Claim
If you would paste the same instructions twice across sessions, that workflow belongs in a skill. The description field is the routing layer. A vague description means the skill does not trigger when it should.
Mechanism
Claude Skills are invoked by description match, not by explicit user call. The description is the signal the model uses to determine whether a skill applies to the current context. A vague or generic description leaves the routing decision to chance. The result: the model produces generic output from the wrong context rather than the specialized output the skill was built for. The paste-twice test is a scope diagnostic: if the same instructions would help across multiple sessions, the value is recurring, and recurring instructions belong in a skill with a precise description.
Conditions
Holds when: the workflow is recurrent across sessions and the instructions are stable enough to codify. Applies to any AI assistant that routes to capabilities via description matching.
Fails when: the workflow is exploratory or one-off. Promoting it to a skill adds overhead without routing benefit.
Evidence
Gupta's finding after 75 tests of Claude Skills. The primary failure mode across all 75 tests was description quality, not instruction quality.
The test: if you'd paste the same instructions twice, it belongs in a skill.
Skills with vague descriptions were not triggered at the right moment. Skills with precise, context-specific descriptions triggered reliably.
Signals
- You find yourself copying the same prompt preamble at the start of multiple sessions
- A skill exists but the model does not invoke it in contexts where it would clearly help
- Output quality varies across sessions for the same type of task
Counter-evidence
The test applies specifically to Claude's description-matching invocation architecture. Other AI systems with different routing mechanisms (explicit tool calls, user-initiated slash commands) do not have the description-as-routing-layer problem in the same form.
Cross-references
- [[ins_bottleneck-is-context-not-capability]] — Sherwin Wu: when the agent isn't doing what you want, fix the context, not the model
- [[ins_dotclaude-as-deployable-artifact]] — on treating .claude configurations as deployable artifacts with named owners
- [[ins_anthropic-skills-as-vertical-distribution-unit]] — Anthropic treating skills as the unit of vertical agent distribution