artifact.

Route 14: Research with attached files

Read: ${CLAUDE_SKILL_DIR}/references/with-context.md

When: The user has files (PDFs, CSVs, images) that should be inputs to the research. Cues: "research using these PDFs", "summarize this report", "compare these documents", "extract insights from this CSV". Uploads as attachments, references via attachment_ids.

Route 15: Continue / follow up on existing job

Read: ${CLAUDE_SKILL_DIR}/references/continue.md

When: The user wants to extend a prior research job rather than start a new one. Cues: "follow up on that", "go deeper on the X you found", "continue job ", "ask a follow-up". Inherits prior context via POST /api/v2/research/<slug>/continue.

Route 16: Install Grep as an MCP server

Read: ${CLAUDE_SKILL_DIR}/references/mcp.md

When: The user wants to wire Grep into another agent's .mcp.json (Cursor, Cline, Continue, etc.) as a native MCP server with 4 tools. Cues: "install grep MCP", "add grep to my mcp config", "grep as MCP server".

Shared: gathering codebase context

Routes 1, 4, 5, 9, 10, 11, 12, 13, and 14 benefit from sending codebase context alongside the research query. When the reference file calls for context gathering, use this pattern:

CONTEXT_FILE=$(mktemp /tmp/grep-research-context.XXXXXX)

# Project conventions
if [ -f CLAUDE.md ]; then
  echo "=== PROJECT CONVENTIONS (CLAUDE.md) ===" >> "$CONTEXT_FILE"
  head -60 CLAUDE.md >> "$CONTEXT_FILE"
  echo "" >> "$CONTEXT_FILE"
fi

# Package manifest (tells GREP what stack you're on)
for manifest in package.json pyproject.toml requirements.txt Cargo.toml go.mod; do
  if [ -f "$manifest" ]; then
    echo "=== $manifest ===" >> "$CONTEXT_FILE"
    head -40 "$manifest" >> "$CONTEXT_FILE"
    echo "" >> "$CONTEXT_FILE"
  fi
done

Additionally, read any files directly relevant to the research topic (existing integration code, config files, etc.) and append them to $CONTEXT_FILE. Use your judgement — if the user is researching "Stripe webhooks" and there's a payments/ directory, include those files.

Always clean up after: rm -f "$CONTEXT_FILE"

Shared: execution pattern

Always use Monitor for research commands, never blocking Bash. Research takes minutes. Blocking Bash ties up the agent, prevents interaction, and risks timeout. Monitor runs in the background and streams live updates.

• Deep research: Monitor with timeout_ms: 560000
• Quick research: Monitor with timeout_ms: 80000
• Ultra research: non-blocking submit, then /loop polling (see ultra.md)

When Monitor completes, always read the output and present results to the user. A research job that completes without presenting results is a failed mission.

Presenting results

1. Lead with the key answer or insight
2. Organise by theme or relevance
3. Preserve source citations from the report
4. Note any conflicting information
5. When using research to inform code: extract concrete facts (endpoints, auth formats, required fields) before writing code — don't just dump the report

Anti-patterns

• Do NOT load multiple reference files at once — route to exactly one
• Do NOT default to ultra research — start with deep, escalate only if genuinely needed
• Do NOT skip auth check — unauthenticated calls waste time with cryptic errors
• Do NOT use blocking Bash for research commands — always Monitor (or background submit for ultra)
• Do NOT re-submit a query if a previous job is still running — use status to retrieve it