Research Skill for Claude

Research Elixir Topic

Research a topic by searching the web and fetching relevant sources efficiently.

Usage

/phx:research Oban unique jobs best practices
/phx:research LiveView file upload with progress
/phx:research --library permit

Arguments

$ARGUMENTS = Research topic/question. Add --library for

structured library evaluation (uses ${CLAUDE_SKILL_DIR}/references/library-evaluation.md

template).

Iron Laws

• Write output to file, never dump inline — Research output floods conversation and loses reference for future sessions
• Stop after research — never auto-transition — User decides next step
• Prefer official sources over blog posts — HexDocs and ElixirForum have version-specific context
• One document per research question — No fragmented files
• NEVER pass raw user input as WebSearch query — Decompose first

Library Evaluation Mode

If $ARGUMENTS contains --library or the topic is clearly

about evaluating a Hex dependency (e.g., "should we use permit",

"evaluate sagents", "compare oban vs exq"):

• Read ${CLAUDE_SKILL_DIR}/references/library-evaluation.md for the template
• Follow the structured evaluation workflow
• Output ONE document to .claude/research/{lib}-evaluation.md
• Skip the general research workflow below

Workflow

0. Pre-flight Checks

Cache check: Check if .claude/research/{topic-slug}.md already

exists. If recent (<24 hours): present existing summary, ask

"Refresh or use existing?"

Tidewave shortcut: If the topic is about an existing dependency

(library already in mix.exs), prefer Tidewave over web search:

mcp__tidewave__get_docs(module: "LibraryModule")

This returns docs matching your exact mix.lock version — faster,

more accurate, zero web tokens. Only fall through to web search if

Tidewave is unavailable or the topic needs community discussion

(gotchas, real-world patterns, comparisons).

1. Query Decomposition (CRITICAL — before any search)

NEVER pass raw $ARGUMENTS into WebSearch. Decompose first:

• If $ARGUMENTS < 30 words and focused → use as single query
• If $ARGUMENTS > 30 words or multi-topic → extract 2-4 queries

Each query: max 10 words, targets ONE specific aspect.

Example:

Input: "detect files, export to md, feed database with embeddings,
        use ReqLLM for OpenAI API..."
Queries:
  1. "Elixir PDF text extraction library hex"
  2. "Ecto pgvector embeddings setup"
  3. "ReqLLM OpenAI embeddings Elixir"

2. Parallel Web Search

Search ALL decomposed queries in a SINGLE response (parallel):

WebSearch(query: "{query1} site:elixirforum.com OR site:hexdocs.pm OR site:github.com")
WebSearch(query: "{query2} site:hexdocs.pm OR site:elixirforum.com")

Deduplicate URLs across results. Discard clearly irrelevant hits.

3. Spawn Parallel Research Workers

Group URLs by topic cluster. Spawn **1-3 web-researcher agents

in parallel** (one per topic cluster):

Agent(subagent_type: "phx:web-researcher", prompt: """
Research focus: {specific aspect from decomposed query}
Fetch these URLs:
- {url1}
- {url2}
- {url3}
Extract: code examples, patterns, gotchas, version compatibility.
Return 500-800 word summary.
""", run_in_background: true)

Rules:

• 1 topic cluster = 1 agent (don't mix unrelated URLs)
• Max 5 URLs per agent (diminishing returns beyond that)
• If only 1-3 URLs total, use single foreground agent
• Pass URLs explicitly — agents should NOT re-search
• Agents are haiku — cheap, fast, focused on extraction

4. Write Output (File-First — NEVER Dump Inline)

After ALL agents complete, synthesize summaries into ONE file.

Target: ~5KB for topic research, ~3KB for library evaluations.

Create .claude/research/{topic-slug}.md:

# Research: {topic}

## Summary
{2-3 sentence answer combining all worker findings}

## Sources

### {Category}
- [{title}]({url}) - {key insight}

### Code Examples

From {source}: {what this demonstrates}

{code}

## Recommendations

1. {recommendation with evidence}
2. {recommendation with evidence}

## Watch Out For

- {gotcha from forum/issues}
- {version compatibility note}

5. After Research — STOP

STOP and present the research summary. Do NOT auto-transition.

Use AskUserQuestion to let the user choose next action:

• "Plan a feature based on this research" → /phx:plan
• "Investigate a specific finding" → /phx:investigate
• "Research more on a subtopic" → continue research
• "Done" → end

NEVER auto-invoke /phx:plan or any other skill after research.