Moonshot's Kimi excels at long-context reasoning. ChooChoo gives it a compiled KIMI.md so that long context is your actual standards, ADRs, and boundaries, not a generic summary.
Kimi works best when its context window starts with structured project facts. ChooChoo generates KIMI.md from your linting config, key files, and ADRs so every conversation starts from the right baseline.
KIMI.md with stack, standards, and architecture decisions
What to avoid: explicit anti-patterns from your real linting rules
Auto-updated when your stack or standards change
# KIMI.md
# Generated by ChooChoo. Do not edit by hand.
## Project: payments-api
Language: Python 3.12
Framework: FastAPI 0.110
Database: PostgreSQL 15 via SQLAlchemy 2.0
## Standards
- Type-annotate every public function (strict mypy)
- Pydantic for inputs/outputs at the API boundary
- Ruff format (line-length = 100), Ruff check passing
- Tests for every business-logic path (≥ 85% cov)
## Avoid
- Raw SQL strings; use SQLAlchemy ORM
- print() in production code; use logger
- Logging PII, payment tokens, or auth secrets
## Architecture Decisions
ADR-001 RS256 JWT for auth
ADR-007 /v1/, /v2/ URL versioning
# SKILLS.md
# Generated by ChooChoo. Do not edit by hand.
## auth
JWT token lifecycle: creation, validation, refresh.
Entry: src/auth/service.py
Tests: tests/auth/
## payments
Stripe integration: charges, subscriptions, webhooks.
Entry: src/payments/processor.py
Note: Never log STRIPE_KEY or card numbers
## database
SQLAlchemy sessions and query patterns.
Entry: src/db/session.py
Rule: Always use get_db() dependency injection
## notifications
Email (SendGrid) and SMS (Twilio) dispatch.
Entry: src/notifications/dispatch.py
# PLAN.md
# Generated by ChooChoo. Do not edit by hand.
## Goal
Refactor the authentication service to support
multi-factor authentication (MFA) via TOTP.
## Steps
1. [ ] Add TOTP secret column to users table
migration: src/db/migrations/0023_add_totp.py
2. [ ] Implement TOTP validation in auth service
src/auth/service.py → verify_totp()
3. [ ] Add /auth/mfa/setup and /auth/mfa/verify routes
src/api/auth.py
4. [ ] Write unit tests for new MFA paths
tests/auth/test_mfa.py (≥ 90% coverage)
5. [ ] Update AGENTS.md with new ADR-012
## Constraints
- Do not break existing JWT flow (ADR-001)
- TOTP secrets encrypted at rest (AES-256)
- Rate-limit /auth/mfa/verify to 5 attempts / 15 min
# ARCHITECTURE.md
# Generated by ChooChoo. Do not edit by hand.
## System Overview
payments-api is a stateless FastAPI service.
All state lives in PostgreSQL (primary data) and
Redis (sessions, rate limiting, event streams).
## Request Flow
Client → API Gateway → FastAPI → Service Layer
→ Repository Layer
→ PostgreSQL / Redis
## Key Modules
src/auth/ JWT issuance + TOTP (ADR-001, ADR-012)
src/payments/ Stripe integration (charges, webhooks)
src/db/ SQLAlchemy sessions + migration runner
src/api/ Route definitions (v1 + v2 namespaces)
## ADR Index
ADR-001 RS256 JWT for authentication
ADR-004 Redis Streams for webhook event bus
ADR-007 URL versioning (/v1/, /v2/)
ADR-012 TOTP-based MFA (in progress)
## Deployment
Docker → ECS Fargate (app)
RDS PostgreSQL 15 (primary)
ElastiCache Redis 7 (cache + streams)
# GOAL.md
# Generated by ChooChoo. Do not edit by hand.
## Current Sprint Goal
Implement MFA support without breaking existing
single-factor authentication flows.
## Success Criteria
- All existing auth tests continue to pass
- New /auth/mfa/* endpoints return correct responses
- TOTP validation rejects tokens older than 30 s
- Coverage for src/auth/ stays above 90%
## Out of Scope
- SMS-based MFA (deferred to ADR-013)
- Admin UI for MFA management (separate ticket)
- Recovery codes (phase 2)
## Definition of Done
- PR reviewed and approved
- CI green (lint + test + coverage)
- AGENTS.md updated with ADR-012 summary
- Deployed to staging and smoke-tested
flowchart TD
A([Client Request]) --> B[API Gateway]
B --> C{JWT Valid?}
C -- Yes --> D[FastAPI Router]
C -- No --> E([401 Unauthorized])
D --> F[Service Layer]
F --> G[Repository]
G --> H[(PostgreSQL)]
G --> I[(Redis)]
F --> J[Stripe API]
J --> K[Webhook Event]
K --> L[(Redis Streams)]
Why use ChooChoo with Kimi?
Long context is only useful when it's the right context. ChooChoo writes the right one.
Consistent Output
Generated code follows your linting rules from the start, reducing cleanup work before code review.
Fewer Iterations
The AI understands your standards upfront, producing acceptable code faster and with less back-and-forth.
Team Alignment
Everyone's AI assistant generates code that matches your team's style guide, automatically.
Zero Configuration
Get started instantly with sensible defaults that work out of the box for most projects.
Type-Safe Code
Strict linting rules catch potential bugs and enforce type safety before they reach review.
AI-Ready Rules
Purpose-built rules that AI models understand and follow consistently, session after session.
Ready to ground Kimi in your codebase?
Compile KIMI.md from your real standards and watch reasoning quality jump from the first prompt.