wintersyntax/ai-assisted-development-workflow

> **Active work in progress.** This repository documents an experimental software-development workflow that I am actively using and refining while building real projects. It is not presented as a finished framework or a universal best practice.
A portable, guardrailed workflow for AI-assisted software development, designed around **durable project memory, model flexibility, bounded implementation and cost-aware execution**.
Why this exists
AI-assisted development is powerful, but many workflows become dependent on one chat session, one model subscription, or one tool's internal memory. That creates practical limits around context, cost, availability and reproducibility.
This project explores a more portable approach:
- keep project memory, decisions and handoff state in version-controlled documents;
- use replaceable model/provider boundaries instead of tying the workflow to one subscription;
- spend stronger reasoning models where they add the most value;
- give implementation agents narrow, explicit contracts rather than the whole project;
- use deterministic tests, guards and CI as completion authority;
- preserve a durable ledger so work can continue across sessions, tools and models.
The goal is not maximum autonomy. It is **more control over context, documents, cost, time and change** while still benefiting from strong AI models where they are most useful.
High-level architecture
Development path
The main path is intentionally linear: **solve → prepare → apply → verify → review → integrate**. The only implementation branch is whether the Host can apply Architect-written exact edits directly or must invoke the thin Cline fallback.
Supporting layers
These layers stay deliberately separate: **the repository owns durable state and policy; OpenRouter only supplies replaceable model access**. Neither chat history nor a model provider becomes the source of truth.
OpenRouter is a **model/provider boundary**, not another agent. TypingMind is the planning/review workspace, the controlled Host owns execution and verification, and Cline is only the bounded fallback when a solved change cannot be applied deterministically.
Interactive demo
The repository includes a **fully synthetic browser demo** of the guarded Host workflow. It contains no private source code, credentials, deployment values, runner identities, or production data.
*Prepare → Run → Verify → Review. Synthetic portfolio data only; the animation mirrors the operator flow without exposing the private project's code or environment.*
Run it locally with:
Then open:
The demo walks through four stages:
1. **Prepare** — SHA-bound Contract + Blueprint and pre-mutation checks;
2. **Run** — Host-owned activity, read-only Context/Changes evidence, and an exact-edit path with **0 executor provider turns**;
3. **Verify** — focused tests, guards, and fail-closed repair boundaries;
4. **Review** — read-only semantic review followed by exact-PR-head CI, human merge approval, and exact-main CI.
Use the stage tabs or left/right arrow keys to step through the flow. The demo keeps a synthetic portfolio presentation layer while following the same four operator phases as the real Host console: Prepare → Run → Verify → Review.
Core ideas
Repository-owned prompts and durable project memory