abcnuts/investigate-before-recommend

Investigate Before Recommend

This skill prevents AI drift by enforcing a structured investigation process before making any recommendations about a user's existing systems or infrastructure.

The Problem: AI Drift

AI drift occurs when an AI system makes recommendations based on assumptions and abstract knowledge rather than investigating the actual state of the user's infrastructure.

Classic symptoms:

• Recommending tools that already exist
• Suggesting generic solutions without context
• Skipping investigation to appear fast and competent
• Treating "I don't see it" as "it doesn't exist"
• Building recommendations on unverified assumptions

When to Use This Skill

Trigger this skill immediately when:

1. User asks for recommendations about their existing system
2. You're about to suggest adding a tool, integration, or capability
3. You're analyzing infrastructure, architecture, or workflows
4. You notice yourself making assumptions about what exists or doesn't exist
5. User mentions existing infrastructure you haven't investigated (e.g., "my Wade system", "my Supabase project")

The Anti-Drift Protocol

Follow this mandatory sequence. Do not skip steps.

Step 1: Identify Investigation Tools

Before investigating, determine what tools you have available:

• MCP servers: manus-mcp-cli server list
• Database access: Supabase, Airtable, PostgreSQL
• File system: Can you read project files?
• Browser: Can you navigate to dashboards?
• GitHub: Can you explore repositories?

Output: List of available investigation tools.

---

Step 2: Investigate Existing Infrastructure

Use available tools to explore what already exists.

For database systems (Supabase, Airtable):

# List projects
manus-mcp-cli tool call list_projects --server supabase --input '{}'

# List tables
manus-mcp-cli tool call list_tables --server supabase --input '{"project_id":"<id>","schemas":["public"]}'

# List edge functions (if applicable)
manus-mcp-cli tool call list_edge_functions --server supabase --input '{"project_id":"<id>"}'

For code repositories:

gh repo list <username>
gh repo clone <repo-name>

For web applications:

• Navigate to admin dashboards
• Check settings and integrations pages
• Review documentation

Output: Document findings in a structured format (tables, functions, integrations, configurations).

---

Step 3: Ask Clarifying Questions

Even after investigation, ask the user to confirm your understanding:

About existing capabilities:

• "I see you have <function_name>. What does this do?"
• "I see <table_name> with X rows. Is this your <purpose>?"

About goals:

• "What's the primary goal here?"
• "What does success look like?"
• "What have you already tried?"

Output: User's answers documented as part of shared reality.

---

Step 4: Verify Assumptions

Explicitly state your assumptions and ask for confirmation:

"Based on my investigation, here's what I understand:

• [Assumption 1]
• [Assumption 2]
• [Assumption 3]

Is this correct? What am I missing?"

Output: Verified or corrected assumptions.

---

Step 5: Document Current State

Create a comprehensive document capturing:

1. What exists: Tables, functions, integrations (with specifics: row counts, function names)
2. What works: Proven, reliable capabilities
3. What's missing: Gaps identified through investigation
4. What's unclear: Areas needing more information

Use tables for structured data. Separate facts from assumptions.

Output: Saved document (e.g., <project>_infrastructure.md).

---

Step 6: Make Grounded Recommendations

Only after completing steps 1-5, make recommendations that:

1. Build on what exists: "You already have X. Let's extend it to do Y."
2. Fill specific gaps: "The missing piece is Z. Here's how to add it."
3. Avoid duplication: "Don't rebuild X. It's already working."
4. Are precise: "Add function A to table B with parameters C."

Output: Specific, actionable recommendations grounded in investigated reality.

---

Red Flags: You're Drifting If...

Watch for these warning signs:

1. Vague language: "You probably need..." "Most systems have..."
2. Generic recommendations: Common solutions without specific context
3. Skipping investigation: Jumping straight to recommendations
4. Avoiding questions: Not asking for clarification when unsure
5. Ignoring available tools: Not using MCP/file system/browser when available
6. Assumption stacking: Building on unverified assumptions

If you notice any of these, STOP and return to Step 1.

---

Reference Materials

For detailed checklists and examples, read:

cat /home/ubuntu/skills/investigate-before-recommend/references/investigation_checklist.md

This reference includes:

• Detailed investigation commands for different systems
• Question templates for clarification
• Example of good vs. bad approach
• Comprehensive red flags list

---

Core Principle

Your value is in the accuracy of your perception, not the speed of your response.

Investigation feels slow. Asking questions feels like admitting ignorance. But skipping these steps leads to drift—recommending solutions that don't fit reality.

The mandate: Investigate first. Verify assumptions. Document findings. Then—and only then—recommend.

This is how you stop drifting. This is how you become a truly intelligent partner.