Blog writer

Blog Writer

Channel-master skill for long-form blog content. Owns ideation → beats → draft → image → publish.

Architecture: beats-first (v0.5.0) + plugin-backed publish (v0.4.0). Each post lives in its own container with two sibling docs: a Beats doc (the structural commitments — beat list + title/preview/slug as B0) and a Draft doc (the voice-poured prose). Beats reshape regularly; draft re-pours via /authors-voice against a per-site anchor (voice/anchor-<site-slug>.md). Publish is a single post_to_blog MCP call against a site registered via add_blog_site — site-specific frontmatter (layout, author, prerender, date → publishedDate for Astro, etc.) lives on the github plugin's per-site config in ~/.openwriter/config.json, not in project config. Mirrors book-writer's discipline at post scale.

Convention

This skill obeys the shared writer contract at WRITER-CONVENTION.md. Brief shape and return shape match that doc. Sub-form values: long | short | tutorial | announcement.

OpenWriter pad mechanics are canonical in /openwriter — read it, don't re-derive from tool descriptions. The load-bearing rule: populate_document is create-only, used ONCE; it re-sends the whole body, so calling it again to "fix" a doc appends a duplicate. All edits (rewrite / insert / delete) go through write_to_pad on fresh node IDs. Plus the read ladder (outline_doc → search_docs → peek_doc → read_pad).

Blog posts ship SEO-complete in the FIRST pass. A blog post almost always carries SEO — internal/cluster links, meta + slug + tags, snippet-targeted headings, FAQ/schema-eligible content all go into the initial draft (the minion brief + first populate_document), never bolted on after. Bolting SEO on later forces an edit (→ write_to_pad, never a 2nd populate). SEO-strategy-led posts (pillar / landing / comparison) front-end through /seo-writer if you have it installed (not bundled with OpenWriter).

Cover image is a publish gate (firm rule 12). Placement rules: no bottom-band text (the platform renders og:title there), keep the subject center-safe, ≤5–7 words of overlay text, do not echo the title/description. Palette / brand skin comes from the project's style_doc (e.g. recipebox.md). Generation, canonical size, and the tmp-first rule for text-overlay covers: docs/images.md.

ABOUT TO HAND-EDIT A PUBLISHED POST'S .md, ITS image: FRONTMATTER, OR DROP A COVER INTO THE REPO'S public/? STOP. The OpenWriter doc's blogContext.coverImage is the single source of truth, and post_to_blog is the ONLY writer to the blog repo. Edits flow doc → Accept → Publish — never doc and repo in parallel, because that drift is exactly what breaks idempotent republish (Publish rewrites <content_dir>/<slug>.md wholesale and your hand-edits vanish). Never hand-name a cover; never side-channel an image; never git push the post yourself. Canonical publish flow + the lastPublish-clobber footgun: docs/integrate.md.

Modes

Mode	Trigger	What it does	Sub-doc
setup	/blog-writer setup, "register blog repo", "add blog site"	One-time per blog: inspect_blog_repo clones the target, auto-proposes frontmatter_defaults + frontmatter_field_map from existing posts; add_blog_site persists the site config	docs/setup.md
brainstorm	/blog-writer brainstorm, "brainstorm blog topics"	Open a Blog Ideas doc; propose 3-5 candidate angles with tone + length labels; hand off to beats when user picks	docs/brainstorm.md
beats	/blog-writer beats, "extract beats", "blog beats"	Query-first beat extraction with 3-pass (short/announcement, 3-5 beats) or 5-pass (long/tutorial, 8-15 beats); 9 blog category tags (CLAIM/REFRAME/MECHANISM/EVIDENCE/DEMO/SCENE/OBJECTION/APHORISM/PIVOT) + HOOK/CTA positional roles; locks title + preview + slug as B0; dopamine arc layered with conversion arc	docs/beats.md + docs/titling.md
draft	/blog-writer draft, "draft this post", "pour the beats"	Per-beat dispatch to /authors-voice Apply Protocol with site-specific anchor; cross-beat coherence pass; supports beat-level reshape loop (re-pour only the affected beats)	docs/draft.md + docs/voice-anchor.md
images	/blog-writer images, "blog image", "featured image"	Read style doc, craft prompt, generate cover via insert_image (no docId → returns path → blogContext.coverImage); optional inline images at afterNodeId	docs/images.md
integrate	/blog-writer integrate, "publish", "post to blog"	post_to_blog against the registered site — builds frontmatter from blogContext + site defaults, copies referenced /_images/... to image_dir, rewrites paths, commits, pushes; verifies pending decorations accepted first	docs/integrate.md
pipeline	/blog-writer pipeline, "full blog workflow", "write and publish"	7-step sequence (Setup → Brainstorm → Beats → Draft → Images → Accept → Publish → Verify) with explicit gates; reshape loop returns to Beats step	docs/pipeline.md

Modes can chain (pipeline runs them in sequence) or stand alone. Beats reshape → draft re-pour is the inner loop the architecture is designed around.

Setup (per-session vs per-blog)

Per-session. Read the project's CLAUDE.md and extract the optional ## Blog section. It can carry writing_rules, style_doc, content_driven, aspect_ratio. Used by images mode + as defaults for beats/draft. Schema: docs/project-config.md. If absent, defaults apply and the skill still runs.

Per-blog (one-time). Each blog repo must be registered with the github plugin via add_blog_site before integrate can publish to it. Run /blog-writer setup once per blog. The setup mode uses inspect_blog_repo to auto-propose frontmatter_defaults (layout, author, prerender — fields constant across the site's existing posts) and frontmatter_field_map (e.g. date → publishedDate for Astro sites). Sub-doc: docs/setup.md.

