Skip to content

Instantly share code, notes, and snippets.

@HelgeSverre
Created March 27, 2026 11:22
Show Gist options
  • Select an option

  • Save HelgeSverre/9d48bc1bcdd6c6dca13edf0cbb3f9710 to your computer and use it in GitHub Desktop.

Select an option

Save HelgeSverre/9d48bc1bcdd6c6dca13edf0cbb3f9710 to your computer and use it in GitHub Desktop.
Agentic Research Skill — structured workflow for parallel AI research with confidence markers and synthesis
name agentic-research
description Use when the user needs to research multiple systems, products, or technologies in parallel and synthesize findings into actionable recommendations. Triggers on competitive analysis, technology comparison, market research, "research X vs Y", architectural research across multiple systems, or any investigation requiring structured multi-source analysis.

Agentic Research Workflow

A structured workflow for researching multiple systems in parallel and synthesizing findings into actionable output. Based on the research brief pattern — define objectives, dispatch parallel agents, gather structured reports, synthesize across systems.

When to Use This

  • Comparing multiple products, tools, or technologies
  • Competitive analysis ("how do others solve this problem?")
  • Architecture research before building something new
  • Investigating a domain you're unfamiliar with
  • Any task where you need to research 3+ systems and draw conclusions

Decision Tree

Research task received → How many systems?
│
├─ 1 system (deep dive)
│  └─ Single agent, use "Research Techniques" below, skip synthesis
│
├─ 2 systems (comparison)
│  └─ Write brief → 2 parallel agents → comparative synthesis
│
├─ 3+ systems (survey)
│  └─ Write brief → N parallel agents → full synthesis with feature matrix
│
└─ Unknown / exploratory ("how do others solve X?")
   └─ WebSearch to identify systems first → then treat as 3+ system survey
Per-system research → What sources are available?
│
├─ Has public API docs / OpenAPI spec?
│  └─ Start here — API shapes reveal the data model directly
│     WebFetch the docs URL, look for entity names and relationships
│
├─ Has help center / knowledge base?
│  └─ Second priority — "How to set up X" articles reveal workflows
│     WebFetch help center pages, follow links 2-3 levels deep
│
├─ Has open-source code / SDK?
│  └─ Check GitHub — types, schemas, and entity definitions are gold
│     Use Bash: gh repo view, gh api to explore
│
├─ Marketing site only?
│  └─ WebSearch for reviews (G2, Capterra), blog posts, comparison articles
│     These often contain more detail than the marketing site itself
│
└─ Very little public info?
   └─ Mark confidence 🔴, state what's unknown, spend less time here
      An honest "unknown" is more useful than a plausible guess

Phase 1: Research Brief

Before touching any system, write a brief that defines:

1. Objective

One paragraph: what are we trying to learn, and why?

## Objective

Research and document how [domain] systems handle [specific problem].
The goal is to understand [what patterns exist / how others solve this]
to inform [our own design / a decision / a recommendation].

2. Key Design Questions

3-9 specific questions that drive the research. These are NOT "tell me about X" — they're decision-forcing questions:

### Key Design Questions

1. **[Specific design decision]**: How do different systems handle [X]?
   Is the approach [A-centric or B-centric]?
2. **[Trade-off]**: When [scenario], do systems choose [approach A] or
   [approach B]? What are the consequences?
3. **[Gap detection]**: What capabilities do users need that current
   systems don't provide?

Bad questions: "What is X?" / "Tell me about Y" (too vague, no decision to inform) Good questions: "Is the data model person-centric or credential-centric?" (forces comparison)

3. Systems to Research

List all systems with URLs and categories:

### Systems to Research

| System | URL | Category |
| --- | --- | --- |
| System A | https://... | Category 1 |
| System B | https://... | Category 1 |
| System C | https://... | Category 2 (adjacent) |

Include adjacent/analogous systems from related domains — they often have the best ideas.

4. What to Look For

Ordered by priority — what information matters most:

### What to Look For (priority order)

1. **Data model** — Core entities, relationships, cardinality
2. **Terminology** — What words does each system use for the same concepts?
3. **Workflows** — Key user journeys, step by step
4. **Integration model** — APIs, webhooks, native integrations
5. **Strengths and limitations** — What works, what doesn't, what's missing

5. Per-System Report Template

Define the structure every report should follow:

### Per-System Report Structure

1. System Overview (what, who, positioning)
2. Glossary (system term → generic concept → description)
3. Data Model (entity-relationship diagram + written description)
4. Key Workflows (as diagrams)
5. Strengths & Limitations
6. Sources (all URLs consulted)

6. Confidence Markers

Require confidence annotations on every section:

  • 🟢 High — API docs, schema, or primary source available
  • 🟡 Medium — Inferred from help docs, UI screenshots, or demos
  • 🔴 Low — Guessed from marketing copy or reviews

"Data model: Not publicly documented. Inferred from UI screenshots and help articles." is more useful than a confident-sounding fabrication.

Phase 2: Parallel Research Dispatch

Each system can be researched independently and in parallel. There are no dependencies between systems.

