Research Gate Skill

Purpose

Enforce knowledge priority by gating expensive research while allowing quick lookups. Context is king - project documentation takes precedence over external sources.

Permissive Mode (Default)

This skill operates in permissive mode:

• Quick searches allowed
• Library documentation allowed
• Only expensive deep research is gated

What Gets Gated vs Allowed

BLOCKED (Requires Explicit Request)

| Tool | Reason | |------|--------| | mcp__perplexity__deep_research | Expensive, slow, often unnecessary |

ALWAYS ALLOWED

| Tool | Reason | |------|--------| | mcp__context7__resolve-library-id | Needed for accurate library lookups | | mcp__context7__get-library-docs | Prevents API hallucination | | mcp__local-rag__query_documents | Project knowledge | | mcp__local-rag__ingest_file | Adding to project knowledge | | mcp__perplexity__search | Quick lookups, low cost | | mcp__perplexity__reason | Reasoning, moderate cost | | WebSearch | Quick web lookups | | WebFetch | Fetching user-provided URLs |

Knowledge Priority (Enforced)

When researching any topic, follow this priority:

1. CLAUDE.md (project conventions) ← MANDATORY
2. .claude/PLANNING.md (architecture) ← MANDATORY
3. examples/ (working code patterns) ← MANDATORY
4. local-rag (indexed project docs) ← ALWAYS QUERY
5. PRPs/ai_docs/ (library guides) ← CHECK
6. context7 (library APIs) ← ALLOWED
7. perplexity search/reason ← ALLOWED
8. perplexity deep_research ← GATED

When deep_research is Blocked

If mcp__perplexity__deep_research is called without explicit request:

Research gate: deep_research blocked

This operation is expensive and usually unnecessary.
Already checked:
- local-rag: [X results]
- context7: [available/not available]

To proceed, explicitly request deep research:
- "Do deep research on [topic]"
- "I need comprehensive research on [topic]"
- Use /maintenance:research command

When deep_research is Allowed

The gate opens when:

• User explicitly says "deep research" or "comprehensive research"
• User runs /maintenance:research command
• User says "research this thoroughly"
• No relevant results from local-rag AND context7

Pre-Research Checklist

Before ANY external research, verify:

• [ ] Checked CLAUDE.md for conventions?
• [ ] Checked PLANNING.md for architecture decisions?
• [ ] Queried local-rag for existing knowledge?
• [ ] Checked PRPs/ai_docs/ for library guides?
• [ ] Used context7 for library APIs?

Only proceed to perplexity if local sources insufficient.

Research Flow

User asks about [topic]
        ↓
1. Check local-rag
   Found? → Use it
        ↓
2. Check PRPs/ai_docs/
   Found? → Use it
        ↓
3. Check context7 (if library-related)
   Found? → Use it
        ↓
4. Use perplexity search/reason (quick)
   Sufficient? → Use it
        ↓
5. deep_research ONLY if:
   - Explicitly requested, OR
   - All above sources insufficient

Integration with Agents

research-expert Agent

• Bypasses gate (designed for research)
• Still follows knowledge priority
• Documents findings in PRPs/ai_docs/

implementation-guide Agent

• Uses local knowledge first
• Quick lookups allowed
• deep_research requires explicit request

pattern-enforcer Agent

• Uses project docs only
• No external research needed
• context7 for library verification

Output When Gated

research-gate: Checking local sources first...

Local results:
- local-rag: Found 3 relevant documents
- PRPs/ai_docs/: Found typer-patterns.md
- context7: Typer library docs available

Recommendation: Use local sources. Deep research not needed.

[If user still wants deep research, they must explicitly request it]

Configuration

Projects can adjust gate strictness in CLAUDE.md:

## Research Gate Settings

- deep_research: gated (default)
- perplexity_reason: allowed (default)
- perplexity_search: allowed (default)
- context7: always_allowed
- local_rag: always_first

What This Skill Does NOT Do

• Does not block context7 (prevents hallucination)
• Does not block local-rag (project knowledge)
• Does not block quick perplexity searches
• Does not require flags for normal lookups

Success Metrics

• 90%+ of queries answered from local sources
• deep_research used only when necessary
• No API hallucinations (context7 always available)
• Fast response times (local sources first)

---

Core Principle: Project documentation is the source of truth. External research supplements, never replaces, local knowledge.