Rl writing craft

Writing Craft — Structure, Edit, Audit, Copyedit

A general writing-quality skill. Four functions, each independently callable, defaulting to a full sequence when invoked generically.

This skill does not set voice. Register, rhythm, and the things that make writing sound like a specific person belong to the author-voice skill layered on top. This skill shapes the architecture, tightens the lines, and removes machine-writing patterns. The author-voice skill always wins where they touch the same sentence.

---

The Four Functions

1. structure — Architectural editing. The bones. Where the piece opens, how it moves, what each section does, how it closes. Run when a draft's argument is sound but its shape is wrong, or before drafting to plan the build.

2. edit — Line-level editing. Readability, cutting, grounding, sentence rhythm. Anti-AI pattern removal is woven through this pass — tells are cheapest to fix while editing the line they live on. Run when a piece is structurally sound and needs tightening.

3. audit — Dedicated anti-AI sweep. Mechanical pattern detection run clean against finished copy. Catches what survived the edit pass, including tells the editing itself introduced.

4. copyedit — Grammar, punctuation, usage, consistency, and proofreading. The mechanical-correctness layer a copyeditor and proofreader provide. Runs last, against near-final copy. Corrects unambiguous errors only — it never imposes style rules the author's voice deliberately breaks (fragments, sentence-initial conjunctions, etc.).

Default: full sequence. When invoked generically, run all four in order — structure, edit, audit, copyedit. Name a single function to run it alone.

Each function's full rules live in its own file next to this one — see Function Reference below. Load the file for the function you're running.

---

Sequencing With an Author-Voice Skill

The full pipeline, when a voice layer is present. The voice layer is your content-generation skill (venture/brand lane) or your author-voice skill (personal lane) — two distinct optional roles; the sequence is the same for both:

1. Draft — content produced
2. Voice layer — sets register and rhythm (your content-generation skill or your author-voice skill)
3. structure — shapes the architecture
4. edit — tightens lines, removes tells, applies grounding rules
5. audit — final mechanical sweep
6. copyedit — grammar, punctuation, usage, consistency, proofreading (errors only)

The voice layer runs first because voice decisions shape what the structure should be. The audit runs last because it catches tells introduced by every prior pass. If no author-voice skill is present, check for VOICE-PROFILE.md at the repo root (produced by rl-voice-discovery) before concluding there's no voice reference at all — treat it the same as an author-voice skill would be treated in step 2. Only with neither present does this skill run with no imposed register.

(The step-by-step mechanics of the four-function run live in "Running the Full Sequence" near the bottom of this file.)

Carve-out: Intentional structural techniques are not AI tells. Section breaks (* * *), one-sentence paragraphs for percussion, sentence fragments for rhythm, em-dashes as breath marks, a cold open that drops the reader into the middle of something — these are craft, not machine patterns. Do not flag them. Flag only overuse (e.g. em-dashes above ~1 per 1,000 words in body prose), never the technique itself.

---

Logical Consistency — Grave Errors

Separate from structure and line quality: does the argument actually hold together? Treat every item below as a blocking error, not a style preference — smooth prose can still be logically broken, and that's a worse failure than awkward prose, because it survives every pass that only reads for tone.

• A label must match what it does. If a passage is called a "test," it has to function as one — a procedure the reader could run. If it's actually a description, an account, or a framework, name it that instead. (Example: "the cleanest test I know" introduced two conditions to weigh, not a procedure — fixed to "the clearest account I've found.")
• A recap must cover what it claims to recap. When a section establishes multiple failure modes, causes, or conditions, any later diagnostic or summary invoking that same framework must include all of them — especially the one the piece's own central example demonstrates. Silently dropping one is a logical gap, not a trim. (Example: Hogarth's framework named feedback that's late, missing, or drowned in too much of it — but the diagnostic recap later listed only the first two, omitting the exact failure mode the piece's opening story was about.)
• Pronouns need an antecedent before they're used, not after. "Arguing opposite sides of it" when "it" is named as "intuition" only in the following clause leaves the reader with nothing to attach the pronoun to in the moment they read it. This is a comprehension gap, not just a style nit. The same applies to definite references — "the structure," "this approach," "that problem" — which assert the noun was already introduced. A section heading does not count as that introduction: "Structure is what does the work" reads fine as a fresh claim under a heading, but "The structure is what does the work" points back at a body sentence that never ran. Drop the article, or introduce the noun in prose first.
• Claims of personal knowledge vs. sourced knowledge must be accurate. "The cleanest test I know" claims lived, personal ownership of something that was actually learned from research. That's a misattribution of where the claim's authority comes from, not a tone issue.
• A causal or comparative chain can't quietly swap its terms. If two distinct conditions are introduced (e.g., "the right training" vs. "the right environment"), later sections can't conflate them or leave it ambiguous which one a diagnosis is actually about.

