feat: add geo-writing skill and content-automation onboarding flow (JARVIS-899)

[alex-nork]

Summary

• Bundle geo-writing skill as a prompt-only bundled skill with listicle and head-to-head article formats
• Rewrite BOOTSTRAP-CONTENT-AUTOMATION.md for skill-first onboarding: load skill immediately, ask one question, ship a draft, learn voice from edits
• Rewrite wake-up greeting on the daemon side for content-automation cohort — bypasses the canned greeting and lets the model respond via normal inference

PRs merged into feature branch

• feat: add GEO article writing skill as bundled skill #31369: feat: add GEO article writing skill as bundled skill
• feat: update content-automation bootstrap for GEO article flow #31371: feat: update content-automation bootstrap for GEO article flow
• fix: geo skill structure cleanup (references dir, orphaned files, compatibility) #31376: fix: geo skill structure cleanup (references dir, orphaned files, compatibility)
• fix: deduplicate geo skill rules from bootstrap template #31377: fix: deduplicate geo skill rules from bootstrap template
• feat: rename geo -> geo-writing, rewrite bootstrap for skill-first onboarding #31380: feat: rename geo -> geo-writing, rewrite bootstrap for skill-first onboarding
• feat: rewrite wake-up greeting to GEO message for content-automation cohort #31384: feat: rewrite wake-up greeting to GEO message for content-automation cohort

Part of plan: content-automation-geo-skill.md

Closes JARVIS-899

[linear]

JARVIS-899

[alex-nork]

Plan papertrail: Content-Automation GEO Skill Auto-Install

Content-Automation GEO Skill Auto-Install

Overview

Bundle a new GEO writing skill in the assistant, update the content-automation cohort onboarding flow to send a GEO-focused first user message instead of the generic "Wake up, my friend!" greeting, and let the model respond naturally (bypassing the canned greeting) using the updated BOOTSTRAP-CONTENT-AUTOMATION.md and the geo skill's article writing framework. This is a cross-repo feature touching both vellum-assistant and vellum-assistant-platform.

Linear: JARVIS-899

PR 1: Add geo skill as bundled skill

Depends on

None

Branch

geo-skill/pr-1-bundle-geo-skill

Title

feat: add GEO article writing skill as bundled skill

Repo

vellum-assistant

Files

• assistant/src/config/bundled-skills/geo/SKILL.md
• assistant/src/config/bundled-skills/geo/geo-article-writer.md
• assistant/src/config/bundled-skills/geo/geo-vs.md
• assistant/src/config/bundled-skills/geo/head-to-head-structure.md
• assistant/src/config/bundled-skills/geo/listicle-structure.md
• assistant/src/config/bundled-skills/geo/qc-checklist.md

Implementation steps

1. Create the directory assistant/src/config/bundled-skills/geo/.
2. Copy all 6 markdown files from ~/Desktop/geo-skill/ into the new directory.
3. Update the SKILL.md frontmatter to match bundled skill conventions:
   • name: geo (must match directory name)
   • description: keep the existing description about GEO/AEO-optimized article writing
   • Ensure metadata.vellum.display-name is set (e.g. "GEO Article Writer")
   • Add metadata.vellum.activation-hints with entries like "GEO content writing", "SEO article optimization", "AI search engine optimization", "blog post writing for AI citations"
   • No feature-flag field — the skill is prompt-only and safe for all users
4. Verify no TOOLS.json is needed (this is a prompt-only skill with no tool executors).
5. No need to regenerate bundled-tool-registry.ts since there are no tool executors.

Acceptance criteria

• Directory bundled-skills/geo/ exists with all 6 markdown files
• SKILL.md frontmatter passes parseFrontmatter() validation (name, description, metadata fields present)
• bun run lint and bunx tsc --noEmit pass
• Skill appears in the catalog when loadBundledSkills() runs (verify with a targeted test or by inspecting the discovery logic)

