pipeline-runner

pipeline-runner

Chains a use case hypothesis from /content-ideas through the AI strategy and pre-sales pipeline. The full chain is:

last30days → GBrain Recall → content-research → vertical-scorer → ai-strategy-brief → branded-pptx-deck → research-to-strategy → presales-deal-prep → GBrain Write-back

Each stage produces a deliverable and gates the next — a PASS verdict at Stage 2 stops the pipeline early, saving time.

The pipeline consumes existing skills as-is via their slash commands. No business logic lives here — this is pure orchestration.

---

Stage 0: Load use case

Resolve $CONTENT_HOME (default ~/Documents/Content). Find the most recent dated subfolder of $CONTENT_HOME/research/ that contains a feed-data.json with a useCases array. If none exists, tell the user to run /content-ideas first (in strategy mode — their Content Goal must mention strategy/pre-sales).

Parse useCases[]. If the user passed a number as an argument, pick that index (1-based). If they passed a name, fuzzy-match against title. If no argument or "list", present the numbered list:

Use cases from {date} feed:

1. {title} ({confidence}, {len(signals)} signals)
2. ...

Pick a number to run the pipeline.

Once selected, confirm:

Running pipeline for: {title} Hypothesis: {hypothesis} Signals: {count} posts, {count} patterns Suggested prospects: {orgs joined}

Proceed? / Pick a different one / Add a prospect

---

Stage 0.5: Last 30 Days Signal (real-time engagement)

Invoke /last30days "{verticalName}".

This runs before GBrain Recall and before any source gathering. It pulls real-time engagement signals from Reddit, X/Twitter, YouTube, TikTok, Hacker News, Polymarket, and GitHub — ranked by actual upvotes, likes, and prediction-market money, not editorial curation. Output is a synthesized research brief in conversation context.

Pass the use case title or verticalName as the query (e.g., AI agent orchestration for ERP or on-premise LLM compliance). For named-entity topics (product names, company names, person names), the skill generates its own query plan internally.

After the skill completes, extract for downstream use:

• Active communities: which subreddits, X handles, and YouTube channels are most active on this topic
• Engagement spikes: posts or threads with unusually high score (signals timing and framing)
• Polymarket odds: if any prediction markets exist on this vertical, record the odds and payout dates
• Competing products or vendors surfaced in discussions
• Practitioner language: the actual words practitioners use (not vendor copy) — use this to improve Stage 1 source queries and Stage 3 brief framing

These signals feed Stage 1 source selection directly: prefer sources that match the active communities and high-engagement threads found here, and supplement the standard architecture/market/compliance search with specific subreddits, repos, or channels the skill surfaced.

If SCRAPECREATORS_API_KEY is not set, the skill runs in degraded mode (web + HN only, no Reddit/X/YouTube API). Still useful — proceed.

Print status:

PIPELINE: {title}
═══════════════════════════════════════
  ✓ Last 30 Days      {N} signals — Reddit/HN/X/Polymarket/YouTube
  ◻ GBrain Recall     (pending)
  ◻ Content Research  (pending)
  ◻ Vertical Score    (pending)
  ◻ Strategy Brief    (pending)
  ◻ PPTX Deck         (pending)
  ◻ Full Strategy     (pending)
  ◻ Deal Prep         (pending)
  ◻ GBrain Write-back (pending)

---

Stage 1: Content Research (augment the use case)

Search for the best available sources on the use case topic across YouTube, web articles, GitHub repos, and documentation. Select 4-6 high-signal sources covering:

1. Architecture/deployment — how this is built in production
2. Market landscape — who is winning, funding, valuations
3. Compliance/regulation — what governs this space
4. Integration points — what systems it connects to
5. Open-source alternatives — self-hostable options

If gbrain is available as an MCP server, use it by default at the start of Stage 1 to recall prior company research, recurring prospects, related vertical work, prior source bundles, and named-account context before repeating research from scratch. Treat it as the durable knowledge layer for recurring entities and themes, not the system of record for deliverables. Treat that recall as embedding-backed semantic retrieval by default, not just keyword lookup. Prefer semantic recall first; use synthesis only when Stage 1 needs merged interpretation rather than simple recall.

Treat this as a named chain stage:

• GBrain Recall must happen before fresh Stage 1 source gathering when prior work may exist
• reflect that stage in pipeline status when it was used successfully

If the host exposes stronger research plugins such as exa, prefer them for discovery in this stage so official product pages, docs, GitHub repos, competitive/vendor signals, and current operator proof points are found faster and with less search noise than generic web search alone.

In terminal-first hosts such as Codex CLI, prefer the closest equivalent: an MCP-connected research server or a local CLI/API wrapper for tools such as Exa when available. Treat that as the terminal analogue to desktop plugin access.

Concrete terminal patterns to prefer when available:

• Exa MCP over remote/HTTP MCP
• a local Exa API wrapper that calls https://api.exa.ai/search

Codex Desktop plugin access is a discovery advantage, not an exception to the pipeline contract. The same local artifact-generation, branded-deck, QA, repo-rule, and source-verification requirements still apply.

