AI Loop Skill — AI skill for Claude Code

AI-Loop

A **gated** multi-agent circuit for Claude Code. The model proposes; the workflow validates and authorizes.

Most spec-driven workflows ask the model to audit itself against a document of principles. Here the gate is not one the model gets to pass:

• Nobody grades their own work. Whoever implements does not verify, and whoever criticizes does not modify.
• The verifier re-measures with its own probes. The implementer's report is an input, never evidence. And it runs on different weights: on the same model it inherits the same blind spots.
• Four verdicts, treated asymmetrically. A FAIL is cheap: one finding with evidence is enough. A PASS is expensive: it requires mapping every acceptance criterion to concrete evidence. BLOCKED never means retry. And UNKNOWN is not PASS, ever.
• No implicit transitions. Compiles ≠ done. Tests green ≠ PASS. PASS ≠ accepted.
• DONE is not an agent's opinion: it is a state granted when every condition holds, human acceptance included.

Install

/plugin marketplace add cRolandoJr/ai-loop-skill
/plugin install ai-loop@ai-loop

Then, with the task as the argument:

/ai-loop fix the played-minutes calculation when the match goes to extra time

What comes with it

Besides the circuit, the document carries what it cost to measure it:

• A context budget. A session's cost is quadratic in its length, because every request re-reads the entire prefix. Measured over a real run: 1,070 M tokens. Replaying those deltas with the session cut per phase, the simulation gives 72% less — simulated, not re-run: the distinction is the same one the circuit demands of every other claim.
• Golden cases. The circuit is validated with the same mechanism it uses to validate tests: a seeded defect that must die. The checks are interpreted by a rule, never by an LLM judge — an LLM judge puts inside the instrument the very variance the instrument exists to detec