Run this check on every pass, not just copyedit — a logical inconsistency can be introduced during structure (a recap that outlives the framework it referenced) or edit (a fix to one sentence that breaks a pronoun's antecedent two sentences later). Surface it explicitly even when the surrounding prose reads cleanly.

---

Function Reference

Each function's full rules live in its own file. Load the one you need; the whole set does not have to be in context at once.

Function	File	What it covers
structure	structure.md	Opening, flow and section logic, the close, format defaults, output shape
edit	edit.md	Cutting, Grounding Rules, throat-clearing, rhythm, vocabulary, lists, audience calibration
audit	audit.md	Severity tiers, the two-pass method, output shape
copyedit	copyedit.md	Grammar, punctuation, usage, consistency, proofreading, and what it never does

The Grounding Rules in edit.md are the ones other skills in this suite cite by name — read that file before drafting, not just before editing.

---

Running the Full Sequence

When invoked generically ("run the writing skill," "edit this," "make this better"), run all four in order and return the result of each stage. (How this sequence slots under a voice layer is covered in "Sequencing With an Author-Voice Skill" near the top of this file.)

1. structure — report architectural issues and apply agreed reorderings/cuts. If the structure needs the writer's decision (which of two openings, whether to cut a section), surface the fork with options rather than guessing.
2. edit — tighten every line, apply grounding rules, remove tells, fix rhythm. Surface grounding-rule flags for the writer's ruling.
3. audit — two-pass anti-AI sweep on the edited copy.
4. copyedit — final mechanical pass: grammar, punctuation, usage, consistency, proofreading. Errors fixed; possible voice choices flagged as questions, not overwritten. Runs last because it works against near-final copy.

Between stages, pause at genuine forks — a missing personal anchor, an unverifiable claim, a choice between two valid structures. Don't fabricate to keep moving. A flagged gap the writer fills beats a smooth invention.

Replacement text inherits the check it replaced. A fix applied mid-sequence — a line rewritten because edit flagged it, a sentence swapped out during a later revision pass, a rewrite made in response to the writer's own redirection — is new text, not a patch exempt from scrutiny. Run it back through whichever check flagged the original problem, Grounding Rules at minimum, before treating the fix as final. A rewrite that trades one violation for a different one (a vague sentence replaced with a fabricated fact, a cliché replaced with an unmeasured quantitative claim) is not a fix. Catching that is the same job audit's two-pass method already does for a whole draft — apply it to the one line that just changed, every time a line changes after the first pass, not just at the end.

This holds above the line level too. A heading renamed mid-conversation, a section added or reordered, a subtitle changed after the headings were already set — each of these is a structural edit, and it inherits structure's checks the same way a line inherits Grounding Rules: run the heading-collision check and the paragraph-reshuffle test again against the whole document, not just confirm the one changed piece reads fine on its own. An edit that only checks itself, in isolation from what it now sits beside, is exactly how a heading rename quietly collides with a heading three sections down that nobody re-read at the same time.

---

What This Skill Does Not Do

• Set voice register or rhythm signature — that's the author-voice skill layered on top. This skill won't make writing sound like a specific person; it makes writing structurally sound, tight, grounded, and free of machine patterns.
• Inject personality to defeat detection — this skill cuts to what's earned; it never adds manufactured asides, vulnerability, or current-events backdrop to seem more human. Humanity comes from real specifics and a real voice (the author-voice skill's job), not from injected tics.
• Fabricate to fill gaps — flags missing anchors and unverifiable claims for the writer; never papers over them.
• Make series or project-level architecture decisions — those belong to the writer and the project brief.

---

Sourcing Note

The structural principles (cold opens, callback discipline, closes that point forward instead of summarizing) draw on general editing craft shared across journalism and writing-craft traditions broadly — not any single author's specific named method. The anti-AI detection draws on systematic pattern research. The copyedit function draws on standard usage and style references (Strunk & White's The Elements of Style, the Chicago Manual of Style, Garner's usage) — applied in service of voice, never over it. This is a general craft skill; it layers under whatever author-voice or closing-polish skill is in use, and doesn't implement any one person's specific named framework (a skill like ann-handley-voice, where one exists, is where that layer belongs).