---

PR 2: Update content-automation bootstrap for GEO-focused first message

Depends on

None

Branch

geo-skill/pr-2-update-bootstrap-geo

Title

feat: update content-automation bootstrap for GEO article flow

Repo

vellum-assistant

Files

• assistant/src/prompts/templates/BOOTSTRAP-CONTENT-AUTOMATION.md

Implementation steps

1. Read the current BOOTSTRAP-CONTENT-AUTOMATION.md at assistant/src/prompts/templates/BOOTSTRAP-CONTENT-AUTOMATION.md.
2. Update the "First turn" section to handle the new first message "I want to write articles that rank better in GEO":
   • Instead of a generic "turn content into a publishable draft" framing, acknowledge the GEO/AEO intent explicitly.
   • The opening should acknowledge the user's GEO goal and establish that the assistant will help create articles optimized for AI search engines (ChatGPT, Perplexity, Claude, Gemini).
   • Keep the existing logic for checking pre-existing state (website URL in user context → data/sanity-connection.json → data/content-source.json).
3. Update the "First draft" section to reference the geo skill's article formats:
   • The model should use the geo skill's framework (listicle or head-to-head format) for the first draft.
   • The model should follow the geo skill's research phase (competitor research, real data) before writing.
   • The model should apply the geo skill's QC checklist after drafting.
4. Keep the existing VOICE.md mechanics, edit loop, publishing, and scheduled drafting sections largely unchanged — these are format-agnostic.
5. Update the "Constraints" section to reference the geo skill's banned words and formatting rules (no em dashes, no buzzwords, HTML tables only).

Acceptance criteria

• Bootstrap template handles "I want to write articles that rank better in GEO" as the expected first message
• Bootstrap references the geo skill's article formats (listicle, head-to-head) and QC checklist
• Existing pre-chat state detection (Sanity, content source, website URL) is preserved
• VOICE.md mechanics are preserved
• Template is valid markdown with no frontmatter parsing errors

---

PR 3: Send GEO-specific first message for content-automation cohort

Depends on

None

Branch

geo-skill/pr-3-geo-first-message

Title

feat: send GEO-focused first message for content-automation cohort

Repo

vellum-assistant-platform

Files

• web/src/app/(app)/assistant/(chat)/AssistantPageClient.tsx

Implementation steps

1. Read web/src/app/(app)/assistant/(chat)/AssistantPageClient.tsx, focusing on the auto-greet effect (around line 1262-1280).
2. At line 1279, the hardcoded sendMessage("Wake up, my friend!") fires for all cohorts. Change this to be cohort-aware:
   • Read pendingOnboardingContextRef.current?.cohort (the ref is already populated at line 1033 before the auto-greet effect fires).
   • If cohort is "content-automation", send "I want to write articles that rank better in GEO" instead.
   • For all other cohorts (or no cohort), keep "Wake up, my friend!".
3. The exact code change at line 1279:

   const greetMessage =
     pendingOnboardingContextRef.current?.cohort === "content-automation"
       ? "I want to write articles that rank better in GEO"
       : "Wake up, my friend!";
   void sendMessage(greetMessage);

4. The onboarding context (including cohort: "content-automation") is automatically attached to the message by the useSendMessage hook via pendingOnboardingContextRef, so no additional changes are needed to thread cohort information.

Acceptance criteria

• Content-automation cohort users see "I want to write articles that rank better in GEO" as their first message in chat
• All other cohorts still see "Wake up, my friend!" as the first message
• The onboarding context (cohort, tools, tasks, tone, sanityConnected, contentSourceUrl) is still attached to the message payload
• npx tsc --noEmit passes
• The message does NOT match isWakeUpGreeting() on the daemon side (verified by text comparison: "i want to write articles that rank better in geo" !== "wake up, my friend"), which causes the daemon to skip the canned greeting and use normal model inference instead

---

Parallelism analysis

Wave 1 (parallel): PR 1, PR 2, PR 3
Total: 3 PRs across 1 wave (vs 3 sequential)

All three PRs are independent and can be implemented in parallel. The complete feature requires all three to be merged:

• PR 1 provides the geo skill capabilities to the model
• PR 2 ensures the bootstrap template handles the GEO-focused first message
• PR 3 sends the GEO-focused first message from the platform

Cross-repo note: PR 3 targets vellum-assistant-platform. PRs 1 and 2 target vellum-assistant (this repo).

[devin-ai-integration]

Devin Review found 2 potential issues.

View 3 additional findings in Devin Review.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 084ca921b6

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[alex-nork]

Cleaned up in #31464 — removed fabricated directives that were hallucinated during the generated PR.