A battle-tested methodology and template set for building production software with AI agents — without losing quality, context, or control.
I used this approach to independently ship three production-grade systems in about a year, alongside a full-time job — including an enterprise observability platform that delivers capability commercial tools charge six-to-seven figures for. Same method each time, across three completely different domains. This repo is that method, generalized so you can use it.
It's deliberately tool-agnostic. I orchestrate Claude, OpenAI Codex, Gemini, and Grok — plus generative pipelines (Meshy, Scenario) and a persistent ops agent — assigning each to the work it does best. The templates work with any AI coding assistant that can read your project files.
AI coding tools are powerful but fail in predictable ways once a project outgrows a single session:
- Context amnesia — every session starts fresh and forgets architectural decisions made days ago.
- Shortcut bias — left unconstrained, the model takes the simplest path, not the correct one.
- Scale blindness — it doesn't inherently know that code built to handle thousands of production systems needs different patterns than a prototype.
- Forgotten wiring — it adds the feature but misses three of the eight files that needed to change.
- False confidence — "done!" when the thing was never actually run and verified.
The fix isn't a better prompt. It's an operating system for the work.
Treat the human as a conductor and the AI tools as a section of players. The conductor never plays every instrument — they decompose the work, assign each part to the right player, and hold the standard. Four pillars make it work:
A primary "Conductor" session decomposes a feature into independent sub-tasks and dispatches each to a sub-agent working in an isolated git worktree with a bounded file scope and a self-contained brief. Task state lives in a manifest (one file per task: status, owner, scope, dependencies). The Conductor reviews each diff, runs acceptance tests, and integrates — and nothing merges to main without explicit human approval. Parallel work, zero collisions, one set of hands on the wheel.
No single model is best at everything, so don't pretend otherwise. Route each task to the strongest tool and let them coordinate. (Real example: in a game project, Claude authors asset briefs, Codex generates, and Claude renders them through Meshy and Scenario — covering image/3D generation that a single model can't.) → methodology/multi-tool-orchestration.md
A .claude-docs/ knowledge layer survives across sessions and tools:
- MEMORY — current state, environment, gotchas (read first, every session)
- STANDARDS — numbered (
§) patterns the AI must follow and can cite - DEPENDENCIES — "if you change X, also change Y" impact chains
- SESSION_LOG / HANDOFFs — what changed, why, and what's next
- ADRs — binding decisions, never re-litigated
- LESSONS — real mistakes, documented once so they never recur
- No partial fixes — a fix covers every code path, or it isn't done.
- Pre-merge self-verification — prove claims by running the command, not re-asserting from memory.
- Automated gates — tests, lint, and security scanning on every change; human-gated merges.
- Cross-project intelligence — proven patterns get ported between codebases instead of reinvented.
→ Full write-up: methodology/the-conductor-method.md
The same method produced three very different production systems:
| Project | What it is | Why it's evidence |
|---|---|---|
| MonLite | Enterprise observability platform (SAP, 6 databases, hosts, web) | Delivers six-to-seven-figure commercial monitoring capability; scale-tested to handle 3,000+ systems; embedded air-gapped AI assistant |
| Pathfinder | AI-augmented SAP-to-cloud migration decision engine | Multi-LLM authoring + a cumulative-intelligence loop; tamper-evident audit |
| Disaster Scenario | Cross-platform multiplayer game (Godot) | Multi-agent asset pipeline; deterministic tested core; proves the method travels beyond enterprise software |
(Case studies are capability-level — architecture and outcomes, no proprietary code.)
The templates/ directory is a drop-in .claude-docs/ system.
Automated setup — give your AI assistant templates/AI_BOOTSTRAP.md and say "Set up .claude-docs for my project." It will explore your codebase and generate project-specific docs.
Manual setup:
cp -r templates/.claude-docs your-project/.claude-docs
cp templates/.claude-docs/CLAUDE.md your-project/CLAUDE.md # or merge into your existing oneThen customize MEMORY.md, STANDARDS.md, and DEPENDENCIES.md for your project, and point your tool's config (CLAUDE.md, .cursorrules, etc.) at them.
→ Details: templates/README.md
Built by Joshua Lans — a 12-year enterprise infrastructure engineer who turned AI-assisted development into a repeatable way to ship production software. github.com/josh-lans · linkedin.com/in/joshualans
License: MIT — use it, adapt it, ship something.