Decoupled AI Engineering Workflow with pdlc-skills

Description and use

The pdlc-skills framework decouples graph, prompt, and loop layers by using a shared state file on disk instead of direct calls. This enables parallel processing of independent features in separate git worktrees, with human oversight for final shipping decisions.

Problem

Traditional AI engineering pipelines often suffer from tight coupling between logic layers, making them fragile and difficult to parallelize. There is a need for a system where the graph (stage order), prompt (model interaction), and loop (iteration control) layers are decoupled to allow for robust, parallel feature development without direct inter-layer calls.

Outcome

The pdlc-skills framework successfully decouples the three engineering layers (Graph, Prompt, Loop) by using a shared state file on disk rather than direct function calls. This architecture enables parallel processing of independent features in separate git worktrees, with clear separation between automated iteration and human shipping decisions.

Workflow

• Lead agent or human splits overall design into N mutually independent features.
• Outer script starts an inner loop for each feature in its own git worktree.
• Graph layer checks preconditions (e.g., requirements exist) at stage boundaries.
• Prompt layer executes the specific command (e.g., write requirements, design, tests) using shared rules.
• Command writes results and metadata to a feature-specific state file on disk.
• Loop layer reads the state file to decide whether to continue iterating, stop, or block.
• Inner loops run in parallel until each reaches 'review complete' or 'blocked'.
• Human collects results from all N features and decides on shipping.

Evidence

The three layers never call each other, not once. They hand off through a single file on disk.

Source

The loop's stop decision reads the first box [command results] and nothing else. Which means the model's self-assessment is excluded from the stop path.

Source

Each feature's inner loop runs in its own git worktree, out of everyone's way, with artifacts going through separate PRs.

Source