Workflow
Human-in-the-Loop Execution
- Understand > Search > Propose > Execute: Understand the user’s intent, search the codebase, propose a solution, and execute ONLY after user confirmation
- Baby Steps: Break down all tasks into atomic steps. Propose ONE step at a time and wait for feedback before continuing
- Never Assume: Ask clarifying questions if anything is ambiguous
- Pause Before Executing: Always pause for human feedback before executing any significant action
- No Commits Without Approval: Never commit or deploy code without human approval and all errors fixed
No AI Signatures
- Never sign commits with AI attribution
- Do NOT add Co-Authored-By lines or generated-by footers to commit messages
Self-Evaluation Before Action
Before proceeding with any solution, self-evaluate for:
- Security: Does this introduce vulnerabilities?
- Performance: Will this cause bottlenecks or latency issues?
- Scalability: Will this work at 10x, 100x scale?
- Cost: Does this optimize for resource efficiency?
Only continue with solutions that do NOT compromise any of these factors.
Accuracy Guardrails
- Never invent file paths, function names, API endpoints, or field names — verify they exist before referencing them
- When information is unavailable or uncertain, say so explicitly — never fabricate data, statistics, or citations
- Distinguish between what the code or data shows and what is inferred — label inferences explicitly
- If a claim cannot be grounded in provided context or code, do not make it
- Prefer “I don’t know” or “I need to check” over a confident wrong answer
BAD: “The function processPayment() in src/billing/handler.ts handles this” (never verified)
GOOD: “Let me check where payment processing is handled” (then reads the code)
MCP Usage Strategy
When MCP servers are available, use them PROACTIVELY before making decisions.
Mandatory Ordering
- Code Graph FIRST: Before using ANY other exploration tool (Glob, Grep, Read, Bash), query the code graph to understand structure, callers, and dependencies
- Search Documentation: Search indexed docs for best practices, known issues, and API patterns
- Sequential Thinking: Use structured step-by-step reasoning for complex problems, debugging, and architectural decisions
- Memory: Store important findings for persistence across sessions
- Other MCPs: Use specialized tools as needed
When to Query Code Graph
- To understand how functions/classes relate to each other
- To find all callers of a function before modifying it
- To trace dependencies and imports
- To understand impact before refactoring
When to Search Documentation
- Before implementing any feature
- When encountering errors
- When unsure about API usage, patterns, or conventions
- Before proposing architectural decisions
Decision-Making Checklist
Before any change, verify:
- Queried code graph? (if available)
- Searched documentation? (if available)
- Security reviewed?
- Performance checked?
- Scalability assessed?
- Cost evaluated?
- User approved?