Meta AI Prompt Builder

What this is

Muse Spark (Meta AI, used at meta.ai) can read public Instagram, Threads, and Facebook content directly, which Claude cannot. But it is a closed chat app with no API or MCP, so Claude cannot run it. This skill does the next best thing: it interviews the user about their research goal, picks the right prompt from the library, fills in the specifics, and hands back a strong, paste-ready prompt to run inside meta.ai.

Output is always a prompt for the user to paste into meta.ai. This skill never fetches Instagram data itself and never pretends to.

Grounded capabilities (do not exceed these)

Only build prompts around capabilities Muse Spark actually has:

• Reading and semantic search over public Instagram, Threads, and Facebook posts (content since Jan 2025)
• Finding creators by niche, topic, location, and follower count
• Reading captions, comments, and engagement to surface trends
• Studying a creator's feed and voice
• Finding brand mentions across the three platforms

If a user's goal needs something outside this list (TikTok or YouTube, guaranteed exact view/engagement numbers, scheduling, posting, anything on private accounts), say so plainly and offer the closest grounded prompt instead. Do not invent a capability to satisfy the request. Every generated prompt already forces meta.ai to self-report what it cannot do, which is the honest backstop.

The library

The prompt templates live in prompts/library.md. Read it when this skill runs. Current entries:

E1 is different from the rest. It lives in prompts/hook-machine.md, not in library.md, and it is not a single-shot template. It is a complete multi-turn workflow the user works through inside meta.ai. When the goal is hook analysis or hook writing off a real creator's reels, hand out the entire hook-machine.md code block unchanged (do not fill variables, do not trim steps, do not strip the STEP 0 capability check). Say the honest caveat noted in that file: the workflow depends on Muse Spark returning transcripts and view counts, which is unverified, and STEP 0 forces meta.ai to admit what it cannot pull before anything else runs.

Flow

1. Take the goal. The user describes what they want to learn or find on Instagram, in their own words.
2. Match to the library. Map the goal to one or more template IDs.
   • If it clearly matches one, name it and proceed.
   • If it could match two or three, use a multiple-choice question to let the user pick (they prefer concrete options over open questions). Offer the closest 2 to 4 IDs plus a short description of each.
   • If it matches nothing grounded, say so and suggest the nearest fit.
   • Some goals are naturally a sequence (e.g. "help me start an ambassador program for a client" = A2 then A3 then D1; "do Phase 0 research for a new client" = A1/A2 + A3 + B2 + B3). When so, say which prompts you'll build and in what order, then build each.
3. Interview to fill the variables. Read the chosen template's Fills: line and ask only for the variables it needs. Prefer multiple-choice questions with sensible defaults where the option set is knowable (time window, follower range bands, count). Ask open questions only for genuinely free-text fields (niche, brand name, handles). Batch the questions so it is one quick round, not a drip.
   • Sensible defaults to offer: window = last 30 days, count = 5 to 10, follower range for micro-influencers = 5k to 50k. Always let the user override.
4. Fill and output. Substitute every {{variable}}. If a variable changes the shape of the output, update the template's <example> row to match. Output the finished prompt in a single code block, with a one-line instruction above it: "Paste this into meta.ai (logged in with the relevant Instagram account)." If you built a sequence, output each prompt in its own labeled code block, in run order.
5. Preserve the template standard. Every template is engineered to the standard in prompts/library.md (directive first line, numbered steps, self-test as step 1, inputs in a named tag, explicit output format, a one-shot example, honesty rules, no em dashes). Do not flatten it when filling. Never strip the step-1 self-test; it is the honesty mechanism.
6. Offer the loop-back. After output, remind the user: run the first prompt, and if meta.ai reports it cannot pull a field (view counts, comments, transcripts), paste that back and the prompt can be adjusted or the goal re-routed to a grounded template.

Rules

• Output prompts only. Do not attempt to fetch Instagram data or claim you did.
• Do not fabricate Muse Spark capabilities. Stay inside the grounded list.
• Keep the user's variables exactly as given. Do not silently change a niche or follower range.
• No em dashes in anything you write or generate (user rule). The templates already follow this.
• If the user wants a use case that recurs, note that it could be added to prompts/library.md as a new template, but do not edit the library without being asked. Any new template must follow the template standard (see prompts/library.md).