Skip to content

Instantly share code, notes, and snippets.

@PriNova
Last active June 26, 2025 07:06
Show Gist options
  • Select an option

  • Save PriNova/7fa11eaf552aee1f91ae465805bf83cf to your computer and use it in GitHub Desktop.

Select an option

Save PriNova/7fa11eaf552aee1f91ae465805bf83cf to your computer and use it in GitHub Desktop.
Amp feature implementation prompt.

How to Implement Features in Unfamiliar Codebases Using Sub-Agents

Overview

This workflow maximizes efficiency and reliability when implementing new features in unknown codebases by using strategic sub-agent delegation and systematic discovery phases.

Complete Workflow Prompt

I need to implement [FEATURE DESCRIPTION] in this unfamiliar codebase. Follow this optimized workflow with strategic sub-agent delegation:

**PHASE 0 - ENHANCED CONTEXT PREPARATION (Main Agent):**
1. Check existing knowledge graph for this codebase
2. **IF SUBSTANTIAL KNOWLEDGE EXISTS:**
   - Review stored architecture, patterns, and conventions
   - Identify knowledge gaps specific to the new feature
   - Create targeted discovery plan focusing only on unknown areas
   - Skip to abbreviated discovery (Phase 1B)
3. **IF MINIMAL KNOWLEDGE EXISTS:**
   - Proceed with full discovery workflow (Phase 1)
4. Create initial todo breakdown using todo_write
5. Quick read of key project files (README, package json, etc..) for immediate context if needed

**PHASE 1A - FOUNDATIONAL DISCOVERY (Single Sub-Agent) - For New Codebases:**
6. Sub-Agent: Comprehensive codebase overview including project structure, tech stack, build system, main entry points, core configuration files, and overall architectural patterns
7. Main Agent: Process foundational findings and create targeted Context DNA for Phase 1B

**PHASE 1B - TARGETED ANALYSIS (3 Parallel Sub-Agents with Context DNA):**
8. Sub-Agent A: Receive foundational Context DNA, focus on similar features and integration points for the target feature
9. Sub-Agent B: Receive foundational Context DNA, focus on testing frameworks, error handling patterns, and validation approaches
10. Sub-Agent C: Receive foundational Context DNA, focus on deployment patterns, environment configuration, and development workflows
11. Main Agent: Validate targeted findings, resolve conflicts, and create Context DNA for Phase 2
12. **CHECKPOINT 1**: If checkpoints enabled, summarize codebase findings and confirm approach

**PHASE 1A-ALT - KNOWLEDGE-BASED DISCOVERY (When Prior Knowledge Exists):**
6. Sub-Agent: Focus only on areas/patterns not previously explored related to new feature, leveraging existing knowledge graph
7. Main Agent: Validate discoveries, merge with existing knowledge, and create Context DNA for Phase 2 (skip Phase 1B)
8. **CHECKPOINT 1**: If checkpoints enabled, summarize new findings and integration approach

**PHASE 2 - SYNTHESIS & PLANNING (Main Agent + 1 Sub-Agent):**
13. Main Agent: Synthesize discoveries, update todos with specific implementation tasks, merge with existing knowledge
14. Sub-Agent: Receive Context DNA from Phase 1, create architecture diagram showing integration approach
15. Main Agent: Plan dependency-ordered implementation, create Context DNA for Phase 3 with specific sub-agent assignments
16. **CHECKPOINT 2**: If checkpoints enabled, present implementation plan and architecture for approval

**PHASE 3 - PARALLEL IMPLEMENTATION (Multiple Sub-Agents):**
17. Deploy independent sub-agents in parallel (each receives targeted Context DNA and strict scope isolation):
    - Frontend Sub-Agent: Receive frontend-specific Context DNA, handle only frontend components and related files
    - Backend Sub-Agent: Receive backend-specific Context DNA, handle only backend/API logic and related files
    - Data Layer Sub-Agent: Receive data-specific Context DNA, handle only data/database changes and related files
    - Config Sub-Agent: Receive config-specific Context DNA, handle only configuration updates and related files
    **Each sub-agent must ignore errors outside their assigned scope**
18. Deploy dependent sub-agents sequentially:
    - Integration Sub-Agent: Receive consolidated Context DNA from parallel agents, connect components after all parallel agents complete
    - End-to-end functionality implementation
19. Sub-Agent: Receive testing Context DNA, implement comprehensive tests
20. Main Agent: Track progress, UPDATE existing knowledge graph with new observations, update todos as completed
21. **CHECKPOINT 3**: If checkpoints enabled, summarize implementation and confirm testing approach

