content-pipeline

# Content Pipeline Orchestrator

Produces 3 article drafts per run: research once, then spawn 3 parallel sub-agents (one per article). Sub-agents handle their own skill reads. This orchestrator only handles the research phase and final reporting.

**Do NOT publish anything.** Everything stays as drafts.

---

## Phase 0 — Deduplication Check (MANDATORY FIRST)

1. Read `~/clawd/memory/content-published.md` — topics already covered are off-limits
2. Run `bird tweets --user $TWITTER_USER -n 30 --plain` — note angles the author published in the last 30 days
3. Check recent Substack: `web_fetch https://$SUBSTACK_URL` — find recent post titles
4. Build an EXCLUSION LIST. Do not pick any topic overlapping with it.
5. After picking topics, append them to `~/clawd/memory/content-published.md`

---

## Phase 1 — Research (run once, takes ~15 min)

Read `~/marketing/WRITING-STYLE.md` first.

### 1a. Gather engagement data

- **X/Twitter**: Pull @$TWITTER_USER's last 30 posts via `bird`. Note which topics/angles got most engagement (likes, replies, retweets).
- **Grok X search**: Run `/grok-search` with `--x` to find what's trending in the author's niches (AI, investing, indie dev, fintech).
- **Trend research**: Run `/last30days` across YouTube, X, Reddit, and Substack to find viral formats, hooks, and topics with momentum.

### 1b. Find 3 topics

Topics should come from what's actually resonating on social, not from a fixed theme list. Look for:

- **High-engagement angles** from the author's own X data (what got traction in the last 30 days)
- **Trending conversations** where the author has a unique perspective (builder, investor, AI power user, solo founder)
- **Viral formats** that can be adapted (data tables, outrage comparisons, contrarian takes, build-in-public threads)

Not every article needs to be about Bloom. Pick topics because they're genuinely interesting and the author has something real to say. Bloom integration only if it's natural and adds to the piece.

**Pick topics where the author has unique credibility.** Avoid angles anyone could write. Validate with X engagement data from Step 1a.

### 1c. Competitive gap check

For each topic: quick web search for what's already been written. Find the contrarian angle or data point competitors missed.

### 1d. Save research

Write all findings to `~/marketing/substack/drafts/[YYYY-MM-DD]-research.md`:
- X engagement patterns (which topics resonated)
- Trending topics and viral formats found
- 3 chosen topics with positioning angle for each
- Key SEO keywords per topic

---

## Phase 2 — Parallel Article Agents

Spawn 3 sub-agents in parallel using `sessions_spawn`, one per topic. Each sub-agent receives:

```
Topic: [topic]
Positioning angle: [angle from research]
X engagement insight: [what angle resonated with the author's audience]
Output directory: ~/marketing/substack/drafts/[slug]/

Your job:
1. Read ~/marketing/WRITING-STYLE.md
2. Run /hooks — generate 5 options, pick strongest → hooks.md
3. Run /outline-generator → outline.md
4. Run /writer → draft.md (must pass "only the author could write this" test: use specific numbers, named frameworks, personal experience. Reference Bloom only if naturally relevant to the topic.)
5. Run /editor-in-chief — max 5 iterations → editing-log.md + draft-final.md
6. Generate a hero image → hero.png
7. Write LinkedIn post directly → linkedin-post.md
   - Re-read ~/marketing/WRITING-STYLE.md kill phrases list before writing
   - Write a native LinkedIn post from draft-final.md. Do NOT reframe into "here's what most people get wrong" or "X isn't Y, it's Z" patterns. Lead with the most specific, surprising fact or data point from the article. Let the story carry the reader — no interpretive sentences telling them what to feel.
   - Run /evaluate-content on the post (Voice + Leanness scores only). If either scores below 4/5, rewrite and re-check. Max 2 revision passes.
   - Save final version to linkedin-post.md (post body only — no metadata headers)
8. Write X thread → x-thread.md
   - Run writer to create 5 standalone tweets from draft-final.md and sequence them as a thread.
   - No promotional framing, no article links in thread body (link goes in reply)
9. Format for Substack → substack-ready.md

Do NOT publish. Do NOT send messages unless blocked.
```

### File structure per article
```
~/marketing/substack/drafts/[slug]/
  hooks.md            # 5 options + winner
  outline.md
  draft.md
  editing-log.md
  draft-final.md
  substack-ready.md
  hero.png
  x-thread.md
  linkedin-post.md
```

---

## Phase 3 — Substack Drafts

After all 3 sub-agents complete, for each article save a Substack draft using the Substack skill with `profile=clawd`.

**NEVER click Publish. Draft only.**

Steps per article:
1. Read `~/marketing/substack/drafts/[slug]/substack-ready.md` for title, subtitle, and body
2. Open Substack editor: `browser action=open targetUrl="https://$SUBSTACK_URL/publish/post" profile=clawd`
3. Paste title, subtitle, body
4. Upload `hero.png` as the cover image
5. Save as draft (never publish)
6. Capture the draft URL and include it in the Phase 4 report

---

## Phase 4 — Typefully Drafts

After Substack drafts are saved, create Typefully drafts for each article:

**IMPORTANT: Save as UNSCHEDULED drafts only. Do NOT schedule or publish anything.**

```bash
# X thread draft — unscheduled draft only
cd ~/clawd/skills/typefully && node scripts/typefully.js drafts:create $TYPEFULLY_SOCIAL_SET_ID \
  --platform x --text "$(cat ~/marketing/substack/drafts/[slug]/x-thread.md)"

# LinkedIn draft — unscheduled draft only
node scripts/typefully.js drafts:create $TYPEFULLY_SOCIAL_SET_ID \
  --platform linkedin --text "$(cat ~/marketing/substack/drafts/[slug]/linkedin-post.md)"
```

---

## Phase 5 — Report

Do NOT send via the message tool. Just output the summary as your reply. Cron delivery handles routing.

Include:
- 3 topics chosen (one line each with positioning angle)
- Substack draft URLs (one per article — ready to review and publish)
- Typefully draft links (LinkedIn + X thread, unscheduled)
- Source data and trending topics that informed each article