Research Online
Research a programming topic from multiple angles using parallel sub-agents, with critical evaluation of source credibility.
Usage
/research-online <library> <what you want to do>
/research-online "<error message>" <library>
/research-online how to implement auth in react-navigation v7
/research-online best practices for state management in React
/research-online Redux vs Zustand for large appGotchas
- Context7 docs may lag behind a recent major release. The indexed version could be outdated — check which version is documented before citing it as authoritative.
- "Prefer recent, then higher authority" can be wrong: an authoritative maintainer comment from 18 months ago may be more correct than a popular blog post from last month. Weigh authority first for stable libraries.
- Quick mode may miss nuance. If a seemingly simple question turns out to have a complex answer (e.g., "default port" but it depends on framework integration), note this in the synthesis and suggest re-running in Standard mode.
Workflow
Step 1: Parse Input
Extract from the user's query:
- Library/framework name (e.g., "react-navigation", "prisma", "next.js")
- Error message (if any, usually in quotes)
- Version (if mentioned, e.g., "v7", "version 14")
- Goal/intent (what they're trying to build or achieve)
- Problem description (if debugging: what's broken or unexpected)
- Comparison targets (if comparing: "X vs Y", "X or Y")
Step 2: Classify Query Depth
Classify the query to avoid over-researching simple questions:
| Mode | When | Behavior |
|---|---|---|
| Quick | Simple factual lookup, single API question, "what version supports X", "how to do X" with well-known library | Skip internal docs check. Spawn only Docs + General (max 2 agents). Skip follow-up loop and critique. Go straight to synthesis. |
| Standard | Comparisons, best practices, errors, complex implementation, "real world experience", debugging | Full workflow including follow-up loop and adversarial critique |
When in doubt, use Standard.
Step 3: Check Internal Documentation First (Standard mode only)
Before external research, use the Grep tool to search for relevant keywords in the project's docs:
Grep pattern: "<relevant_keywords>" path: "docs/" and "*.md"If found, read and include internal patterns/conventions in the synthesis. Internal docs often contain project-specific decisions that external research won't cover.
Step 4: Determine Which Agents to Spawn
| Agent | Spawn When | Purpose |
|---|---|---|
| Docs | Library/framework mentioned | Context7 documentation lookup |
| GitHub | Library with known repository | Search issues AND discussions |
| General | Always | Broad "how to" web search |
| Specific | Error message provided | Search for exact error text |
| StackOverflow | Common problem/implementation pattern | Community Q&A solutions |
| Changelog | Version mentioned OR "stopped working" / "after upgrade" | Breaking changes, migration guides |
| Best Practices | Feature implementation (no error message) | Architecture patterns, recommended approaches |
| Comparison, best practices, or "real world experience" queries | Candid developer opinions, warnings, real-world experience | |
| Comparison | Query contains "vs", "or", "compare", "which", "best library" | Compare options, pros/cons |
Step 5: Spawn Agents in Parallel
Use the Task tool to spawn ALL relevant agents in a single message (parallel execution). Each agent uses subagent_type: general-purpose.
Every agent must capture metadata for each source: URL, date, and source type (official docs / GitHub issue / blog / SO answer / forum).
| Agent | Tool | Search Strategy |
|---|---|---|
| Docs | Context7 | resolve-library-id then query-docs. If resolve-library-id returns no match, fall back to WebSearch for {lib} official documentation {goal} |
| GitHub | WebSearch | site:github.com {lib} "{terms}", then WebFetch top 2-3 |
| General | WebSearch | how to {goal} {lib} |
| Specific | WebSearch | "{exact_error_message}" {lib} |
| StackOverflow | WebSearch | site:stackoverflow.com {lib} {keywords}, then WebFetch top answers |
| Changelog | WebSearch | {lib} {version} changelog breaking changes migration |
| Best Practices | WebSearch | {lib} best practices {goal} + {lib} recommended architecture {goal} |
| WebSearch | site:reddit.com {lib} {keywords}, then WebFetch top 2-3 threads | |
| Comparison | WebSearch | {option_A} vs {option_B} {context} |
For full agent prompt templates with detailed instructions, see references/agent-prompts.md.
Step 6: Collect and Deduplicate Results
Wait for all agents to complete and gather their findings with metadata. Before evaluation, deduplicate: if multiple agents found the same URL or GitHub issue, keep the entry with the richest metadata and merge any unique context from the duplicates. Note which agents independently found the same source — convergence from multiple agents increases confidence.
