The PIV Loop: Build Reliable AI Software in Three Steps

Get AI income methods before they spread.

Free weekly intelligence newsletter.

The PIV loop — Plan, Implement, Validate — solves the most common failure mode in AI-assisted coding: implementing without a clear plan and then spending more time fixing than building. Paired with the Planotator browser extension for interactive plan review, the loop produces reliable AI software in three structured steps, with per-node model switching to keep costs low.

What Is the PIV Loop and Why Does It Work?

PIV breaks each coding task into three non-negotiable steps. Planning happens before any implementation. Validation happens before any merge. The loop does not advance until each step is complete.

The reason this produces better software than unstructured prompting: AI coding fails most often at the boundary between understanding and execution. An LLM that starts coding immediately often codes a solution to the wrong problem. The planning step forces the problem to be stated clearly before any code is written.

No code is written until the plan is approved. This single constraint changes everything.

Validation closes the other common gap: code that runs but does not satisfy the requirement. The validate step checks the output against the original acceptance criteria — not against what the implementation did, but against what it was supposed to do.

What Is Planotator and How Does It Change the Planning Step?

Planotator is a browser extension that renders the implementation plan as a web page for review. Instead of reading a plan in a terminal or chat interface, you see a formatted document with each planned step laid out clearly. You can add inline comments directly in the browser, and Planotator writes those comments back to a plan.md file.

The practical benefit: you can review and annotate a plan without typing in the terminal. For complex plans with many steps, this is significantly faster than reading plan output inline and typing corrections back as prompts.

Reviewing plans in a browser with inline comments takes half the time of reviewing them in a terminal.

How Does Pi Coding Agent Fit Into the PIV Loop?

Pi is a minimalist open-source coding agent designed for full system prompt control. It supports all major LLM providers, which makes it compatible with Archon's per-node model switching. In the PIV loop, Pi handles the implementation step — it receives the approved plan, follows it precisely, and reports back.

The key Pi feature: you control the system prompt completely. Unlike some agents that add their own instructions on top of yours, Pi runs exactly what you tell it to. For the implementation step in a governed codebase, this matters — the agent follows the governance rules without overriding them with its own defaults.

How Does Per-Node Model Switching Reduce Costs?

The three steps in PIV have different capability requirements. The clarify step — asking questions to understand the task — can run on a small, fast model. The plan step needs moderate reasoning. The implement step benefits from the strongest available model. Validation can run on a fast model with access to the test results.

In Archon, each node specifies its own model. The implement node uses Claude. The classify and validate nodes use Minimax M2.7 or another cheap model. The cost of a full PIV loop is a fraction of running Opus for all three steps.

Match the model to the cognitive requirement of each step, not to the whole loop.

Frequently Asked Questions

How is the PIV loop different from just using plan mode in Claude Code?

Claude Code's plan mode (Shift+Tab) is a single-session feature that asks Claude to plan before acting. The PIV loop is a structured workflow across multiple steps and potentially multiple tools. PIV uses Planotator for browser-based plan review, Archon for orchestration, and Pi for implementation. Plan mode is a habit; PIV is a system.

What is Planotator and where do you get it?

Planotator is an extension for Archon that renders plan.md files in a browser interface with inline commenting. It is available from the Archon repository. It requires Archon to be running — it is not a standalone tool. Install it after setting up Archon and configure it to open automatically when a plan is generated.

Does the PIV loop work for bug fixes, or just new features?

Both. For a bug fix, the plan step describes the root cause and the intended fix before any code changes. This catches the common case of fixing a symptom rather than the cause. The validate step confirms that the original bug is gone and no regressions were introduced. The loop applies to any task where correctness matters.

How long does a complete PIV loop take for a medium-complexity feature?

Plan step: 5 to 10 minutes including Planotator review and any clarification comments. Implementation: 15 to 45 minutes depending on complexity. Validation: 5 to 15 minutes including test runs. Total: 30 to 70 minutes for a medium-complexity feature. That is comparable to doing the same work manually, but the output is more reliable and the approach is documented in the plan file.

Watch the full PIV loop demo: The PIV Loop on YouTube