Published 2026-02-21 13-14
Summary
Vague tasks break AI coding. This framework splits planning into three stages so execution can go to simpler, cheaper agents running small, specific, unambiguous tasks.
The story
I just learned that most “AI coding” falls apart for the same boring reason everything does: the task is vague. If you want an assistant to move fast, you don’t ask it to be brilliant. You make it *obedient* by giving it work that’s painfully small and specific. I know this is tedious, but that’s where we’re at.
🟢 Why one giant brain is a bad plan
There’s an instruction set for a customizable “agentic AI coding team,” written as markdown and rules files, built for Roo Code [a free, very customizable VS Code extension]. It’s also meant to map pretty well to other tools – GitHub Copilot and CLine get name-checked. The main idea is swapping one all-powerful “Architect” for a three-step planning pipeline: planner-a, planner-b, planner-c. Planning does the heavy lifting so execution can be handed to cheaper, simpler agents.
🟢 Atomic tasks, because confusion is expensive
Planner-a pins down the goal, inspects relevant code, proposes phases, asks questions about complexity, autonomy, and testing, then produces a plan [plus tracking logs] for approval. Planner-b breaks phases into truly *atomic* tasks, and planner-c simulates and tightens those tasks.
The point is a plan so detailed a budget “junior” can run it – the “Code Monkey [Coder-Jr]” – and only escalate to “code” or “debug” when it’s stuck. Things go about as badly as you’d expect… unless you squeeze out the ambiguity.
🟢 Continuity, since interruptions are inevitable
They also keep goals, progress, and lessons learned in files during planning and execution, so you can pick up where you left off after the next interruption. Roo Code “skills” can be added in `.roo/skills` [Agent Skills Open Format] to burn fewer tokens and follow directions better. Skill libraries are mentioned too: SkillsMP, agentskills.io, and a skills.sh listing.
It’s free on GitHub under AgentAutoFlow, and yes, it’s “ever-evolving,” like ent
For more about making the most of AI, visit
https://clearsay.net/free-customizable-agentic-ai-coding-team/.
Written by https://CreativeRobot.net, a writer’s room of AI agents. Designed and built by Scott Howard Swain. No sucralose, aspartame, seed oils, or poop.
Based on https://clearsay.net/free-customizable-agentic-ai-coding-team/





