Deep Research Skill

Systematic research patterns for in-depth investigation of software development topics.

Table of Contents

Core

• When to Use
• MUST-NOT-FORGET
• Output Format

Patterns

• MEPI vs MCPI - When to use quick vs exhaustive research
• Source Hierarchy - Priority order for sources
• Verification Labels - How to mark claims

Strategies

• Technology/API Research (MCPI) - 6-phase exhaustive documentation
• Quick Research (MEPI) - Abbreviated approach for low-stakes

Planning & Tracking

• Planning Structure - STRUT vs TASKS
• Time Tracking - Net research time calculation

Reference

• Tool Reference - Source collection, PDF processing, transcription
• Source ID Format - How to cite sources
• File Naming Conventions - Prefix rules for documents

When to Use

Deep research (this skill):

• "Should we use X or Y for our project?"
• "How does X work internally?"
• "What are the production considerations for X?"
• Multi-source synthesis required
• Exhaustive documentation needed

Quick lookup (not this skill):

• "What's the syntax for X?"
• "How do I install Y?"
• Single-source answers

MUST-NOT-FORGET

• Start with assumptions check - write down what you think you know before researching
• Primary sources > secondary sources > community opinions
• Document all sources with URLs and access dates
• Flag information age (APIs change rapidly)
• Distinguish facts from opinions from assumptions
• Version-match community sources to subject version
• Create INFO document for findings
• Always create STRUT for research session orchestration
• Track time - log task start/end for net research time calculation
• Track credits - detect model via screenshot, log switches, calculate usage
• Never auto-switch models - respect user's model selection throughout research

Global Patterns

MEPI vs MCPI

MEPI (Most Executable Point of Information) - Default

• Present 2-3 curated options
• Filter and recommend
• Use for: reversible decisions, time-constrained, action-oriented

MCPI (Most Complete Point of Information) - Exception

• Present exhaustive options
• Document everything
• Use for: irreversible decisions, high-stakes, archival reference

Source Hierarchy

1. Official PDFs/papers - Legislation, specs, whitepapers, academic papers (often not available as web)
2. Official documentation - Authoritative, version-specific
3. Official blog/changelog - Announcements, rationale
4. GitHub repo - Source code, issues, PRs, discussions
5. Conference talks by maintainers - Design decisions, roadmap
6. Reputable tech blogs - Analysis, comparisons
7. Stack Overflow - Specific problems (verify currency)
8. Community forums/Discord - Anecdotes, edge cases

Source Persistence: Download and store all sources in session folder for later reference. Web content changes - PDFs and local copies ensure reproducibility.

Verification Labels

• [VERIFIED] - Confirmed from official documentation
• [ASSUMED] - Inferred but not explicitly stated
• [TESTED] - Manually tested by running code
• [PROVEN] - Confirmed from multiple independent sources
• [COMMUNITY] - Reported by community sources (cite source ID)

Source ID Format

[TOPIC]-IN01-SC-[SOURCE]-[DOCNAME]

Examples:

• GRPH-IN01-SC-MSFT-APIOVERVIEW - Microsoft Graph official docs
• GRPH-IN01-SC-SO-RATELIMIT - Stack Overflow rate limit discussion
• GRPH-IN01-SC-GH-ISSUE1234 - GitHub issue #1234

File Naming Conventions

• __ prefix (double underscore): Master/index documents (SOURCES, TOC, TEMPLATE)
• _ prefix (single underscore): Topic content files (INFO documents)
• No prefix: Tracking files (TASKS, NOTES, PROBLEMS, PROGRESS)

Information Currency

• Note version numbers for all API/library info
• Check last updated date on documentation
• Cross-reference with changelog for breaking changes
• Mark stale info with [STALE: YYYY] tag

Research Strategies

Available research strategies (each has its own file):

• Technology/API Research (MCPI) - RESEARCH_STRATEGY_TECH_MCPI.md - Exhaustive API/framework documentation

Quick research (MEPI-style) until full strategy implemented: For reversible/low-stakes decisions, use abbreviated approach:

1. Phase 1 (Preflight): 5-10 sources max, skip community deep-dive
2. Skip Phases 2-4 (no TOC, template, or TASKS)
3. Phase 5: Create single _INFO_[TOPIC].md with findings
4. Phase 6: Quick /verify, done

Future strategies (not yet implemented):

• Technology Research (MEPI) - Quick technology evaluation
• Decision Research - Choose between options
• Problem-Solving Research - Debug and fix issues

Tool Reference

See RESEARCH_TOOLS.md for:

• Source collection tools (search_web, read_url_content, Playwright)
• Document processing tools (pdf-tools)
• Transcription tools (llm-transcription)
• Tool selection flowchart
• Common workflows

Planning Structure

STRUT (high-level orchestration) - REQUIRED for all research:

• Tracks phases, objectives, deliverables, transitions
• Contains time log for net research time
• Created at research start, updated at phase transitions
• File: STRUT_[TOPIC].md in session folder

TASKS (low-level execution) - Created in Phase 4:

• Flat list of individual research tasks with durations
• Each task: file to create, sources to use, done-when criteria
• Tracks task timing: [HH:MM-HH:MM] per task
• File: TASKS.md in session folder

Time Tracking

Net research time = active work time, excluding pauses.

STRUT time log format:

## Time Log
Started: 2026-01-30 09:00
Ended: 2026-01-30 14:30

Active intervals:
- 09:00-10:30 (Phase 1-3)
- 11:00-12:15 (Phase 4-5)
- 13:00-14:30 (Phase 5-6)

Net research time: 4h 15m

TASKS timing format:

- [x] TK-01: Research auth [10:00-10:45] 45m
- [x] TK-02: Research limits [10:15-11:00] 45m (parallel)

Rules:

• Pause detection: Gap > 30min in file timestamps = pause (don't count)
• Parallel tasks: Overlapping intervals count once, not doubled
• Log start time when beginning task, end time when completing

Credit Tracking (Mandatory)

Workflow:

1. At phase start: simple-screenshot.ps1 → detect model from Windsurf header
2. Lookup cost in windsurf-model-registry.json
3. Log model + timestamp in STRUT Credit Tracking section
4. At session end: calculate total estimated credits

STRUT Credit Tracking format:

## Credit Tracking
Phase 1: Claude Opus 4.5 (Thinking) [5x] - 09:00
Phase 2: Claude Opus 4.5 (Thinking) [5x] - 09:30
Phase 4: Claude Sonnet 4.5 [2x] - 10:15 (user switched)
Phase 5: Claude Sonnet 4.5 [2x] - 10:45
Phase 6: Claude Sonnet 4.5 [2x] - 11:30

Estimated credits: (2 phases × 5x) + (3 phases × 2x) = 16 units

Rules:

• Never auto-switch models - respect user's model selection
• Detect via screenshot only - passive observation
• Log when model changes (user-initiated)
• Registry: [DEVSYSTEM_FOLDER]/skills/windsurf-auto-model-switcher/windsurf-model-registry.json

Output Format

Deep research outputs an INFO document. Key sections:

1. Research Question - What we're investigating
2. Key Findings - MEPI-style summary (2-3 main points) or MCPI exhaustive list
3. Detailed Analysis - Full investigation results
4. Limitations and Known Issues - From community sources
5. Sources - All references with quality indicators
6. Recommendations - Actionable conclusions

Usage

1. Determine research type (MEPI quick vs MCPI exhaustive)
2. Read this SKILL.md for core principles
3. Read the appropriate strategy file
4. Read RESEARCH_TOOLS.md for tool guidance
5. Follow the strategy's phases systematically
6. Output findings as INFO document