# Blog Post > Write a new blog post for cryptoflexllc.com: 5-agent captain pipeline with cover graphic, diagrams, voice QA, and CI-matched validation - Skill: `chris2ao/blog-post` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add chris2ao/blog-post` - Raw SKILL.md: https://api.skillmd.com/api/skills/chris2ao/blog-post/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: DevOps & Infra - Author: chris2ao (https://skillmd.com/u/chris2ao) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/chris2ao/blog-post --- # /blog-post - Blog Post Production Pipeline Single entry point for producing a cryptoflexllc.com blog post. Coordinates the Blog Captain (opus), which runs a specialist team: writer, voice agent, senior editor, UX agent, diagram author, and brand-graphics cover artist. This skill is callable from ANY repo or directory. Never rely on the current working directory: resolve all paths absolutely. ## Step 0: Repo Resolution ```bash BLOG_REPO="$HOME/GitProjects/cryptoflexllc" [[ -d "$BLOG_REPO" ]] || BLOG_REPO="$HOME/Github_Projects/cryptoflexllc" ``` Every file operation in this pipeline uses absolute paths under `$BLOG_REPO` or `~/.claude/`. If neither repo path exists, stop and tell the user. ## Step 1: Post Inventory ```bash bash ~/.claude/scripts/blog-inventory.sh --minimal ``` The output now includes `series`, `seriesOrder`, `coverImage`, and `featured` per post, plus a `series_summary` block with each series' post count and highest seriesOrder. Series matching is quote-normalized by the script; never grep frontmatter for series values yourself. From the inventory, derive: - The 3 most recently active series (for the series question below) - The next seriesOrder for whichever series is chosen (`series_summary[].max_order + 1`) - The 2 most recent post paths (calibration posts for the captain) ## Step 2: Voice Profile Read `~/.claude/skills/blog-voice-profile.md` in full. Pass the complete text to the captain (subagents may not be able to read `~/.claude/skills/` themselves). ## Step 3: User Discovery Use AskUserQuestion (one call, four questions): 1. **Destination** ("Destination"): "Production" (live after deploy, `src/content/blog/`) or "Backlog" (draft, `src/content/backlog/`, published later via the /backlog admin page). 2. **Source material** ("Material"): what the post is built from: - "This session": what was accomplished in the current Claude Code session - "Today's work": everything done today across repos (git logs since midnight) - "Last 24 hours": same sweep with `--since="24 hours ago"` - "A specific feature or topic": deep dive on a named skill, feature, incident, or subject (user supplies specifics via Other or a follow-up) 3. **Series** ("Series"): offer the 3 most recently active series from the inventory plus "None (standalone)". The user can type any other series name (including a brand-new one) via Other. When a series is chosen, use the next seriesOrder computed in Step 1; for a new series, seriesOrder is 1. 4. **Tone** ("Tone"): "Educational and friendly" (default voice), "Witty and accessible" (humor, GIFs, Info boxes for every concept; for general readers), or "Technical reference" (documentation style). If the answers leave the topic ambiguous, ask a plain follow-up question before spawning the captain. Never guess at source material: the pipeline must not fabricate. ## Step 4: Research Recipes (pass to captain) Repos for research, all absolute: - `$BLOG_REPO` (the site) - `$HOME/GitProjects/CJClaude_1`, `$HOME/GitProjects/cryptoflex-ops` (private; never linked in posts) - `~/.claude` (claude-code-config) By material type: - **This session:** review the current conversation; identify the most educational moments, real commands, real errors. - **Today's work / Last 24 hours:** parallel Explore agents (haiku) running `git log --since=... --stat` across the repos above, plus CHANGELOG/MEMORY updates. - **Specific feature/topic:** read the actual source files, config, and history behind it. Quote real code only. ## Step 5: Spawn the Captain Spawn ONE Task agent: - **subagent_type:** general-purpose - **model:** opus - **name:** blog-captain Pass it, in the prompt: 1. "You are the Blog Captain. Follow the instructions in ~/.claude/agents/blog-captain.md" 2. `BLOG_REPO` (resolved absolute path) 3. The inventory JSON from Step 1 and the 2 calibration post paths 4. The user's answers: destination, material, series + seriesOrder, tone 5. The full voice profile text 6. The applicable research recipes from Step 4 7. If destination is Backlog: "Write to src/content/backlog/. Commit prefix 'chore: add backlog draft'. Note the backlog runtime differences in your instructions." The captain runs the full pipeline: research, draft, parallel review + diagram authoring, revision, cover graphic, build, and user approval. It returns a JSON report. ## Step 6: After the Captain Returns 1. Display the post details, scores, revision history, cover/diagram outputs, and any unresolved issues. 2. If the captain did not commit (approval pending), offer to commit and push. Production commits include the MDX, `public/blog//`, and any new diagram/registry files. 3. Report the URL: `https://cryptoflexllc.com/blog/` (production) or "Draft saved to backlog; publish via the /backlog admin page." 4. Backlog publishes (user-triggered via the admin page) rename the file on the remote with 2 commits; after one, run `git pull --ff-only` in `$BLOG_REPO`. 5. Backlog drafts in a series carry the seriesOrder assigned at draft time, which goes stale as the series grows. When a draft with `series:` is about to be published, re-check the inventory and bump its `seriesOrder` to `max_order + 1` first. ## Known Gotchas - **Cover uniqueness bar**: every cover is a bespoke composition derived from the post's story, sharing branding only; the contract and composition register live in `$BLOG_REPO/docs/cover-graphics-standards.md`, and the captain reviews the render as an art director (unique vs recent covers, one focal point, balanced, filled panels, legible at card size) before accepting it. - **Diagram visual bar**: inline SVG diagrams follow the editorial diagram system, matching the cover-infographic aesthetic. Contract: `$BLOG_REPO/docs/editorial-diagram-standards.md`; primitives: `src/components/mdx/diagram-editorial.tsx`; exemplar: `ReviewPipelineDiagram` in `diagrams-security-review-round-two.tsx`. Plain outlined boxes joined by crossing diagonal lines are below the bar, and the diagram author must run the standards doc's screenshot verification loop before handing off. - **Tailwind v4 dynamic class purging** (matters for diagram TSX): never interpolate class fragments (`bg-${accent}-600`). Use a static `as const` class map. Hit in TerminalPromo.tsx 2026-07-21. - **MDX runtime traps** (caught by `~/.claude/scripts/validate-mdx.sh`, but know them): bare `<` before digits (`<100ms`) parses as a JSX tag and breaks at render with HTTP 200; nested double quotes inside JSX attribute values render as an error boundary with no build error; slugs are `[a-z0-9-]` only, no dots. - **Series data**: quoting is normalized to unquoted in frontmatter (`series: Claude Code Workflow`). The inventory script is the only source of truth for seriesOrder.