Tool/Orchestration/Adapted Open Source

SkillOpt

Offline agent review cycle consolidating recurring session corrections.

What It Does

Parses historical agent session transcripts offline, detects repeated human corrections, and proposes surgical skill refinements behind an evaluation gate.

Origin and Adaptation Notes

Repeatedly correcting the same model behaviors wastes effort. Consolidating corrections into permanent skill modifications improves future performance.

Attribution: Adapted from Microsoft Corporation's SkillOpt project. Distributed under the MIT License.

When to Activate

•Running nightly or offline agent self improvement routines
•Consolidating repeated engineering instructions into persistent rules

When to Avoid

•Active live coding sessions

Skill Specification and Runtime Setup

Full Instruction Specification

# SkillOpt Specification

## Overview
Parses historical agent session transcripts offline, detects repeated human corrections, and proposes surgical skill refinements behind an evaluation gate.

## Rationale
Repeatedly correcting the same model behaviors wastes effort. Consolidating corrections into permanent skill modifications improves future performance.

## Core Principles
* Always verify environment constraints and dependencies before executing.
* Enforce negative constraints and reject speculative abstractions.
* Validate output against stated acceptance rubrics.

## Execution Workflow
1. Discovery: Map requirements and inspect active context.
2. Decomposition: Break down the task into discrete, verifiable phases.
3. Execution: Implement surgical changes preserving existing code architecture.
4. Verification: Test all criteria before declaring completion.