I've been using Claude Code for the last 4 months to build SaaS backends. Love it. Until I don't.
You know the pattern. Day 1: Claude writes beautiful auth logic. You're impressed. Day 3: Ask it to add Stripe webhooks. Day 5: Auth is broken. No idea what changed. Day 7: Context window full. Start new session. Day 8: "Wait, what database schema are we using again?"
Every. Single. Time.
I'd spend more time re-explaining my project than actually building it. The "brilliant colleague with amnesia" metaphor is painfully accurate.
The Context Loss Problem Nobody's Solving
Here's what I kept hitting:
Mid-session drift. Claude would start with async/await, then randomly switch to .then() chains 200 lines later. Why? Context degradation. The model "forgets" earlier patterns as the conversation grows.
Schema amnesia. I'd define a users table with specific columns in message 5. By message 40, Claude's suggesting queries for columns that don't exist.
Security regression. RLS policies carefully set up in Phase 1? Completely ignored when adding features in Phase 3.
The Groundhog Day effect. Close laptop Friday. Open Monday. Spend 30 minutes re-explaining the entire project before Claude can write a single line.
I tried everything the internet suggested:
- ✗ Longer prompts with full context (hit token limits, quality degraded anyway)
- ✗ Custom instructions (too vague, didn't persist across sessions)
- ✗ Separate chats for each feature (lost the big picture, broke dependencies)
- ✗ Manual "memory dumps" (exhausting, error-prone)
Nothing worked. The fundamental issue is that LLMs have working memory, not long-term memory. They're brilliant in the moment, terrible at maintaining state.
What Actually Fixed It: Multi-Agent Orchestration
I realized the problem isn't the AI. It's the workflow.
Human developers don't keep entire codebases in their heads either. They use documentation. Design docs. Database schemas. API specs. External references that persist.
So I built a system that orchestrates Claude through specialized agents, each with fresh context windows and specific jobs.
The Four Files That Maintain State
1. PROJECT.md - The vision document
- Problem being solved (plain English)
- Target users and workflows
- Core value proposition
- Success criteria
2. REQUIREMENTS.md - Traceable feature definitions
- Every requirement has a unique ID (AUTH-01, PAY-02, etc.)
- v1 scope (must have), v2 scope (future), out-of-scope (won't do)
- Acceptance criteria for each
3. ROADMAP.md - Phased execution plan
- Phase 0: Infrastructure
- Phase 1: Core feature
- Phase 2: Supporting features
- Phase 3: Polish
- Each requirement mapped to specific phases
4. STATE.md - The living memory
- Completed phases (locked from modification)
- Current phase (only modifiable code)
- Database schema (exact DDL)
- API routes built (paths, methods, business logic)
- Architectural decisions
These files are sized to avoid context degradation (under 10k tokens each) and serve as a single source of truth for both humans and AI.
The Multi-Agent System
Instead of one long Claude conversation, the system spawns specialized parallel agents:
Research agents (4 running in parallel before coding):
- Stack researcher → best technologies for your domain
- Features researcher → table stakes vs differentiators
- Architecture researcher → system design patterns
- Pitfalls researcher → common mistakes to avoid
Execution agents:
- Planner → creates verified task plans
- Executor → runs plans with atomic commits
- Verifier → tests and auto-debugs
- Mapper → analyzes existing codebase
Each agent gets fresh context. No degradation. No drift.
The Workflow Cycle
1. INITIALIZE
Describe vision → AI creates PROJECT.md, REQUIREMENTS.md, ROADMAP.md
2. DISCUSS (each phase)
Shape implementation preferences before committing
3. PLAN
Research domain patterns → create verified execution plan
4. EXECUTE
Run plans in parallel waves with fresh contexts → atomic git commits
5. VERIFY
User acceptance testing with automatic debugging
Repeat 2-5 for each phase
Critical rule: Completed phases are locked. The AI can only modify code in the current phase. This prevents the "adding payments breaks auth" problem entirely.
Boilerplate-Aware Intelligence
The AI knows what's already built in the boilerplate (auth, Stripe, Razorpay, Supabase, multi-tenancy, emails, admin panel). It only plans what's custom to your domain.
This means:
- Zero time wiring auth to database
- Zero time setting up payment webhooks
- Zero time building admin panels
- Pure focus on your unique business logic
The Results (Why I'm Sharing This)
Last 30 days, I built 3 production SaaS backends using this system:
Analytics Dashboard (13 hours total, across 4 sessions)
- Custom analytics schema (metrics, data_points, aggregations)
- Ingestion API with validation
- Time-series calculations (daily, weekly, monthly)
- CSV export with date filtering
- Now has 8 paying users making $96/month
Feedback Widget (11 hours, 3 sessions)
- Feedback schema with metadata
- Widget embedding API (iframe + script tag)
- Admin CRUD with filtering
- Email notifications on submission
- Webhook system for integrations
- 5 signups in first week
Content Calendar (9 hours, 2 sessions)
- Content schema with scheduling
- CRUD API with role-based access
- Publishing logic with timezone handling
- Calendar view backend
- En route to production
All production-ready. All built with AI orchestration. All using persistent state across weeks.
Commands That Run It
After building this system for myself, I packaged it:
/propelkit:new-project
This master command:
- Asks deep questions about your project
- Spawns research agents for your domain
- Creates PROJECT.md, REQUIREMENTS.md, ROADMAP.md
- Generates phased execution plan
- Hands you off to phase-by-phase building
Then for each phase:
/propelkit:discuss-phase 1 # Shape your preferences
/propelkit:plan-phase 1 # Research + create execution plan
/propelkit:execute-phase 1 # Build with parallel agents
/propelkit:verify-work # Test with auto-debugging
The system maintains STATE.md automatically. Close laptop. Come back days later. Resume exactly where you left off.
PropelKit - The Packaged System
After the third project, I productized it.
What you get:
Production Next.js boilerplate (saves 100+ hours):
- Auth (email, OAuth, sessions)
- Stripe + Razorpay payments
- Supabase (PostgreSQL with RLS)
- Multi-tenancy (organizations, teams, roles)
- Credits system (usage-based billing)
- Email templates (8 pre-built)
- Admin panel (user management, analytics)
- 26 AI PM commands
Stack: Next.js 16, TypeScript, Supabase, Stripe, Razorpay
One-time purchase. You own the code. Build unlimited products.
The AI PM uses the exact multi-agent orchestration system described above. Persistent state. Parallel research. Boilerplate-aware. Atomic commits.
Demo: propelkit.dev (watch the AI questioning, research, roadmap generation, and execution)
Why This Approach Works
Context engineering - Separate files under degradation thresholds, not one massive chat
Multi-agent orchestration - Fresh contexts per agent, no drift accumulation
Boilerplate awareness - AI knows what exists, only builds what's custom
Atomic commits - One feature per commit, precision rollback
Phase locking - Completed code stays completed, no random rewrites
Domain research - AI understands your industry before writing code
This isn't just for PropelKit. The principles work anywhere - you need persistent state files and fresh context windows per task.
What's your experience with AI code context loss? Have you found other systems that work?
Top comments (0)