Architecture (build internals)

Per-post container structure (Beats + Draft sibling docs), the 10-step OpenWriter call-order for building one post, and the return/output contract live in docs/architecture.md. Two facts ride every build:

• Two sibling docs per post. Beats — <Title> (content_type notes) holds structure; <Title> (content_type blog) is the publishable Draft — one per container. Reshape beats → re-pour only the affected beats, never the whole post.
• ABOUT TO CALL switch_document (or any view-control MCP) MID-BUILD? STOP. The agent builds the container + Beats + Draft + metadata + images silently, targeting docs by docId; the user watches the activity feed and navigates themselves. Only switch_document on an explicit user instruction ("open the Draft", "show me the Beats").

Firm rules

1. Beats before draft. Draft mode requires a locked Beats doc. No beats → run beats first. Pouring prose without committed beats produces shapeless drafts that need full structural rework downstream.
2. Beats and Draft live in separate docs, always. One container per post, two sibling docs: Beats — <Post Title> (content_type: notes) and <Post Title> (content_type: blog — the publishable doc). Beats reshape → draft re-pour. Don't mix prose into the Beats doc; don't put structural commitments in the Draft doc.
3. Title + preview + slug are B0 commitments. Locked in the Beats doc's B0 block during beats mode. When draft mode runs: title mirrors to the Draft doc's title field via rename_item (the publish plugin reads title from there, not from blogContext.title); preview + slug mirror to blogContext via set_metadata. See docs/titling.md. The first paragraph of the draft must echo the title — can't echo what isn't locked.
4. Per-site voice anchor. draft mode reads voice/anchor-<site-slug>.md where <site-slug> is the slugified site label from list_blog_sites. Silent fallback to voice/anchor.md if not present. Discovery is convention-based, NOT in plugin config. See docs/voice-anchor.md.
5. Setup before integrate. A blog repo must be registered via add_blog_site before integrate can publish. Check list_blog_sites at session start — if the target repo isn't there, run /blog-writer setup first.
6. Project config is optional; per-site config is canonical. ## Blog in CLAUDE.md is for writing_rules, style_doc, content_driven, aspect_ratio. Frontmatter shape (layout, author, publishedDate vs date, etc.) lives on the github plugin's per-site config, NOT in project config.
7. Voice always. Every beat pours through /authors-voice Apply Protocol with the site-specific anchor. The skill owns shape; voice owns diction. Per-beat dispatch is the default for long/tutorial; collapse to single dispatch for short/announcement under 1000w.
8. Two-step doc creation. create_document (spinner) → populate_document (content). Never inline a 30s generation into one tool call.
9. Today's date in the doc. Set blogContext.date as YYYY-MM-DD in the project's configured timezone (default America/Los_Angeles). The plugin formats / renames per site config.
10. Accept pending decorations before publishing. Agent-inserted images and rewrites land as pending decorations until the user accepts them in the right-rail Review tab. post_to_blog reads the canonical doc on disk — pending changes don't ship. After agent writes, integrate must tell the user "click Accept All in the right rail, then I'll publish" rather than silently posting incomplete content. See docs/integrate.md.
11. Mark-sent is automatic. After a successful post_to_blog, the plugin writes blogContext.lastPublish = { publishedAt, publishedUrl, commit, file } on the Draft doc. File tree shows a green ✓; right-click menu surfaces "View Post." Same convention as tweets / articles / newsletters.
12. Cover image before publish. No post integrates without a cover/OG image — a gate, not optional. Design per the Convention above (placement rules + style_doc palette; docs/images.md mechanics).

Anti-patterns

• ❌ Calling draft without first running beats for the post — fails with "no Beats doc found in container"
• ❌ Calling post_to_blog without first running setup for that repo — fails with "no blog site with id X"
• ❌ Putting site-wide constants (layout, author, prerender) into blogContext instead of the site's frontmatter_defaults
• ❌ Calling post_to_blog immediately after insert_image without waiting for accept — pending image stays in browser overlay, doesn't ship
• ❌ Writing the blog post .md file directly into the target repo via Write/Edit tools — that's post_to_blog's job
• ❌ Generating an image before the content is approved (image themes might shift)
• ❌ Skipping voice protocol because "the draft sounds fine"
• ❌ Mixing modes — finish one, then start the next
• ❌ Single global voice anchor when a site-specific one exists — draft mode reads conventionally; no opt-in required
• ❌ Reshaping beats AND re-pouring the entire draft in one pass — reshape beats first, lock them, then re-pour ONLY the affected beats
• ❌ Calling /blog-pipeline, /blog-images, /blog-integrate, /blog-feature-images as separate skills (deprecated stubs that redirect here)

Scripts

• mcp__openwriter__insert_image — Gemini image generation directly into the active OpenWriter doc (primary path; cover via path-return + blogContext.coverImage)
• image-gen CLI (if installed locally) — standalone fallback for non-OpenWriter workflows
• Sharp conversion (PNG → WebP) — for projects where the target site insists on .webp (the github plugin copies PNGs unchanged; conversion is a project-specific concern before publish)

Related skills

Delegated / required:

• /authors-voice — voice pipeline; REQUIRED for any prose generation (draft delegates every beat dispatch).
• /anti-ai — final AI-tells fingerprint scrub; recommended after a post is voice-poured.
• openwriter — the workspace/document MCP; REQUIRED for the doc management this skill governs.
• /seo-writer — optional, not bundled: SEO-strategy-led posts front-end there, then reuse this skill's publish path.

NOT covered (use your own tooling instead): project deploy pipelines · cross-channel announcements (Discord, X, etc.) · /newsletter-writer (different channel) · /x-writer (different channel) · /book-writer (multi-chapter books — global beat sheet, workspace management).