From Prompt Files to Agent Skills: Rebuilding Content Automation
Three separate Copilot prompt files were merged into one portable agent skill, simplifying content publishing and making it work with any AI agent.
A developer had been using three separate VS Code Copilot prompt files to add videos, podcasts and blog posts to a website. The prompts worked, but they duplicated the same rules across files, lacked any verification or PR-creation step, and only worked inside Copilot's agent mode. Real-world testing exposed a long list of practical issues the prompts never covered, including cookie consent dialogs, relative date formats, and shell PATH/environment quirks.
Those lessons drove a migration to 'agent skills' — a portable instruction format. By consolidating shared logic (environment setup, git workflow, browser automation, PR creation) into one reference file and splitting content-type-specific details into separate files, roughly 70% duplicated logic was eliminated. The resulting structure also follows a progressive-disclosure pattern, loading only the reference files actually needed for a given task, which keeps the agent's context lean.
The broader takeaway for engineers is that fully autonomous workflows require documenting operational details—authentication, environment variables, UI edge cases—that a human previously handled implicitly. Moving from IDE-specific prompt formats to agent-agnostic skills also makes the same automation reusable across different AI coding tools, such as Goose or Claude Code, rather than being locked into one assistant.