weekly-plan-draft

name	weekly-plan-draft
description	Generate caption + hashtags + media plan + product_id for one schedule slot. Subagent-friendly: orchestrator spawns one per slot in parallel.

Common rules: docs/OPERATING_RULES.md.

Input

• slot: one row from weekly-plan-schedule.schedule[] (slot_id, weekday, time_local, platform, mode, theme; product_id arrives as null and this skill fills it).
• available_products: full products[] from active brand YAML.
• topic_research_rows: subset of topic_research_rows filtered to this slot's platform.
• brand_yaml: the active brand profile (full).
• reference: optional — fetch-reference output if the slot is news/event-driven.

Output (JSON)

{
  "slot_id": "2026-W19-ig-mon-1230",
  "platform": "instagram",
  "mode": "post",
  "product_id": "md-905",
  "caption": "...",
  "hashtags": ["#...", "#..."],
  "media_path": "media/assets/<brand>/<...>.jpg",
  "reference_urls": ["https://..."],
  "draft_id": "weekly-2026-W19-ig-mon-1230",
  "rationale": "one or two sentences on product/tone/hook choice grounded in research_rows"
}

draft_id is synthesized from slot_id so /publish-from-plan can write a normal draft JSON for tracing.

Behavior

1. Load the matching draft-<platform> SKILL.md mentally — it sets the rules for length, hashtag count, link policy. Follow them.

2. Pick product_id from available_products:

   • For each product, score slot.theme against product.suggested_themes (substring or stem match counts). Pick the highest-scoring product.
   • If two products tie or slot.theme is generic, pick the product with fewer slots already assigned this week (alternation tie-break — orchestrator passes a weekly_assignments_so_far count if available; otherwise pick alphabetically).
   • If available_products is empty, set product_id: null and proceed.

3. Filter research for SKU:

   • From topic_research_rows, keep rows where relevant_to == product_id or relevant_to == "both".
   • Use these to abstract patterns (tone, opening hook style, layout) — do not lift sentences.

4. Pick media_path:

   • Look in media/assets/<brand>/<product_id>/ for an image / video matching mode.
   • If nothing fits → fall back to media/assets/<brand>/ library roots.
   • If still nothing → return media_path: null and add to rationale "needs media (orchestrator: route to image-generator)".

5. Honor voice.forbidden_words and voice.forbidden_behavior from the brand YAML.

Don'ts

• Do not call publish-* or image-generator directly — the orchestrator does that on the back of media_path: null.
• Do not write files — return JSON only.
• Do not duplicate captions across slots — vary hooks even when the theme is identical.
• Do not invent engagement metrics or research data. Use what topic_research_rows provides; if a slot's platform has no research data, draft conservatively from brand YAML alone and say so in rationale.
• Do not pick a product_id not present in available_products.