a builder's codex
codex · operators · Aakash Gupta · ins_aakash-gupta-paste-twice-test

If you would paste the same instructions twice, that workflow belongs in a skill, not session memory

By Aakash Gupta · Product growth advisor and AI tools practitioner · 2026-05-12 · essay · Claude Skills: 7 Laws

Tier B · TL;DR
If you would paste the same instructions twice, that workflow belongs in a skill, not session memory

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

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

Open the interactive view → View original source → Markdown source →