Skill: /workflow-stage-draft

Four-stage drafting pipeline. Outputs a clean, voice-corrected, style-checked draft that's ready for /publish-post to take over.

When to use

• User says "draft a post on X" / "write this up as a blog" / "prepare a draft on Y".
• A /learned or /weekly-learnings run surfaced a publishable thread and the user wants to develop it.
• An existing draft needs a voice + style pass before publishing.

When NOT to use

• The piece is already drafted AND voice-corrected → call /publish-post directly.
• Capturing notes, not drafting → use /log or the obsidian skill.
• Ghost-writing in a different voice → use /ghost first, then this workflow.

Steps

Stage 0: Grounding — step interview — invokes writing-interview

Runs only when the piece needs material Shane alone can supply and no draft or source text exists yet. Asks, then waits. No prose is written in this stage.

Skip when the input is an existing draft, or when the piece needs no personal material (a doc rewrite, a definition, a status note).

Output: Shane's own language, examples, and uncertainties, plus a list of vault notes backing the piece. Checkpoint: the interview is answered. Never proceed on an unanswered interview.

Stage 1: First draft — step draft

Two paths depending on the input:

• From an idea / outline (no existing draft): write a first draft in Shane's voice, anchored on whatever vault evidence exists. ~600–1200 words. Save to Inbox/Draft - <slug>.md.
• From an existing draft (user pastes or names a vault note): read it; skip drafting; proceed to humanize.

Before writing prose, write the outline with a word budget per H2 — tables and code blocks are their own line items, not free. Check the running count at each section boundary, not at the end. Over budget means cut a section or merge two, never trim adjectives: lexical trimming preserves the structure that caused the overrun.

Grounding rule — never invent a specific. Write only what the vault evidence and Shane's own input support. Profiles/voice.md binds here: "The voice is portable. The thinking isn't." Where the piece needs a fact, number, incident, or opinion the source doesn't contain, leave a marked gap:

[TK: how many reviews had you run before you noticed? Rough number is fine.]

Never fill a [TK] with a plausible specific. List every open [TK] at the end of the run. Three honest gaps is staged work; three invented numbers is a retraction waiting to happen.

Output: draft file in vault Inbox. Checkpoint: confirm draft direction before voice pass.

Register check before drafting. Name which of these the piece is, because the rules differ and applying essay rules to a guide damages both:

Only an Argument owes a disputable thesis and a closing punch. A Guide may use predictable headings and may simply stop when the steps end — voice.md's punch-ending law is an essay law, not a universal one. Most of the Agentic and DevRel series is Explanation, and the beat doc's archetype split reflects that.

Stage 2: Voice pass — step humanize — invokes humanize

Removes AI-typical patterns (hedging, list-iness, em-dash overuse, throat-clearing) and restores Shane's conversational rhythm. Preserves technical precision.

Output: humanized draft (in-place edit to the Inbox file). Checkpoint: skim before style pass.

Stage 3: Style pass — step style — invokes ms-style-pass

Applies Microsoft Writing Style Guide term preferences, bias-free language rules, and heading conventions. Doesn't touch voice.

Output: style-corrected draft (in-place edit). Checkpoint: confirm draft is publish-ready before handoff.

Handoff (NOT a step — explicit to user)

After style, print:

─── Draft staged ───
File: Inbox/Draft - <slug>.md
Word count: <N>
Open gaps: <N> [TK] items — list them, or "none"

Provenance:
  Author material: <which sections or language came from the vault or from Shane>
  Model contribution: <organization, structure, connective prose, or none>
  Open items: <the [TK] list, or none>

Ready for /publish-post when you are. That skill will:
  - generate brand artifacts (blog hero, OG card, LinkedIn companion)
  - run any final style checks
  - publish to shane.logsdon.io
  - refuse to publish while any [TK] is unresolved

Or, to iterate first: edit the draft directly and re-run /workflow-stage-draft --resume-from humanize.

Resume protocol

Default: run interview → draft → humanize → style in order, with checkpoints between each. interview is skippable (see below); the other three are not.

--resume-from <step>: skip every step before <step> and begin from <step>. The named step itself runs.

If <step> is not in {interview, draft, humanize, style}, print:

Invalid resume-from step: <step>
Valid steps: interview, draft, humanize, style

…and stop.

Required state when resuming

Before running the resume step, confirm the listed inputs exist. If anything is missing, prompt the user for it.

• --resume-from interview — needs only the topic. Runs the interview and waits for an answer before anything is drafted.
• --resume-from draft — same as a fresh run; needs the topic / outline / source note. Skip the interview only when the piece needs no personal material.
• --resume-from humanize — needs an existing draft file (in Inbox/ or pasted in conversation). If the user names a different file, accept it; otherwise default to Inbox/Draft - <slug>.md. Confirm the file path before running humanize.
• --resume-from style — needs a draft that has already been through humanize (or that the user explicitly says is voice-correct). Confirm the file path; warn if the file looks unedited (no recent changes since draft).

Examples:

/workflow-stage-draft --resume-from humanize  # user edited the draft manually, re-run voice pass
/workflow-stage-draft --resume-from style     # voice is fine, just want the style pass

Other flags

• --no-pause to run end-to-end (use only when you've already iterated and just want a final pass).

Session open

Check Inbox/STATE.md or the draft project's directory for a STATE.md. If present, read General rules and Known failure modes — these carry learned patterns (e.g. which humanize anti-patterns recur in Shane's technical writing) so they don't need to be re-discovered.

Stage quality gate

After each stage: does the output meet the stage goal (draft = core argument present and in Shane's voice; humanize = AI-typical patterns removed without losing technical precision; style = terms and headings conform)? If it partially misses, retry once with an explicit note on the gap before presenting or continuing. On second failure, surface the issue rather than auto-continuing.

Humanize and style passes should be validated by re-reading with a critic framing (does this read like a human wrote it?) rather than the same pass that produced it.

Piece-shape check, once, at the end of draft. The word budget per H2 makes uniform sections easy, and uniform sections are their own tell. Read the draft for these and fix the dominant one only:

• sections of near-identical length, in identical internal order
• paragraphs that could trade places without changing the argument
• headings doing all the organising, so the prose is a table of contents
• every paragraph ending on a line built to be quoted

One section should be allowed to carry more weight than the others. Do not introduce a tangent, an unresolved thread, or sentence-length wobble merely to look irregular.

Session close

If the humanize or style pass hit recurring issues, append them to this SKILL.md under a ## Known failure modes section. Patterns that compound here improve every future draft run.

Failure handling

If humanize flags content it can't safely de-AI without losing technical accuracy, it stops and surfaces the conflict to the user. style then operates on whatever humanize produced. On any failure, print the failed step plus the resume command (/workflow-stage-draft --resume-from <step>).

Declined imports (checked, and deliberately not adopted)

Rejected after checking against Profiles/voice.md, the Personal Brand Operating Doc, and Gate ∞. Do not re-import these:

• Applying "say something only you could say" to short-form. Gate ∞ already sizes rigour to reach and permanence. Essay-grade authorship demands on a LinkedIn reply produce invented personal material.
• Requiring a disputable thesis on every piece. Only an Argument owes one. See the register table above; most of the published series is Explanation.
• "Stop where the thought stops" on its own. Taken bare it suppresses the closing punch that voice.md mandates. It needs the calibration that goes with it: change repeated use, not the device.
• Replacing Gate ∞ with genre-based medium routing. Reach, permanence, and retractability are the better dial, and they are already tied to the accountability log.