Draft and Revise

Most drafts get one pass: the model writes, the user receives, the user lives with whatever shape the first attempt produced. The result is competent prose that does not get better between turn 1 and turn 10 because the model is not given the affordance to criticize itself.

This skill builds that affordance into the protocol. Draft and revise are two distinct cognitive moves; doing them in one breath blurs them both. The skill splits them: draft once with full attention on generation, then revise with full attention on the rubric. The two phases use different mental modes — generative open-ended writing in the first, narrow checklist-driven editing in the second — and separating them lets each one work at full strength.

When to Apply

• The user asks for a draft of any prose artifact (email, message, blog post, memo, README, design doc, PR description, summary, response).
• The user asks for a revision of a draft they already have.
• The user explicitly asks for iteration: "draft this, then improve it," "give me a v1 and a v2," "show me your edits."
• The output is going to be read by someone other than the model; quality is non-trivial.

When NOT to Apply

• The user asks for a one-shot reply to a simple question (small talk, factual lookup, single-line confirmation).
• The user asks for code; the code-planning and implementation skills are the right tools.
• The artifact is so short (one sentence, one line) that there is no meaningful revision dimension.