Research

Ask scoping questions one at a time or in small batches:

• Main research questions / what specifically to explore
• Timeframe (last 2 years? historical?)
• Geographic scope (if relevant)
• Audience (who will read this?)
• Purpose (what decisions will this inform?)
• Source preferences (academic, industry, news, docs)
• Specific angles, perspectives, or controversies to explore

If the request is already clearly scoped (e.g. "research the impact of tariffs on EU automotive exports since 2024"), skip redundant questions — just confirm scope and proceed.

Step 2: Route

Once intent is clear, ask: "Should I run this research now, or generate an optimized prompt you can use in another tool (OpenAI, Gemini, etc.)?"

• Run now → proceed to Step 3 (configure) then Step 4 (execute pipeline)
• Generate prompt → proceed to Prompt Output section below

Step 3: Configure

Before dispatching, allow overrides:

• Output path — default: reports/{topic-slug}-{YYYY-MM-DD}/
• Mode — deep (default) or quick
• Creative — true or false (default: false)

Default to deep mode unless the user signals quick: "quick look at", "brief overview of", "what's the deal with".

Create the output directory and a research/ subdirectory inside it using Bash before starting the pipeline.

Step 4: Execute Pipeline

Run the multi-agent pipeline. Each agent communicates through files in the output directory.

4.1: Plan

1. Read planner-prompt.md from this skill directory
2. Inject into the template: research brief, mode, creative flag, output path
3. Dispatch via Agent tool → wait for completion
4. Verify {output-path}/research/plan.md exists and contains sub-questions

4.2: Research

1. Read researcher-prompt.md from this skill directory
2. Inject into the template: research brief, mode, output path, path to research-recipes.md
3. Dispatch via Agent tool → wait for completion
4. Verify {output-path}/research/sources.md and {output-path}/research/notes.md exist

4.3: Source Review Gate

1. Read source-reviewer-prompt.md from this skill directory
2. Inject into the template: mode, output path
3. Dispatch via Agent tool → wait for completion
4. Parse verdict from agent response
5. If FAIL with CRITICAL issues:
   • Re-read researcher-prompt.md, inject original context + reviewer's CRITICAL issues into {REVIEWER_FEEDBACK}
   • Re-dispatch researcher → wait → re-dispatch source reviewer
   • Repeat up to 3 times. If still failing, present remaining issues to the user and ask whether to proceed or manually intervene
6. If PASS: continue to 4.4

4.4: Write

1. Read writer-prompt.md from this skill directory
2. Inject into the template: research brief, mode, creative flag, output path, path to report-template.md
3. Dispatch via Agent tool → wait for completion
4. Verify {output-path}/report.md exists

4.5: Report Review Gate

1. Read report-reviewer-prompt.md from this skill directory
2. Inject into the template: mode, creative flag, output path, path to report-template.md
3. Dispatch via Agent tool → wait for completion
4. Parse verdict from agent response
5. If FAIL with CRITICAL issues:
   • Re-read writer-prompt.md, inject original context + reviewer's CRITICAL issues into {REVIEWER_FEEDBACK}
   • Re-dispatch writer → wait → re-dispatch report reviewer
   • Repeat up to 3 times. If still failing, present remaining issues to the user and ask whether to proceed or manually intervene
6. If PASS: continue to 4.6

4.6: Present

Report the output path and a brief summary to the user.

Prompt Output

When the user wants a prompt for external tools, generate a structured prompt:

### TASK
[Clear, specific research objective]

### CONTEXT/BACKGROUND
[Why this matters and how it will be used]

### SPECIFIC QUESTIONS OR SUBTASKS
1. [First specific question]
2. [Second specific question]
...

### KEYWORDS
[Relevant search terms and concepts]

### CONSTRAINTS
- Timeframe: [specified timeframe]
- Geography: [specified scope]
- Source Types: [preferred sources]

### OUTPUT FORMAT
[Preferred format with specific requirements]

### FINAL INSTRUCTIONS
Remain concise, reference sources accurately, and provide evidence-based analysis.

Output directly in conversation. Optionally save to a file if the user requests it.

Self-Healing

• Agent dispatch fails for any step: Fall back to running that step inline (in the main conversation). Warn the user about context usage.
• Review loop exhausted (3 iterations): Present the reviewer's remaining issues to the user and ask whether to proceed or manually intervene.
• Artifact file missing after agent completes: Re-dispatch the agent once. If still missing, report the error and which file is absent.
• Output directory not writable: Check permissions, suggest an alternative path, or ask the user where to save.

Behavioral Guidelines

• Trigger on research intent (investigate, deep dive, report, sources) — not simple factual questions Claude can answer from training data
• When in doubt: "Would you like me to do a thorough research investigation, or just answer from what I know?"
• If the user asks for both a research run AND a prompt, do both — run the pipeline and also output the prompt
• Each prompt template has a {REVIEWER_FEEDBACK} placeholder. On first dispatch, leave it empty. On fix-mode re-dispatch, inject the reviewer's CRITICAL issues.
• Never pass session history to agents. Construct each dispatch prompt fresh from the template + injected values.
• See report-template.md for report structure and research-recipes.md for search patterns