social-copy

The fast, batchable, on-voice social post copy layer — description + trending hashtags for a finished piece, shaped for TikTok / Reels / Shorts in one invocation. The skill DRAFTS; ralphy unit caption PERSISTS the copy into unit.json.

ALSO FIRE

• The user just rendered / formed a unit and asks "and now? what do I write under it?" (any language).
• The user drops a niche + "give me trending hashtags" with no caption ask — still this skill (hashtag bank + format shaping).
• A batch of reels just finished and the user wants copy for all of them → bulk mode.

DO NOT FIRE

• The user wants aligned-to-VO subtitles / SRT burned into the video — that is ralphy generate captions (a different surface). This skill is the social POST copy.
• The user wants the script / scenario itself — that is the scenarist playbook.
• The user wants to publish to the library — that is templater / dev-publish-template; caption is an input that path can carry.

Hard invariants

1. Target-audience language drives the copy, NOT the chat language. Same rule as intake: if the audience is German, the post is German even if the user is chatting in English / Russian. Offer EN + the audience language. Pass --language <lang> to the verb.
2. Source-ground every copy block. Pull from the unit's title / blurb / tags / provenance, the project brief, and — when present — the reference reel's meta.info.json / transcript / on-screen text. Copy must describe the ACTUAL content, never a generic guess.
3. Match the niche's native voice. Meme / brainrot register for the aura format; professional for a brand ad. The niche is read from the matched #412 content-mode / the unit's tags / provenance, then carried into the draft.
4. Hashtags come from the maintained bank, not invented. The per-niche trending-tag bank lives at cli/lib/social/hashtag-bank.ts (niche spine + format tags + broad-reach, deduped + ordered). Trend tags rot — see the staleness note below.
5. Append-only on the persisted caption. ralphy unit caption never clobbers an existing caption: re-drafting requires --force, which archives the prior into caption_versions. AGENTS.md invariant #14.

Platform shaping (one invocation, all three)

Platform	Shape	Tags
TikTok	one punchy hook line (1 sentence, scroll-stopping)	3-5 inline tags
Reels	a fuller caption (2-4 sentences, on-voice, soft CTA)	10-15 tags
Shorts	a title, ≤ 40 chars	tags optional

The drafter writes the three copy bodies; the hashtag bank supplies the merged tag set (the verb caps it ~18, niche-first → format → broad-reach).

The bulk-mode loop

For N units / a batch, do NOT hand-write each. Run once across all units:

ralphy unit caption <project> --bulk --language English

This drafts + writes one caption block per unit (skipping any that already have one unless --force). Single unit:

ralphy unit caption <project> <slug> --language English --niche aura

Useful flags: --niche <hint> (overrides the auto-detected register / tag spine), --brief "<on-screen text / source caption>" (extra grounding), --force (re-draft, archive prior).

Source grounding

Before drafting, gather:

• The unit's unit.json — title, blurb, tags, provenance (style / template / recipes / assets).
• The project brief / STORYBOARD.md / PRODUCTION_PLAN.md if present.
• The reference reel's own caption + on-screen text (we store the source reel + its meta.info.json / transcript). The copy should echo the actual hook, not a generic line.

The verb folds the unit fields + --brief into the draft automatically; pass --brief when the strongest grounding (the source reel's caption / on-screen text) lives outside unit.json.

The trending-hashtag bank + staleness

cli/lib/social/hashtag-bank.ts maps niche → curated tag spine (aura, analog-horror, ps1-horror, tech-explainer, ugc-ad, unboxing, beauty, fitness, food, saas, general) plus format tags (reel / short / tiktok / carousel / …) and a stable broad-reach set (#fyp etc.).

Trend tags rot. The bank stamps LAST_REVIEWED; the verb's JSON output reports bank_stale: true once it is older than 60 days. When stale (or for a hot niche), refresh BEFORE shipping:

• light path — run the researcher skill / a trend scrape on the niche + platform, read the current top tags, update the bucket in hashtag-bank.ts, bump LAST_REVIEWED.
• manual path — open the niche's top recent posts and copy the tags they rank under.

Broad-reach tags rarely rot; format + niche buckets drift on a ~monthly cadence.

Worked example (one reel → TikTok / Reels / Shorts block)

A finished aura-moment-001 unit (PS1-horror "aura" meme reel, format video, tags ["aura","ps1core","meme"]), audience English:

ralphy unit caption aura-proj aura-moment-001 --language English

Drafts (niche auto-resolves to aura) and writes into unit.json:

{
  "caption": {
    "platform": {
      "tiktok": "POV: your aura hits 999,999 and the room goes PS1",
      "reels": "When the aura is too strong the graphics downgrade to PS1. Caught it on camera before reality re-rendered. Watch till the end. Follow for more aura moments.",
      "shorts": "Aura: 999,999 (PS1 mode)"
    },
    "hashtags": ["#auramoment","#aura","#aurapoints","#ps1core","#brainrot","#meme","#sigma","#edit","#reels","#reelsinstagram","#instareels","#fyp","#foryou","#foryoupage","#viral","#trending"],
    "language": "English",
    "niche": "aura"
  }
}

To also offer the audience-language variant, re-run with --language <lang> --force (the EN block is archived into caption_versions).

CLI cookbook

• Draft + persist one: ralphy unit caption <project> <slug> [--language <lang>] [--niche <hint>] [--brief "<text>"]
• Bulk across the project: ralphy unit caption <project> --bulk [--language <lang>]
• Re-draft (append-only): add --force (prior caption → caption_versions).
• Read it back: ralphy unit show <project> <slug> (the caption field).

Never write copy straight into unit.json by hand — the verb owns the write (logs the LLM draft to generations.jsonl, merges the bank, enforces append-only).