format-for-publish

Convert the cited draft into a Strapi-ready package — clean markdown body + Strapi-shaped JSON payload + a README with paste/publish instructions. Pushes directly to Strapi via API when STRAPI_API_TOK

SKILL.md

GitHub

Format for Publish Skill (Strapi)

Transform the cited draft into a publish-ready package for Strapi. The skill produces three files: a clean markdown body (for pasting into Strapi rich-text fields), a Strapi-shaped JSON payload (for API publish), and a README with step-by-step publish instructions.

Input

For slug {slug}:

  • content-pipeline/6-drafts-cited/{slug}.md (the cited draft)
  • references/strapi-format.md (application rules)
  • ../../templates/strapi-format.md (canonical Strapi reference)

Process

  1. Read the cited draft. Make a working copy in memory; don't edit the source.
  2. Strip the editor-notes section if present (everything from ## Editor notes onward). That's internal pipeline metadata, not for publishing.
  3. Extract the H1 as the title and remove it from the body. Strapi renders the title separately; an H1 in the body causes duplicate-title bugs.
  4. Apply callout transformations per references/strapi-format.md:
    • **Tip:** ... and **Pro tip:** ... → :::tip ... :::
    • **Note:** ... and **Sidenote:** ... → :::note ... :::
    • **Editor:** ... → :::editor ... ::: 4b. Convert GFM tables for the live renderer (PLEAA-567 workaround — publish boundary ONLY). The draft/preview keep real markdown tables; the public Next.js renderer can't render GFM tables yet (CTO ticket pending), so at THIS stage convert each table in the working copy: render it to a legible table-card PNG (same matplotlib table-card path /generate-visuals uses for type=table text-tables), upload as a media block where the table stood, and keep a compact bulleted text version beneath the image (one bullet per row, bold first-column label) so the content stays crawlable. When the renderer fix ships, delete this step — nothing upstream changes.
  5. Build the Strapi v5 JSON payload with title, slug, description (first 1–2 sentences of intro, ≤80 chars — also serves as <meta name="description">), blocks[] (single shared.rich-text component holding the markdown body), category (documentId resolved via /api/categories), publishedAt. The schema is strict — author_name, read_time, cover_image_url, tags, excerpt, content, seo are NOT in the Article content-type; Strapi 400s on them. Author is attached manually in admin if needed. The article-level cover relation is auto-attached on --auto-publish to the first uploaded image (PLEAA-570, 2026-05-11) so RSS feeds, OG tags, social cards, and admin preview resolve the hero too. SEO surface is title + description only (PLEAA-457 DOD#4 resolved 2026-05-07).
  6. Run the formatter script:
    python .claude/skills/format-for-publish/scripts/format_for_strapi.py "<slug>"
    
    The script writes the three files to content-pipeline/8-publish/{slug}/.
  7. If the user has wired up Strapi API access via Doppler AND wants direct publish, append --publish:
    doppler run -- python .claude/skills/format-for-publish/scripts/format_for_strapi.py "<slug>" --publish
    
    Requires STRAPI_BASE_URL and STRAPI_API_TOKEN env vars. Article is created in Strapi as a DRAFT (publishedAt = null) — editor publishes manually after review.
  8. Tell the user the output paths and which option (paste vs API) is active.

Output

content-pipeline/8-publish/{slug}/

  • article.md — clean markdown body (paste in Strapi rich-text field)
  • article.json — Strapi-shaped payload
  • README.md — paste-or-publish instructions with title, slug, excerpt, SEO fields, suggested categories/tags

Quality checklist

  • H1 stripped from body (title lives in JSON only)
  • No **Tip:** / **Note:** / **Editor:** markdown prefixes remaining in body — all converted to ::: callouts
  • Editor-notes section excluded from output (it's pipeline metadata, not content)
  • description is real prose from the intro, not the title or first heading
  • description ≤ 80 chars (Strapi v5 cap)
  • blocks[0].__component == "shared.rich-text" and blocks[0].body non-empty
  • No legacy v4 fields (excerpt, content, seo, categories[]) in payload
  • category (when set) is a documentId STRING, not an array
  • Slug matches the input slug exactly
  • publishedAt: null for draft mode; ISO timestamp only when --auto-publish is set

Customizing for your Strapi schema

The default JSON shape assumes a typical "Article" content type with: title, slug, excerpt, content, seo (component), categories (relation), tags (relation), publishedAt. If your Strapi has different field names or component shapes:

  1. Open format_for_strapi.py
  2. Edit the build_payload() function — adjust field names, nest components per your schema
  3. The transformation logic (callouts, excerpt extraction) works regardless of payload shape

When direct API publish fails

Common causes:

  • 401 Unauthorized: API token missing or wrong scope. Check Strapi admin → Settings → API Tokens; needs at least "Authenticated" or "Full access" for content creation
  • 400 Bad Request: Payload doesn't match your Strapi schema. Check article.json against your real content type — likely a missing required field or extra unknown field
  • 404 Not Found: STRAPI_BASE_URL is wrong or the endpoint isn't /api/articles (e.g., your content type is named differently — check Settings → API Tokens → endpoint list)

When --publish fails the script keeps the local files so you can paste manually.