Plugin-assisted research does not replace:

• local file generation
• branded PPTX build and QA
• repo-specific workflow rules
• verifying that final cited sources are primary and current

If the use case involves agent orchestration, coding automation, MCP servers, skills, or OpenHands is named in the source material, include the OpenHands GitHub repo and docs in the Stage 1 source set and treat them as the source of truth for implementation details:

• https://github.com/OpenHands/OpenHands
• https://docs.openhands.dev/

Use those sources to upgrade stack[], architecture notes, and implementation snippets with verified OpenHands primitives rather than generic agent-platform descriptions.

Run /content-research with the selected URLs. This produces:

• Second-brain notes in SECOND_BRAIN_DIR when configured
• Obsidian vault entries under OBSIDIAN_VAULT_DIR/content-research/ when configured
• Knowledge graph entries via /graphify

If SECOND_BRAIN_DIR or OBSIDIAN_VAULT_DIR are unset in this host, skip those exports rather than inventing machine-specific fallback paths. The local run folder remains the system of record either way.

After research completes, update the use case in feed-data.json:

• Upgrade stats[] with research-backed numbers (not estimates)
• Upgrade stack[] with specific vendor/technology names
• Add new signals[] entries from the research sources
• Add new patterns[] from cross-source analysis
• Expand sourceUrls[] with all researched URLs
• Upgrade confidence if signal count increased (3+ = high)

Also write durable findings back to GBrain when they are likely to matter again across sessions: named prospects, validated source bundles, recurring vertical theses, and implementation-stack notes.

This is the closing chain stage:

• GBrain Write-back happens after the run once durable findings are stable
• write-backs should cover reusable vertical theses, validated source bundles, named accounts, implementation-stack notes, and other cross-run memory

Print status:

PIPELINE: {title}
═══════════════════════════════════════
  ✓ Last 30 Days      {N} signals — Reddit/HN/X/Polymarket/YouTube
  ✓ GBrain Recall     semantic retrieval seeded Stage 1
  ✓ Content Research  {count} sources → second-brain + feed-data updated
  ◻ Vertical Score    (pending)
  ◻ Strategy Brief    (pending)
  ◻ PPTX Deck         (pending)
  ◻ Full Strategy     (pending)
  ◻ Deal Prep         (pending)
  ◻ GBrain Write-back (pending)

---

Stage 2: Vertical Scorer (GO/WAIT/PASS gate)

Invoke /vertical-scorer "{verticalName}".

The scorer now benefits from the research gathered in Stage 1 — source URLs and second-brain notes provide grounded evidence for each scoring dimension.

This is a gate:

• GO (score 25+/35) → proceed automatically to Stage 3.
• WAIT/CONDITIONAL (score 18–24) → warn the user, ask whether to continue.
• PASS (score <18) → stop the pipeline. Report the verdict and suggest the user pick a different use case or refine the vertical.

Print status:

PIPELINE: {title}
═══════════════════════════════════════
  ✓ Last 30 Days      {N} signals
  ✓ GBrain Recall     semantic retrieval seeded Stage 1
  ✓ Content Research  {count} sources
  ✓ Vertical Score    {score}/35 — {verdict}
  ◻ Strategy Brief    (pending)
  ◻ PPTX Deck         (pending)
  ◻ Full Strategy     (pending)
  ◻ Deal Prep         (pending)

---

Stage 3: AI Strategy Brief

Invoke /ai-strategy-brief "{verticalName}".

Pass the hypothesis, source URLs, and second-brain note paths as additional context so the brief is grounded in the research from Stage 1, not just web search. The brief should reference the specific market data, competitor landscape, and cost figures from the research notes.

When the brief is produced, update status:

PIPELINE: {title}
═══════════════════════════════════════
  ✓ Last 30 Days      {N} signals
  ✓ GBrain Recall     semantic retrieval seeded Stage 1
  ✓ Content Research  {count} sources
  ✓ Vertical Score    {score}/35 — {verdict}
  ✓ Strategy Brief    {filename}.docx
  ◻ PPTX Deck         (pending)
  ◻ Full Strategy     (pending)
  ◻ Deal Prep         (pending)

---

Stage 4: Branded PPTX Deck

Invoke /branded-pptx-deck to generate a multi-slide presentation from the use case data. The deck uses pptxkit from the branded-pptx-deck skill and follows the Canva-adapted use case realization layout.

This is a hard requirement for client-facing output. Always use the branded PowerPoint template workflow (/branded-pptx-deck, backed by BRANDED_PPTX_TEMPLATE, falling back to ~/.claude/templates/branded-template.pptx when unset). Do not substitute a hand-built python-pptx deck or a blank presentation theme for external/client delivery. If the branded workflow is unavailable, stop and report the deck stage as blocked.