Step 7: Critical Evaluation
Before synthesizing, evaluate each source:
Recency (adjust thresholds by library velocity):
| Age | Fast-moving libs (React, Next.js, etc.) | Stable libs (Express, lodash, etc.) |
|---|---|---|
| < 6 months | High | High |
| 6-18 months | Medium | High |
| 18-36 months | Low | Medium |
| > 3 years | Very Low | Low |
Authority:
| Source Type | Score |
|---|---|
| Official docs, changelogs, core team posts | High |
| GitHub issues (maintainer response) | High |
| GitHub issues (community), recent blogs (known author) | Medium |
| SO answers (accepted + >10 votes), comparison articles | Medium |
| Reddit threads (>50 upvotes or multiple experienced replies) | Medium |
| SO answers (not accepted, <10 votes), old blogs, old comparisons | Low |
| Reddit threads (<10 upvotes, few replies), random forums | Very Low |
Relevance: Exact error/goal match = High. Same library, similar task = Medium. Related concept = Low.
When sources conflict: Prefer more recent, then higher authority. Note conflicts in synthesis. If official docs conflict with recent issues, the issue may reveal a bug or undocumented behavior.
Step 8: Follow-Up Search Loop (Standard mode only)
After critical evaluation, check if any topic area has fewer than 2 sources or if the query's core question remains unanswered. If so:
- Identify the gap (e.g., "no sources found for how to handle token refresh specifically")
- Generate 1-2 targeted delta queries — more specific terms, alternative terminology, or broader scope
- Spawn 1-2 follow-up agents with the delta queries
- Merge new results into existing findings, deduplicate, re-evaluate
Max 1 follow-up cycle. If the gap persists, note it in the synthesis as a low-confidence area rather than searching again.
Step 9: Adversarial Critique (Standard mode only)
Brief self-challenge before presenting. Ask:
- What would someone who disagrees with this conclusion say?
- Are we over-weighting one source type? (e.g., all findings from blog posts, no official docs)
- Could any "independent" sources actually trace back to the same original? (e.g., 3 blogs all citing one tweet = 1 source, not 3)
- Is the recommended approach the simplest option, or are we over-engineering?
If the critique reveals a blind spot, adjust the synthesis and lower the confidence level. This is a brief internal review pass — do not spawn a separate agent.
Step 10: Present Results
Lead with the synthesis, not the raw data. The user wants the answer first, with supporting evidence.
Structure the output as:
- Synthesis — goal, recommended approach, confidence level, key findings weighted by credibility. Include the 1-3 most influential references with URLs — the sources that most shaped the conclusion.
- Supporting Details — only include sections relevant to the query, and only findings not already covered in the synthesis. Skip sections that would just repeat what's in the synthesis.
Available detail sections (include only those relevant):
- Documentation — Context7 findings
- GitHub Issues & Discussions — with date, type, authority
- Reddit — real-world experience, warnings, opinions with upvote context
- Comparison — (comparison queries only) pros/cons/best-for per option
- Specific Error Matches — (error queries only) causes and fixes
- Version/Changelog — (version queries only) breaking changes
- Conflicts — only if sources disagree on something material, with resolution
For the full output format template, see references/output-format.md.
Examples
Example 1: Feature Implementation
/research-online how to implement authentication in Next.js 14Spawns: Docs, GitHub, General, Best Practices, StackOverflow
Example 2: Comparison Query
/research-online Redux vs Zustand for large React appSpawns: Docs (both), General, Comparison, Reddit, StackOverflow
Example 3: Specific Error
/research-online "Cannot read property 'navigate' of undefined" react-navigationSpawns: Docs, GitHub, General, Specific, StackOverflow
Example 4: Version-Specific
/research-online auth navigation not working in react-navigation v7Spawns: Docs, GitHub, General, Changelog, StackOverflow
Example 5: Best Practices
/research-online best practices for folder structure in Express APISpawns: Docs, General, Best Practices, Reddit, StackOverflow
Example 6: Quick Mode (Simple Lookup)
/research-online what's the default port for Vite dev serverQuick mode: Spawns Docs + General only. Returns direct answer without follow-up loop or critique.
Troubleshooting
Agent fails or times out
Solution: Continue with remaining agents. Note the gap in the synthesis and which source types are missing. The research is still useful with partial results.
No results found for a query
Solution: Widen search terms — try without the library name, use alternative terminology, or search for the underlying concept rather than the specific implementation.
All sources are outdated
Solution: Flag explicitly in the synthesis. Note the dates and recommend the user verify against current documentation. Prefer official docs over old blog posts.
Sources conflict with each other
Solution: Weight by recency and authority. Note the conflict clearly in the Source Evaluation Summary with a resolution explaining which source to trust and why.
Notes
- All agents run in parallel for speed
- Each agent should complete in under 60 seconds
- Always capture source metadata for critical evaluation
- Weight findings by credibility in synthesis — a recent GitHub issue from a maintainer outweighs a 5-year-old blog post
- For comparisons, be especially careful about recency — library landscapes change quickly