Auto-publish vs draft modes

The skill supports two publish modes:

Draft mode (default; what --publish does without --auto-publish):

  • publishedAt: null in the payload
  • Article enters Strapi as a draft
  • Editor reviews in admin, hand-edits, captures any remaining screenshot placeholders into media library, then clicks Publish
  • The human checkpoint is where the last 20% of quality lives

Auto-publish mode (--auto-publish flag, or BLOG_AGENT_AUTO_PUBLISH=1 env var):

  • publishedAt set to the current ISO timestamp in the payload
  • Article goes live on Strapi immediately
  • Required for /auto-blog-loop autonomous mode (no human in the loop, by design)
  • Compensating control: a quality precondition gate (see "Quality precondition" below) refuses to publish if the upstream /quality-check verdict is FAIL — belt-and-suspenders even if the orchestrator is misconfigured

Pick auto-publish for the cron-driven /auto-blog-loop path. Pick draft mode for one-off manual runs where an editor is going to review.

Quality precondition (auto-publish only)

Before issuing the Strapi POST when --auto-publish is set, the script re-reads content-pipeline/quality-checks/{slug}.md and parses the verdict line. If the verdict is FAIL OR if the file is missing, the script refuses to publish, exits non-zero, and prints the reason. The orchestrator will then quarantine the slug. This is a safety net — /blog-pipeline should never call /format-for-publish --auto-publish on a FAIL'd article, but the precondition makes a misconfigured orchestrator harmless.

Publish gate (PLEAA-581) — artifact must be on origin/main

--publish and --auto-publish both run an artifact precondition before any HTTP write to Strapi (POST /api/articles, PUT /api/articles/{id}, POST /api/upload). The script verifies:

  1. content-pipeline/8-publish/{slug}/article.json exists locally, AND
  2. The same path is present in origin/main HEAD (git ls-tree -r origin/main -- <path> returns a non-empty line).

Both must hold. If either check fails the script exits non-zero with a publish gate REJECT message that names the missing artifact and links /blog-pipeline. This is a HARD gate, not a warning — --auto-publish is no longer allowed to write to Strapi for a slug whose package isn't committed and pushed.

Override: --human-approved "<reason>"

The only escape hatch is --human-approved "<reason>" (required, non-empty). When set, the script:

  • skips the artifact check,
  • appends one tab-separated line to content-pipeline/audit/publish-overrides.log with: ISO timestamp, slug, reason, git config user.email, current branch, HEAD sha,
  • prints publish gate OVERRIDDEN — reason=... to stderr,
  • proceeds with the publish flow.

Use the override only for legitimate one-off operations like a manual unpublish/reseed or a hotfix where the package can't be committed first. Every override is forensically recoverable from the audit log. The log file is gitignored (per-environment append-only); rotate it to long-term storage via your own retention policy if you need an organization-wide trail.

Why this exists

PLEAA-577 (root cause + audit) traced an incident where eight Strapi articles went live without ever passing through /blog-pipeline — the Strapi UI / API token had publish authority with nothing local to gate against. Layer 1 of PLEAA-581 is the client-side belt: a --publish call without an artifact in origin/main is now refused at the script level. Layers 2 (Supabase approval registry + edge-function gate) and 3 (Strapi token surface reduction) live server-side and continue to enforce the same rule even when this script is bypassed.

manifest.json (Layer 2 hook)

Every run of format_for_strapi.py writes content-pipeline/8-publish/{slug}/manifest.json alongside article.md / article.json / README.md. The file is small (slug, title, pipeline_run_id, approved_by, generated_at) and is the source-of-truth artifact that the sync-publish-approvals.yml GitHub Action upserts into the Supabase blog_publish_approvals registry on push to main. The edge function (sync-blog-posts) then refuses any Strapi entry.publish webhook whose slug has no registry row.

Whiteboard staging (PLEAA-448)

After the publish package is written (and any Strapi POST has run), the script bakes the GitHub-Pages whiteboard artifacts for the slug:

  1. Calls scripts/bundle_viewer.py <slug> → writes docs/run-<slug>.html (self-contained offline viewer with every stage's content inlined).
  2. Idempotently appends "<slug>" to the fallback runs array in docs/index.html (the catch (e) { return [...] } block at ~line 197). If the slug is already there, the file is not rewritten.
  3. Best-effort git add docs/run-<slug>.html docs/index.html so one git commit && git push from the operator surfaces the run at https://lionelndong.github.io/blog-agent-2/ within ~1 minute.

This step is best-effort — any failure prints a warning and the operator can re-run python scripts/bundle_viewer.py <slug> manually. Set BLOG_AGENT_SKIP_WHITEBOARD=1 to disable (e.g. for one-off non-board runs).