Autonomous Roadmap Execution System
The Execution Loop that Runs ForeverHow to run an AI development loop that stays on track. One task at a time, without drift, without fake progress.Autonomous development loops break down when state files drift from the repo, or when the AI doesn't know which single task to run next. These four prompts fix that: initialize your project structure, realign state when it drifts, run one task per loop using phase files as the guide, and recover when checkboxes no longer match git reality. The Big Problem Autonomous coding loops sound great until the third session: the AI marks a task done that isn't, skips context from the phase file, adds scope it wasn't asked for, and leaves your roadmap.md out of sync with the code.The result is a completion percentage that lies and a rogue loop that builds the wrong thing.The Big PromiseRun these four prompts as your project stack, and you get: a clean, schema-correct initialization of all state files; a read-only normalizer that realigns state with the actual codebase; a one-task-at-a-time execution loop that documents blocks honestly instead of faking progress; and a recovery prompt for when drift happens anyway.What You GetFour ready-to-paste markdown prompt files: Initializer — Bootstraps or resets prd.md, roadmap.md, changelog.md, system-state.json, ui-state.json, and a2ui-screen.json. Defines phases and small tasks with IDs, titles, descriptions, and checkboxes. Builds a declarative A2UI screen that surfaces project status, roadmap, progress, and changelog. Normalizer — Read-only alignment pass. Makes the roadmap match the codebase, fixes wrong completions, ensures tasks are small and ordered, updates all three state JSON files with accurate progress and current task. Explicitly does not build features. Execution Loop — One task per run. Finds the next incomplete task, loads the phase file, gathers context from PRD and version-scope, and implements only that task, verifies, then updates the changelog and all state files. Documents blocking with blocked: true instead of faking progress. Recovery — Audit code versus roadmap. Finds false completions and inconsistencies, repairs roadmap and state files, optionally refines next tasks — without expanding scope. How It Works Run Initializer once — get a clean project structure with valid JSON and a real roadmap. Run Normalizer whenever you suspect drift — before trusting the completion percent for reporting. Run Execution Loop as your repeating implementation prompt — one task, then verify, then update state. Run Recovery after merges, manual edits, or when tasks were marked done incorrectly. Who This Is For Teams running (or adopting) a single-task execution loop with explicit checkpoints and state files Anyone who needs recovery steps when roadmap checkboxes no longer match git reality Projects using the RPKF roadmap-tech pattern (docs/roadmap-tech/) What This Is Not A replacement for code review or tests — "verification" is whatever your process defines A planning kit — use the Roadmap Architect kit to build the phases and tasks these prompts execute Frequently Asked QuestionsWhat if my project is already underway — can I use the Initializer? Yes. Run the Initializer with your existing roadmap and check that completed tasks are accurately marked. The prompt tells you to reconcile with the actual state of the repo before trusting any checkbox.Do I need the Roadmap Architect kit? They work as a pair. The Architect creates and expands the phases; Execution runs them. You can use either independently if you have comparable inputs, but the Execution Loop specifically reads phase files that the Architect generates.What happens when the AI marks something as done that isn't? Run the Normalizer. It does a read-only pass that compares the codebase to the roadmap and corrects wrong completions. Run it before any reporting pass.How is this different from just giving an AI a task list? A task list is stateless. These prompts maintain structured JSON that records what phase is active, which task is current, the blocked flag, and the completion percent. The loop reads the state before it acts and writes the state after it finishes.What if my tasks are too big? The Normalizer checks for this. If tasks are oversized, it flags them for splitting before the loop runs. Small tasks are a requirement of the execution pattern, not a suggestion.What should I do after the Execution Loop marks a task as done? Open the updated roadmap.md and system-state.json and verify the changes look right. Then run the loop again for the next task.What format are the files? Markdown (.md). Paste into your AI coding tool and attach the state files and phase files each prompt lists.What do I do after I download? See the post-purchase section below.Is there a refund policy? Gumroad's standard 30-day refund policy applies.
Get it → danejw.gumroad.com