Content design for AI
Skills: instructions an AI agent follows the same way twice
- Agent instructions
- Prompt engineering
- Evaluation
- Content specs
- Markdown
- MCP
- Problem
- An agent does what its instructions say. A loose prompt gives a different answer on every run, and nobody can review a prompt that lives in one person’s chat history.
- My role
- Content designer and builder. I wrote each skill, tested it and published it.
- What I made
- Skills written as documents with a fixed structure: when to apply, inputs accepted, output contract, voice and style, and a verification gate the output must pass.
- A two-stage research pipeline. The first skill turns a brief into a structured research prompt. The second turns the finished research into a longform article, an infographic spec and social copy.
- Skill Memory Bank, which compiles a working conversation into a SKILL.md package that loads in Claude Code, Cursor, Copilot, Codex and Gemini CLI.
- Result
- Best in Show, Lane 2, at the Open Accelerator Hackathon (Red Hat, IBM and the Mass AI Hub), Boston, June 27, 2026.
- In Skill Memory Bank’s evaluation harness, retention on a fixed input went from 0 of 3 without the skill to 3 of 3 with it.
- The open-source skills are published under Apache 2.0.