Agent Dispatch Pattern

For each system in the research brief:
  Launch a sub-agent with:
    - The system's entry URL
    - The research brief's "What to Look For" section
    - The per-system report template
    - Instructions to use web search, web fetch, and any available tools
    - The confidence marking requirements

Research Techniques (Per Agent)

Each agent should use these tools in priority order:

Priority Source Type Tool What to Extract
1 API docs / OpenAPI specs WebFetch on docs URL Entity names, field types, relationships, endpoint shapes
2 Help center / knowledge base WebFetch on help articles, follow links 2-3 deep Workflows ("How to set up X"), terminology, form fields
3 Open source code / SDKs Bash with gh api / gh repo view Type definitions, schema files, entity structures
4 Third-party analysis WebSearch for "[system] review", "[system] vs [competitor]" Features, limitations, user pain points
5 Product pages WebFetch on main site Positioning, pricing, feature lists
6 Demo videos WebSearch for "[system] demo" on YouTube UI structure, workflows invisible in docs
7 Job postings WebSearch for "[system] engineer" Tech stack, internal terminology

Concrete patterns:

# Find API docs
WebSearch: "[system name] API documentation"
WebSearch: "[system name] OpenAPI swagger"

# Deep-read help center
WebFetch: https://docs.example.com/getting-started
# Then follow links found in the content

# Check GitHub for schemas/types
WebSearch: "[system name] github"
# If repo found:
gh api repos/org/repo/contents/src/types --jq '.[].name'

# Find reviews with specific details
WebSearch: "[system name] review G2 Capterra 2025"
WebSearch: "[system name] vs [competitor] comparison"

Per-System Report Output

Each agent produces a standalone report following the template. Reports should be:

  • Self-contained — Readable without context from other reports
  • Evidence-linked — Every claim links to a source URL
  • Confidence-marked — Every section has a 🟢🟡🔴 marker
  • Honest about gaps — "Unknown" > plausible guess

Phase 3: Synthesis

After all individual reports are complete, synthesize across systems.

Cross-System Comparison

For each Key Design Question from the brief:

## Question 1: [The question]

| System | Approach | Details | Confidence |
| --- | --- | --- | --- |
| System A | Approach X | [specifics] | 🟢 |
| System B | Approach Y | [specifics] | 🟡 |
| System C | Approach X (variant) | [specifics] | 🟡 |

**Pattern**: Most systems use Approach X because [reason].
System B's Approach Y is interesting because [reason] but has the downside of [limitation].

**Recommendation**: [Specific recommendation with reasoning]

Pattern Identification

Look across all reports for:

  • Convergence — Where do most systems agree? This is likely the right approach.
  • Divergence — Where do systems disagree? This is where interesting design decisions live.
  • Gaps — What do users need that no system provides well? This is opportunity.
  • Terminology patterns — What words does the industry use? Adopt the dominant vocabulary unless there's a good reason not to.

Feature Matrix

| Capability | System A | System B | System C |
| --- | --- | --- | --- |
| Feature 1 | ✅ Full support | ⚠️ Partial | ❌ Missing |
| Feature 2 ||||
| Feature 3 || ✅ Best in class | ⚠️ |

Phase 4: Actionable Output

Research is only valuable if it produces decisions. The final output should include:

Executive Summary (1-2 pages)

  • Top 3-5 findings that should influence design decisions
  • Feature matrix comparison table
  • Recommended approach for each Key Design Question, with citations

Detailed Recommendations

For each key decision:

### Recommendation: [Decision]

**Adopt**: [Specific approach], as used by [System A] and [System C]
**Why**: [Reasoning based on evidence from research]
**Avoid**: [Alternative approach] because [evidence-based reasoning]
**Open question**: [What we still don't know and how to find out]

Extensions (Optional, Based on Scope)

Research → Data Model

When research informs a system you're building, extend findings into an entity-relationship sketch:

## Proposed Data Model (informed by research)

Based on [System A]'s approach to [concept] and [System C]'s handling of [concept]:

[Mermaid erDiagram or written description]

Key decisions:
- [Entity X] is the central entity because [evidence from research]
- [Relationship Y] uses [pattern] based on [System B]'s approach

Research → Spec Draft

## Feature Spec: [Feature Name]

### Background
[Summary of research findings relevant to this feature]

### Requirements (informed by research)
- [Requirement derived from competitive analysis]
- [Requirement addressing gap identified in research]

### Design
[Design decisions justified by research findings]

Research → Prototype

For validation before committing to full implementation:

## Prototype Scope

Based on research, validate these hypotheses with a minimal prototype:
1. [Hypothesis derived from research] → Build: [minimal UI/API to test]
2. [Hypothesis] → Build: [minimal test]

Quality Standards

A completed research project must have:

  • Research brief with clear objectives and key design questions
  • Individual reports for every system (no gaps, no skipped systems)
  • Confidence markers on every section of every report
  • Cross-system comparison answering every key design question
  • Feature matrix with all systems and capabilities
  • Specific, actionable recommendations (not just "it depends")
  • Honest handling of unknowns ("we couldn't determine X" > guessing)
  • All sources cited with URLs

Anti-Patterns

Anti-Pattern Why It's Bad Do Instead
Starting research without a brief You'll waste time on irrelevant details Write the brief first, even if brief
Researching systems sequentially 5x slower than parallel Dispatch all agents at once
Over-researching one system Depth without breadth Equal effort per system, synthesis is where value lives
Fabricating when info is unavailable Undermines trust in all findings Mark confidence 🔴, state what's unknown, move on
"It depends" conclusions Not actionable Recommend a specific approach with reasoning
Skipping adjacent domains Misses the best ideas Always include 2-3 analogous systems from related industries
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment