Use Claude to Evaluate Whether a Project Should Become a Skill
The Insight
When Claude Skills was announced, most people’s question was “what should I build as a skill?” Donald Kihenja asked a better question: he took an existing project and asked Claude itself to evaluate whether it should be migrated to a Skill.
This is a meta-prompting move with broad applicability: use the AI system to audit your existing AI assets against the capabilities of a new paradigm, and let it tell you where migration makes sense.
Donald shared the resulting conversation publicly: https://claude.ai/share/e83a93cf-979d-48f0-bd5a-101f349438fa
The Approach
Rather than theorizing about where Skills add value, Donald:
- Described an existing Claude Project — its purpose, structure, and how he currently uses it
- Asked Claude to evaluate whether it would benefit from migration to a Skill
- Used the output as a decision brief
The conversation produced a useful framework for distinguishing Custom Projects (good for: persistent context, ongoing conversation, document storage) from Skills (good for: reusable procedures, orchestrated workflows, multi-role pipelines, persistent memory across sessions).
When a Project Should Become a Skill
Based on the evaluation framework that emerged, a Project is a candidate for Skill migration when:
- The core value is a repeatable procedure rather than a document collection
- You want the behavior to run without manual orchestration on each use
- The workflow involves multiple defined roles (researcher, writer, editor)
- You want memory to persist and compound across sessions
- The asset has enough polish to potentially share or transfer to others
A Project should stay a Project when:
- The primary value is the knowledge files uploaded into it
- The interaction is primarily conversational and exploratory
- The workflow varies significantly each time
Why This Matters
This is an example of using AI to navigate AI architecture decisions — a genuinely useful application of the tool on itself. As the ecosystem grows more complex (Projects, Skills, Agents, MCP servers, custom GPTs, Gems), the decision of where to house a given capability becomes a recurring challenge. The habit of asking Claude to evaluate its own capabilities against your requirements short-circuits a lot of manual research.
Donald’s comment in chat captured it well: “I found it very helpful so I’m sharing the chat in case anyone finds it helpful.”