sohangujari/fable-to-opus-skills

Fable to Opus Skills is a small project that captures how Claude's Fable 5 model plans and checks its own work, then turns that process into a reusable Claude Skill so a cheaper model like Opus 4.8 can follow the same workflow without needing Fable's more expensive per token pricing every time. It is not a fine tuning project. Instead it uses Anthropic's built in Agent Skills feature, which lets Claude load instruction files from the filesystem on demand. The repository contains two prompts meant to be used together. The first is run once inside a brand new Fable 5 chat session and extracts Fable's general planning and verification habits into a portable skill file that works across any codebase. The second prompt is run inside Claude Code within a specific project, and it captures only what is unique to that project, such as its test commands, conventions, and known gotchas, so it complements the general skill rather than repeating it. Setting things up means saving the first prompt's output as a SKILL.md file in a global folder under the user's home directory so it applies to every project, and saving the second prompt's output inside the specific repository's own skills folder so it only applies there. A skill file needs to start with a small block of frontmatter giving it a name and description, followed by plain markdown instructions, and Claude then loads and applies whichever skills are relevant automatically without needing to be told to use them by name. The project includes a real example output from running the first prompt, shown unedited, so someone can see what a finished skill file looks like before generating their own. The author is upfront that this does not make Opus 4.8 as capable as Fable 5 at difficult long reasoning tasks. It captures Fable's process of breaking down problems and double checking its own work, which accounts for much of what made its answers better, but it is not a full replacement for the underlying model's raw ability. The author also notes that prompts touching on security or vulnerability topics may get flagged by Fable's own safety systems, which is expected rather than a bug. The project is released under the MIT license.