**PHASE 4 - VERIFICATION & RECOVERY (Sequential Sub-Agents):**
22. Sub-Agent: Run full test suite, linting, build verification
23. Sub-Agent: Validate feature works end-to-end
24. Main Agent: Handle any failures with rollback strategy, final knowledge graph updates and enhancements

**DELEGATION RULES:**
- **Main Agent**: Quick reads, coordination, decision-making, conflict resolution, todo management, Context DNA creation and curation
- **Sub-Agents**: Detailed analysis, file modifications, comprehensive testing, diagram creation, Context DNA generation
- **Dependency Handling**: Clearly specify which sub-agents must complete before others start
- **Error Recovery**: Main agent validates sub-agent results at each phase before proceeding

**SUB-AGENT SCOPE ISOLATION:**
- **CRITICAL**: Implementation sub-agents must NEVER run global checks (pnpm check, tsc, npm test, etc.)
- Each sub-agent only works on their assigned files/components and validates only their own changes
- Sub-agents must IGNORE TypeScript/lint errors from other parts of the codebase that are outside their scope
- Sub-agents should ONLY fix errors in files they are explicitly assigned to modify
- If a sub-agent encounters errors in files they didn't create/modify, they must IGNORE them - another sub-agent will handle those
- Use file-specific validation only: check syntax in individual files, not project-wide compilation

**IMPLEMENTATION PHASE BOUNDARIES:**
- **Frontend Sub-Agent**: Only modify/validate frontend files (components, styles, frontend utils)
- **Backend Sub-Agent**: Only modify/validate backend files (API routes, server logic, backend utils)  
- **Data Layer Sub-Agent**: Only modify/validate database/data files (models, migrations, queries)
- **Config Sub-Agent**: Only modify/validate configuration files (package.json, tsconfig, env files)
- **Integration Sub-Agent**: Only handles connecting the above components, runs after all others complete

**VALIDATION RESTRICTIONS DURING PHASE 3:**
- Sub-agents must NOT run: `pnpm check`, `tsc`, `npm test`, `npm run build`, or any global validation
- Sub-agents CAN run: syntax validation on their specific files only
- Sub-agents must focus on completing their assigned scope, not fixing the entire project
- Global validation happens ONLY in Phase 4 by dedicated verification sub-agents

**CONTEXT DNA COMMUNICATION PROTOCOL:**
- **Main Agent Responsibilities**: Create targeted Context DNA for each sub-agent containing only relevant information for their specific task
- **Sub-Agent Task Instructions**: Every sub-agent task must include the Context DNA template and instructions
- **Sub-Agent Requirements**: When receiving Context DNA, acknowledge and use it to inform your approach without redundant discovery
- **Context DNA Format**: Human-readable summary including architecture insights, code patterns, file locations, and task-specific guidance
- **Context DNA Generation**: Sub-agents must end their final summary with Context DNA for the next phase when applicable

**CRITICAL: Include this in every sub-agent task:**

**CONTEXT DNA INSTRUCTIONS:**
End your response with Context DNA for the next phase using this format:

**CONTEXT DNA FOR NEXT PHASE:**

**Key Findings:**
[Most important discoveries from your investigation]

**Architecture/Patterns Identified:**
[Code patterns, tech stack details, conventions found]

**Important Files/Locations:**
[Specific files with line numbers, key integration points]

**Guidance for Next Phase:**
[Specific recommendations for what the next agent should focus on]

**HUMAN FEEDBACK OPTIONS:**
- **Default**: Fully autonomous execution with final summary
- **Add "WITH CHECKPOINTS"**: Pause after each major phase for review and approval
- **Add "REVIEW PLANNING"**: Pause only after Phase 2 for architectural approval before implementation
- **Add "REVIEW IMPLEMENTATION"**: Pause only after Phase 3 to review changes before testing

**CHECKPOINT BEHAVIOR:**
When pausing, provide:
1. Clear summary of work completed in current phase
2. Key findings, decisions, or architectural choices made
3. Next steps planned with specific sub-agent assignments
4. Specific questions or concerns requiring input
5. Option to continue autonomously or modify approach

**FEATURE TO IMPLEMENT:**
[FEATURE DESCRIPTION] [OPTIONAL: WITH CHECKPOINTS | REVIEW PLANNING | REVIEW IMPLEMENTATION]

Begin with Phase 0 and proceed systematically through each phase with proper validation checkpoints.

