blogpost
Emily's blog post pipeline: capture, draft, revise, finalize posts; maintain posts/INDEX.md; delegate research and images to sub-agents.
SKILL.mdskills/blogpost/SKILL.md
GitHubBlog post workflow
Folder contract lives in AGENTS.md. Two rules override everything:
- NEVER edit or overwrite
initial-draft.md. - Log feedback to
feedback-log.md(with the passage it refers to) BEFORE editing the draft.
Before drafting or revising, read skills/blogpost/references/voice-style-guide.md if it exists.
Intake interview
Use this when Emily brings a new rough idea, voice memo, or rambly draft and the main point is not clear yet.
Goal: help Emily figure out what she is actually trying to say, both for herself and so the writing workflow has a clear target.
Ask one concise question at a time. Do not interrogate. Stop when the answer is clear enough to write:
- One-sentence main point
- Intended reader/audience
- Hook or tension
- What makes the idea interesting or different
- Desired reaction: what the reader should think, feel, or do
- Likely format: longform-substack, short-substack, or tiktok
Good first question after capture:
- "What's the one sentence takeaway you want someone to repeat after reading this?"
Write the current answer in main-point.md. If the main point changes during discussion, update main-point.md and record Emily's correction there.
Protect against scope creep: if new side ideas appear, capture them in tasks.md or create a sparked-by post, then return to the current piece.
posts/INDEX.md
Master table, one row per post:
| slug | working title | type | status | started | last worked | next action | sparked by |
|---|---|---|---|---|---|---|---|
| creativity-walks | Why my ideas only come on walks | longform-substack | revising | 2026-07-02 | 2026-07-05 | awaiting Emily feedback on section 2 | — |
- type: longform-substack | short-substack | tiktok. Ask if unclear.
- status: idea | drafting | revising | researching | final | posted | parked.
Update ONLY on: (1) folder creation -> add row; (2) status change; (3) session end ("end session", "done for now") -> update last worked + next action for posts touched; (4) resuming a post -> correct stale row. NOT on every file edit.
"What's in flight?" -> read INDEX.md aloud: slug + status + next action. Skip posted/parked unless asked.
Research delegation
When Emily flags something to research, or a claim needs verification, spawn a sub-agent (do not research in this session):
sessions_spawn with task: "Read workspace/skills/blogpost/references/research-worker.md and follow it exactly. Post slug: <slug>. Research question: <one specific question>. Main point of post: <2 sentences from main-point.md>. Existing findings: <list titles already in research.md, or 'none'>."
One question per spawn; parallel spawns are fine. When a worker announces completion, read the new entries in research.md and give Emily a one-paragraph spoken summary. Do not block waiting; keep the conversation going.
Clean-room draft delegation
Use this only when Emily wants an independent agent draft for comparison, not when the main agent is drafting directly from Emily's words.
sessions_spawn with task: "Read workspace/skills/blogpost/references/draft-worker.md and follow it exactly. Post slug: <slug>. Main point: <summary from main-point.md>. Interview notes: <notes only; do not include Emily's own draft text>. Recommended format: <longform-substack | short-substack | tiktok>."
The draft worker must not see initial-draft.md; it writes only to agent-draft.md.
Image delegation (Phase 5 only)
sessions_spawn with task: "Read workspace/skills/blogpost/references/image-worker.md and follow it exactly. Post slug: <slug>. Final draft: posts/<slug>/final-draft.md. Title chosen: <title, or 'not chosen'>."
The image worker should report local image paths and, when possible, browser-openable GitHub preview/raw links without exposing credentials. The detailed image workflow, reusable image-direction prompt, and browser-link rules live in references/image-worker.md.
Phase 1 - Capture (new idea or voice memo)
- Create
posts/<slug>/. - Save her words verbatim ->
initial-draft.md. - Write your best current read of the main point ->
main-point.md; mark uncertainty if unclear. - Add INDEX.md row (status: idea).
- git add -A && git commit -m "capture <slug>".
- Reply: your read of the main point + at most ONE intake question. STOP and wait.
Phase 2 - Clarify / think it through
- Continue the intake interview when the argument is still fuzzy.
- Ask one question at a time until the main point, audience, hook/tension, and desired reaction are clear.
- Update
main-point.mdwhen Emily clarifies or corrects the idea. - Research flags -> spawn per Research delegation (or add to tasks.md if she says later).
- Side idea -> new folder + initial-draft.md + INDEX.md row with
sparked by, or add totasks.mdif it is not a post yet. Return to current post.
Phase 3 - Draft
- Read
main-point.md,initial-draft.md, andvoice-style-guide.mdif it exists. - Write
working-draft.mdfrom her draft + conversation. Preserve her phrasing; she will rewrite most of it. - Status -> drafting. Commit.
- Reply: word count + readback options (by section | straight through | changes only). STOP and wait.
Phase 4 - Revise (loop)
- Read back in her chosen unit. After each chunk, STOP for feedback.
- Each piece of feedback ->
feedback-log.md: passage reference, feedback, status (pending | applied | discussed-not-applied). - "Change X" -> apply. Question/reaction -> answer, log as discussed-not-applied, do NOT edit.
- Before each substantial edit pass: git add -A && git commit -m "pre-edit <slug>".
- Status -> revising.
Phase 5 - Finalize (on explicit approval)
- Write
final-draft.md, fully formatted (headers, paragraph breaks, bullets), Substack-ready. - Generate 3-5 title/subtitle pairs ->
titles-and-images.md. Spawn image worker per Image delegation. - Status -> final. Update INDEX.md. Commit and push.
- Reply: file paths, open Emily tasks, spawned post slugs.