tomicz/fable-5-train-opus-skills-after-it-retires

This repository contains a single long prompt rather than application code. It is meant to be given to an AI coding assistant, written from the point of view of a senior engineer who is about to retire from a project, with the goal of building a complete library of documented skills before that happens. The prompt asks the AI to first investigate an existing codebase like an incoming lead would: reading the README, build system, test suite, CI setup, documentation, and git history, looking at what changed, what got reverted, and any stalled work. After that research, it is told to ask the human at most five clarifying questions, only about things the repository itself cannot answer, such as the hardest current problem or unwritten rules that are not written down anywhere. Once that groundwork is done, the prompt directs the AI to write out ten to sixteen separate skill files, one at a time, covering things like how changes get reviewed, a debugging playbook of common failure symptoms, a record of past failed attempts so they are not repeated, the reasoning behind key design decisions, how to set up the environment, how to run and operate the project, and how to validate that a change actually works. A more advanced set of skills covers the single hardest live problem, research methods used in the project, and open problems where the project could improve on the current state of the art. The prompt sets strict authoring rules: skills should be written for a less experienced engineer or a smaller AI model, use plain imperative instructions with copy-pasteable commands, define every technical term once, and clearly state when a given skill should not be used. Every command, path, or claim must be checked against the real repository rather than guessed, and it must not conflict with the project's existing rules. After all the skill files exist, the prompt asks for three separate reviews, checking facts, checking for contradictions, and checking how usable the writing is, followed by one pass that fixes the important problems found. In short, this is a reusable prompt template for turning a senior engineer's undocumented knowledge of a codebase into a structured set of skill files that a less experienced person or a cheaper AI model can follow later.