CONTEXT DNA COMMUNICATION SYSTEM:

  • Purpose: Efficient context transfer between phases without token bloat
  • Format: Human-readable summaries of key findings, patterns, and guidance
  • Flow: Main Agent curates Context DNA for each sub-agent based on their specific needs
  • Content: Architecture insights, code patterns, file locations, implementation guidance
  • Benefit: Sub-agents start with relevant context without redundant discovery

KNOWLEDGE ENHANCEMENT STRATEGY:

  • Throughout implementation, UPDATE existing entities rather than create duplicates
  • Add new observations to existing architectural components
  • Create new relations between known and newly discovered components
  • Enhance existing diagrams rather than creating entirely new ones
  • Build upon previous implementation patterns and decisions

Usage Instructions

  1. Replace [FEATURE DESCRIPTION] with detailed description of the feature you want to implement
  2. Choose feedback level by adding one of the optional modifiers:
    • No modifier = fully autonomous
    • WITH CHECKPOINTS = pause after each phase
    • REVIEW PLANNING = pause only after planning phase
    • REVIEW IMPLEMENTATION = pause only after implementation phase
  3. For subsequent features in the same codebase, this workflow will automatically leverage existing knowledge
  4. Paste the complete prompt to initiate the workflow

Key Benefits

  • Systematic Discovery: Prevents misunderstanding unfamiliar codebases
  • Knowledge Reuse: Leverages previous implementations to accelerate subsequent features
  • Parallel Execution: Maximizes efficiency through proper sub-agent delegation
  • Scope Isolation: Prevents sub-agents from interfering with each other's work
  • Context Preservation: Keeps main agent context clean while delegating detailed work
  • Efficient Context Transfer: Context DNA system eliminates redundant discovery and token bloat
  • Flexible Feedback: Allows human input at critical decision points
  • Risk Mitigation: Validates findings before proceeding to implementation
  • Cumulative Learning: Each feature implementation builds upon previous knowledge

Context DNA Examples

Phase 1A → Phase 1B Context DNA:

**CONTEXT DNA FROM FOUNDATIONAL DISCOVERY:**

**Codebase Architecture:**
React frontend with Next.js 13 (app router), Express API backend, PostgreSQL with Prisma ORM. Authentication uses NextAuth.js with JWT strategy. Monorepo structure with /web and /api directories.

**Tech Stack & Build System:**
- TypeScript throughout, built with Next.js build system
- Package manager: pnpm with workspaces
- Testing: Jest + React Testing Library
- Linting: ESLint + Prettier with custom configs in .eslintrc.js

**Core Entry Points:**
- Frontend: /web/src/app/layout.tsx (root layout)
- API: /api/src/server.ts (Express app entry)
- Database: /api/src/db/client.ts (Prisma client)

**Main Configuration Files:**
- /package.json (workspace config)
- /web/next.config.js (Next.js config)
- /api/tsconfig.json (TypeScript config)
- /prisma/schema.prisma (database schema)

**Overall Patterns:**
Follows domain-driven design with clear separation between web and API. Uses middleware pattern for cross-cutting concerns. Database-first approach with Prisma migrations.

Phase 1B → Phase 2 Context DNA:

**CONTEXT DNA FROM TARGETED ANALYSIS:**

**Similar Features Found:**
User management system in /web/src/app/users demonstrates standard CRUD pattern. Auth flow in /web/src/app/auth shows form handling with validation.

**Testing Patterns:**
Jest tests in __tests__ directories alongside source files. Integration tests use @testing-library/react. API tests use supertest with test database.

**Error Handling & Validation:**
- Frontend: react-hook-form + zod schemas for validation
- API: express-validator middleware + global error handler
- Error boundaries implemented at page level in /web/src/components/ErrorBoundary.tsx

**Development Workflow:**
- Local development uses docker-compose for database
- Hot reload enabled for both frontend and API
- Pre-commit hooks run linting and type checking
- CI/CD pipeline defined in .github/workflows/

Common Pitfalls Avoided

  • Sub-agents running global checks and seeing other agents' errors
  • Parallel agents working on the same files simultaneously (eliminated by sequential→parallel discovery)
  • Overlapping discovery work between parallel sub-agents (solved by foundational→targeted approach)
  • Insufficient discovery leading to architectural mismatches
  • Context window exhaustion from detailed analysis in main thread
  • Proceeding with implementation before understanding existing patterns
  • Rediscovering known architecture in subsequent feature implementations
  • Creating duplicate knowledge instead of enhancing existing understanding
  • Context loss between sub-agent phases causing redundant work
  • Token bloat from passing full discovery results to implementation agents
@PriNova
Copy link
Author

PriNova commented Jun 16, 2025

Add Context DNA Communication Protocol for shared context between main agent and sub agents.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment