zhizdev/grow

Overgrow — Grow (Bonemeal Action Queue)

This skill is the integration point between the Bonemeal MCP server and the rest of overgrow. It surfaces the open action queue from the user's Bonemeal project, lets them pick what to build, dispatches each item to the existing spawn-pages or spawn-blogs skill with the action's brief as input, then marks each one completed back on Bonemeal — attributing the completion to "Overgrow" in the Bonemeal UI.

Prerequisites

1. The user must already have authenticated the bonemeal MCP server. If mcp__overgrow__list_projects is unavailable, tell them to run /mcp and authenticate the overgrow server, then re-invoke this skill.
2. Two state files. Create on first run.

State files

State is split into project-scoped (committed to git, shared across collaborators) and developer-scoped (per-machine).

.overgrow/bonemeal.json (in repo root, committed)

{
  "project_id": "<uuid>",
  "project_name": "<for log lines>"
}

Identifies which Bonemeal project this repo represents. Same answer for every developer who clones the repo, so it belongs in version control.

~/.config/overgrow/preferences.json (per-developer, never committed)

{
  "auto_complete": true | false | null
}

Each developer's personal answer to "should /grow auto-mark actions complete in Bonemeal?". null until they answer, then true ("always") or false ("ask each time").

If $XDG_CONFIG_HOME is set, prefer $XDG_CONFIG_HOME/overgrow/preferences.json over ~/.config/.... If running on a system without ~/.config/ (e.g. some Windows setups), fall back to a sensible per-user location and tell the user where you wrote it.

Create either file (and parent directories) on first read if missing. Treat unreadable / malformed JSON as "uninitialized" — re-prompt the user for what's needed and rewrite cleanly.

Flow

1. Resolve the project

Read .overgrow/bonemeal.json. If project_id is null/missing:

• Call mcp__overgrow__list_projects (no arguments).
• If exactly one project is returned, use it; tell the user "Linking this repo to Bonemeal project: {name}".
• If multiple, present a numbered list with name, workspace.name, id, and call AskUserQuestion to ask which one this repo represents.
• Persist project_id and project_name to .overgrow/bonemeal.json.
• Suggest the user commit the file so collaborators don't get re-prompted.

If project_id is set, skip the resolve step. (User can edit the file by hand to relink.)

2. Pull the queue

Call mcp__overgrow__get_project_actions with:

• project: <state.project_id>
• status: "proposed" (default — the open queue)
• action_type: filter based on $ARGUMENTS:
  • "landing" → landing_page
  • "blogs" → blog_post
  • empty → call once with landing_page, once with blog_post, merge results.
• limit: 50 (generous; the daily generator caps lower than this anyway).

If the queue is empty, tell the user "Your Bonemeal queue is clear — no proposed landing pages or blog posts." and stop.

3. Present the queue

Render a numbered list grouped by action_type. Each row: [N] [type icon] {title} — {opportunity_overview}.

opportunity_overview is markdown with bold on the hook; render as-is.

Then call AskUserQuestion asking which to build. Accept these reply forms:

• all — build everything in the list.
• 1, 3, 5 — comma-separated numbers (or ranges like 1-3).
• landing / blogs — only that type from the list.
• skip / cancel — exit without building.

4. Build each chosen action

Process the selected actions in order. For each:

4a. Landing page (action_type === "landing_page")

The action's metadata is shaped:

{
  "target_keywords": ["..."],
  "gap_prompts": ["..."],
  "competing_domains": ["..."]
}

Invoke the spawn-pages skill with a brief built from:

• title (the action title; this is the page intent)
• description and opportunity_overview (rationale and 1-line hook)
• metadata.target_keywords (primary + secondary keywords)
• metadata.gap_prompts (the AI-search prompts the page must answer; pass these into spawn-pages's query-fanout step instead of generating new ones)
• metadata.competing_domains (sites currently winning the keyword; reference for differentiation)

Phrase the dispatch as a natural-language sub-task to spawn-pages so it triggers the existing skill machinery — don't try to call it as a tool.

4b. Blog post (action_type === "blog_post")

The action's metadata is shaped:

{
  "topic": "...",
  "target_keywords": ["..."],
  "gap_prompts": ["..."],
  "competing_content_urls": ["..."]
}

Invoke the spawn-blogs skill with a brief built from:

• title (post title)
• description and opportunity_overview
• metadata.topic (the pillar / cluster anchor)
• metadata.target_keywords
• metadata.gap_prompts (sub-intents to cover)
• metadata.competing_content_urls (for the "what to do differently" pass)

Same dispatch style as 4a.

5. Mark completed

After each successful build (file written and saved):

1. Compute the relative URL of the created page. Examples:

   • app/foo/page.tsx → /foo
   • app/blog/(post)/foo/page.mdx → /blog/foo
   • content/blog/foo.md → /blog/foo

   If the routing convention is ambiguous (custom framework), use the file path the user can git diff to find it; better than no URL.

2. Determine whether to confirm. Read auto_complete from the per-developer preferences file (~/.config/overgrow/preferences.json or $XDG_CONFIG_HOME/overgrow/preferences.json):

   • If auto_complete === true, skip the prompt.
   • If auto_complete === false, prompt once per action: "Mark {title} complete in Bonemeal? [y/n]".
   • If auto_complete is null/missing, prompt once with three options via AskUserQuestion: Yes, No, Always (don't ask again).
     • On Always, write auto_complete: true to the preferences file before calling.
     • On Yes, leave auto_complete unset for this run, but ask again next run (current-run yeses don't change the persistent setting).
     • On No, leave auto_complete unset and skip the call for this action.

3. If confirmed, call mcp__overgrow__mark_action_complete:

   • action_id: <action.id>
   • output_url: <relative URL from step 1> (omit if the URL really can't be derived)

4. If the call returns isError, log the message but continue to the next action — Bonemeal-side failures shouldn't block the local build queue.

6. End-of-run summary

Print:

• Built X / Y selected actions.
• For each successful action: ✓ {title} → {output_url} (and · marked complete in Bonemeal if applicable)
• For each failed action: ✗ {title} — {short reason}
• Z proposed actions still in your queue. Re-run /grow to refresh.

Failure modes

• MCP not authenticated: tell the user to run /mcp. Don't try to skip and rebuild blindly.
• Build failed for an action: do NOT mark complete. Leave it in proposed state so it shows up next run.
• Action already marked complete in Bonemeal between fetch and mark: the MCP tool's UPDATE will no-op and return isError: true; treat as a benign skip.
• State files corrupted or unreadable: treat as uninitialized — re-prompt and rewrite cleanly. The two files are independent; a broken preferences.json doesn't invalidate the project link, and vice versa.

Conventions

• Always tell the user what you're doing (one-line status: "Pulling queue…", "Building landing page X of Y…", "Marked complete in Bonemeal").
• Don't batch silent calls — each MCP call is a remote write and the user should see them happen.
• If $ARGUMENTS is landing or blogs, pass through to the action_type filter and skip the other type entirely.