Every slide in the deck must carry structured content, not sparse placeholders. At minimum:

• action-title cover with premise + evidence anchors
• detailed use-case realization slide in the branded layout
• structured market/proof slide
• structured architecture/stack slide
• structured scorecard or risk/controls slide
• structured roadmap/next-step slide

PPTX QA is required before this stage is considered complete:

• Deck.save() / branded builder validation must pass
• text overlap, overflow, and collisions must be checked
• if preview_pptx.py is available, review the contact-sheet output
• if preview tooling is unavailable, report the deck as unreviewed for visual QA rather than presenting it as final client-ready output
• use explicit deck status: draft, reviewed, or blocked
• keep the branded builder script with the run artifacts so QA fixes are reproducible
• deliver the reviewed filename, not an earlier draft
• visual QA checklist:
  • no red overflow boxes in preview_pptx.py
  • no title/subtitle collisions
  • no clipped text in stat bars, callout strips, or side panels
  • footer/page number present on each slide

Slides to generate:

1. Cover — title, premise, date, source count, vertical score
2. Use Case Realization — the Canva-adapted layout: teal left panel with challenge/solution cards, stat boxes, how-it-works, solution stack, systems/users bars; navy right strip with organizations
3. Market Landscape — who is winning table (from strategy brief research)
4. Architecture Patterns — deployment options with cost tiers (from content-research notes)
5. Vertical Scorecard — 7-dimension score visualization with bars
6. Risks & Controls — failure modes and mitigations
7. 90-Day Roadmap — phased entry plan
8. Closing — bottom line takeaways

The deck pulls data from:

• feed-data.json useCases[] → slides 1-2
• Content-research second-brain notes → slides 3-4
• Vertical scorer output → slide 5
• Strategy brief research → slides 6-7
• All of the above → slide 8

Save to: $CONTENT_HOME/research/{date}/{topic-slug}-deck.pptx Copy to: CLIENT_DELIVERY_DIR when the host has a configured user-facing delivery location.

Recommended filename convention:

• {topic-slug}-deck-draft.pptx while content/layout is still changing
• {topic-slug}-deck-reviewed.pptx after visual QA passes
• {topic-slug}-deck-blocked.txt if the PPTX stage cannot be completed cleanly

If CLIENT_DELIVERY_DIR is unset, keep the reviewed deck in the run folder and report that no host delivery directory was configured.

Update status:

PIPELINE: {title}
═══════════════════════════════════════
  ✓ Last 30 Days      {N} signals
  ✓ GBrain Recall     semantic retrieval seeded Stage 1
  ✓ Content Research  {count} sources
  ✓ Vertical Score    {score}/35 — {verdict}
  ✓ Strategy Brief    {filename}.docx
  ✓ PPTX Deck         {filename}.pptx
  ◻ Full Strategy     (pending)
  ◻ Deal Prep         (pending)

---

Stage 5: Research-to-Strategy (optional, user-gated)

Ask: "Generate full strategy research + council + deck? This takes 5–10 minutes and produces a 30-page research doc, knowledge graph, and slide deck."

Options: Yes — run it / Skip — move to deal prep / Stop here

If yes, invoke /research-to-strategy "{verticalName}" {sourceUrls joined by space}.

Update status on completion.

---

Stage 6: Pre-Sales Deal Prep (optional, per prospect)

Ask: "Prep for a specific prospect?" Show the orgs from the use case, plus an option to type a different company name.

If the user picks one, invoke /presales-deal-prep "{prospectName}".

The strategy brief, deck, and vertical context are already in the conversation, so the deal prep skill can reference them.

Update final status:

PIPELINE: {title}
═══════════════════════════════════════
  ✓ Last 30 Days      {N} signals
  ✓ GBrain Recall     semantic retrieval seeded Stage 1
  ✓ Content Research  {count} sources
  ✓ Vertical Score    {score}/35 — {verdict}
  ✓ Strategy Brief    {filename}.docx
  ✓ PPTX Deck         {filename}.pptx
  ✓ Full Strategy     {filename}.pptx + research.md
  ✓ Deal Prep         {prospect}-deal-prep.md

---

Notes

• No downstream skills are modified. This skill invokes /last30days, /content-research, /vertical-scorer, /ai-strategy-brief, /branded-pptx-deck, /research-to-strategy, and /presales-deal-prep exactly as a human would — by their slash commands with string arguments.
• Last 30 Days is the signal layer. Stage 0.5 runs first and surfaces what practitioners are actually saying — the subreddits, X threads, YouTube channels, and prediction markets with real money on the topic. It shapes Stage 1 source selection and Stage 3 brief framing with practitioner language before any formal research begins.
• Content research is the foundation. Stage 1 grounds every subsequent stage in real data — the scorer uses researched evidence, the brief cites specific numbers, the deck renders verified facts, and deal prep references actual market players. Without it, downstream stages rely on web search alone.
• Re-runnable. The user can run /pipeline-runner on the same use case again (e.g., to add a second prospect in Stage 6) without re-running the scraper.
• Multiple use cases. The user can run /pipeline-runner 1, then /pipeline-runner 2 in the same session to evaluate multiple verticals.
• Direct skill invocation. The user can skip the pipeline and invoke any downstream skill directly: /vertical-scorer "On-premise LLM for healthcare" works standalone. The pipeline is a convenience, not a requirement.
• Stage skipping. If the user has already run /content-research for this topic in a prior session, Stage 1 can be skipped — check if second-brain notes already exist for the use case's verticalName before re-researching.