Draft Writing

Draft Writing

Use this skill when the user wants help with a draft, section, report, article, paper prose, grant text, response letter, or other writing where the wording may change but the intended claims must stay controlled.

Do not use this as an AI detector. This is a procedural workflow for preserving claims through drafting and revision.

Trigger Examples

• draft this section from notes
• rewrite this paragraph without changing meaning
• polish this draft but keep my claims intact
• compare two draft versions for claim drift
• identify unsupported claims before revising
• turn this outline into prose while tracking claims

For paper/book review requests, use the relevant review workflow unless the user is asking to rewrite or prepare draft text.

Core Workflow

1. Define the writing scope and audience.
2. Inspect the local context needed to write in the requested form: current draft, outline, notes, source material, prior posts, templates, house style, venue instructions, and supplied examples. If expected context is absent, say so and state the style/content assumption before drafting.
3. Extract atomic claims from the current draft, outline, notes, or source material.
4. Classify each claim as contribution, evidence, assumption, caveat, comparison, recommendation, definition, result, limitation, or transition.
5. Map each substantive claim to support: source, experiment, theorem, data, author note, prior section, or missing.
6. Freeze the intended claim ledger before substantial rewriting.
7. Identify the active writing-style settings before rewriting:
   • always load writing-style-settings.md
   • load math-manuscript-style.md for mathematical, TCS, graph-theoretic, formal-proof, Lean-synchronized, or LaTeX manuscript prose
   • record style_profile_ref, active_overlays, active_requirement_ids, session_local_additions, and style_applied in the draft ledger or revision map
8. Rewrite for structure, clarity, and style without adding unsupported claims.
9. Audit the revision delta:
   • added claim
   • removed claim
   • strengthened claim
   • weakened claim
   • changed caveat
   • unsupported claim introduced
10. Report remaining gaps before presenting the draft as ready.

Use the installed templates when available:

• draft-claim-ledger.md for claim extraction and support mapping
• draft-revision-map.md for before/after revision audits

Use the instruction doc claim-preserving-writing.md for detailed guidance when the task involves multiple sections, citation-sensitive prose, or repeated revision rounds.

Always apply the installed instruction doc writing-style-settings.md. For mathematical or LaTeX manuscripts, also apply math-manuscript-style.md through the compatibility router language-style-rules.md if older installed targets still refer to it. In particular, check that concepts are defined before use, notation is not defined inside statements, unnecessary local terminology is removed, result introductions explain each statement's role, and long proofs begin with a clear proof idea.

Output Rules

• Separate author-provided intent from model-inferred improvements.
• Label unsupported or newly introduced claims instead of smoothing them into polished prose.
• Preserve caveats unless the user explicitly asks to remove or revise them.
• Do not generate a blog post, article, report, or other format-matched draft before inspecting available prior examples/templates/style artifacts. If the repository or workspace has no such artifacts, say that explicitly.
• When making a rewrite, include a short claim-change note if the change is substantive.
• For finalizable draft work, record a style block or style record with style_profile_ref, policy_hash, active_overlays, active_requirement_ids, and style_applied: true. A bare assertion of style_applied is not enough when the workflow did not load the policy.
• If material evidence remains unchecked, say incomplete analysis before any final readiness claim.

Boundary

This workflow tracks what is being said and whether declared support is present. It does not independently prove claims true unless paired with verification, review, citation lookup, tests, experiments, or formal checks.