content-month

30-day Instagram content engine. Produces ./growth/02-content/.

When to invoke

• User asks: "generate a content calendar", "write a month of Instagram", "build a launch campaign".
• Invoked by /growth orchestrator as phase 2 of 3.
• Standalone via /growth-content.

Prerequisites

This skill assumes market-validate has already run OR equivalent context exists:

• ./growth/references/icp.md — required.
• ./growth/references/positioning.md — required.
• ./growth/references/competitors.md — required.

If any are missing, ABORT and prompt user to run /growth-market first (or /growth for full pipeline).

Three sub-phases

Phase A: Voice ingestion (first run per project)

Output: ./growth/references/voice.md

If voice.md already exists, skip this phase (or run with --reset-voice flag to redo).

Interactive flow (default)

1. Auto-extract from repo + live URL:
   • All README files, hero copy, in-repo marketing files, blog posts if any.
   • If live URL was fetched in research phase, use that copy too.
   • Synthesize a draft voice profile: tone adjectives, sentence rhythm, lexical register, signature phrases, do/don't list.
2. Ask user for samples:
   • AskUserQuestion: "Paste 3-5 sentences that nail your voice. (Bullet list ok.)"
3. Ask user for taboos:
   • AskUserQuestion: "What tones to AVOID? (e.g. 'no Duolingo bro-marketing, no fake urgency, no hustle-porn aesthetic')"
4. Synthesize + confirm:
   • Write voice.md with: tone profile, do/don't list, signature phrases, sample-sentence canon, taboos list.
   • Also write a paired visual_style section: palette, type feel, photo treatment, recurring motifs allowed, forbidden visuals. Source this from the codebase design tokens detected in the research phase.
   • AskUserQuestion: "Voice profile drafted at references/voice.md — confirm / edit / regenerate?"

Under --auto mode

Skip steps 2, 3, and 4's confirmation prompt. Execute:

1. Auto-extract voice profile from repo + live URL (same as step 1 above).
2. Skip the sample-paste prompt — use ONLY auto-extracted signal.
3. Skip the taboos prompt — apply the conservative default taboo set:
   • No Duolingo / bro-marketer hype
   • No fake urgency ("Only X left!")
   • No emoji-spam openings
   • No AI-tell phrases ("In today's fast-paced world", "Let's dive in", "Game-changer", "Level up", "Unlock")
   • No performative hustle / struggle valorization
   • No bait-and-switch hooks
4. Write voice.md with these YAML headers prepended:

   ---
   auto_mode: true
   confirmation: pending
   generated_at: <ISO-timestamp>
   needs_review: true
   review_note: "Voice profile auto-extracted in --auto mode without user sample-paste or taboo confirmation. Refine and re-run /growth-content --reset-voice for personalized regeneration."
   ---

The downstream content-week-writer agents read these headers and add a confirmation: auto_mode_pending flag to each piece's secondary_tags[], surfacing in the final report.md for user review.

Phase B: Calendar skeleton

Output: ./growth/02-content/calendar.json (initial skeleton, fleshed out in Phase C)

1. Ask user language(s) if not already set:
   • Interactive: AskUserQuestion with options he, ar, en, multi.
   • --auto: use the language(s) inherited from the orchestrator's auto-detection (passed via --lang). Never prompt under --auto.
2. Build the day-by-day skeleton. For each day in --days N (default 30):
   • Compute weekday based on current date.
   • Assign theme from the weekday × theme map:
     • Mon → Teaching/data (trust)
     • Tue → Pain/story (engagement)
     • Wed → Behind-the-scenes (trust)
     • Thu → Social proof (trust)
     • Fri → Trick/hack (trust)
     • Sat → Founder/personal (engagement)
     • Sun → Product+CTA (product)
   • Assign bucket per the theme.
   • Enforce 60/30/10 ratio across the month: count buckets and rebalance by selectively reassigning theme on overflow days. Trust days = 60% target. Engagement = 30%. Product = 10%.
   • For each piece type that day (3 stories, 1 carousel, 1 post, 1 reel), assign a primary pattern from ${CLAUDE_PLUGIN_ROOT}/references/virality.md § "Weekday × pattern preferences".
3. Write initial calendar.json with all days + pieces but with empty file pointers until Phase C completes.

Phase C: Parallel content fan-out

For each of 4 weeks (1-7, 8-14, 15-21, 22-N):

Spawn a content-week-writer agent (${CLAUDE_PLUGIN_ROOT}/agents/content-week-writer.md) with these inputs in the prompt:

• Week number + day range.
• Theme + bucket + pattern assignments for each day (from Phase B skeleton).
• Path to ./growth/references/icp.md, voice.md, positioning.md, competitors.md.
• Path to ${CLAUDE_PLUGIN_ROOT}/references/virality.md, image-prompt-cookbook.md, output-schemas.md.
• Output path: ./growth/02-content/per-day/.
• Language(s) from Phase B.

The 4 agents run in parallel (single message, 4 Agent calls).

Each agent generates day-NN.json files conforming to output-schemas.md §3 for its 7 days × 6 pieces = 42 pieces.

Coherence guarantees:

• All 4 agents read the same references/ lib, so ICP + voice + virality stay consistent.
• After fan-out completes, this skill runs a deduplication + narrative coherence pass:
  • Read all 30 day-NN.json files.
  • Detect hook duplication (cosine similarity > 0.85 → regenerate later one).
  • Detect repeated image_description motifs (same scene > 3x in a month → regenerate).
  • Verify bucket ratio holds (60/30/10 ± 5%).
  • Verify every piece has virality_score >= 6. Anything below → flag for regeneration; if more than 5% are below, regenerate the failing ones in a final pass.

Phase D: Assemble + render

1. Update calendar.json with final pattern tags, virality scores, and file pointers.
2. Write calendar.md — human-readable day-by-day summary table.
3. Render dashboard.html — codebase-styled overview:
   • Header with campaign meta (product, days, languages, targets 2500 users / 1M impressions).
   • Grid of 30 day-cards, color-coded by bucket.
   • Click/hover-expand to see the 6 pieces' hooks + virality score.
   • Bucket ratio donut chart (rendered as inline SVG, no external libs).
4. Write report.md — summary stats: bucket ratio, virality score distribution, top 5 highest-scoring pieces, lowest 5 with notes.

Per-piece content rules

Anchored on:

• voice.md — tone, register, signature phrases.
• virality.md — pattern grammar + scoring rubric.
• image-prompt-cookbook.md — two image fields per piece.
• icp.md — audience, pains, gains.
• positioning.md — value prop to weave into product+CTA pieces.

Hashtag rules

• 10-20 hashtags per piece.
• Mix: 30% large (>500k posts), 50% medium (50k-500k), 20% niche (<50k).
• Always include: at least 1 brand/campaign hashtag (e.g. #<product-name>), 2 audience hashtags (e.g. #cs-bagrut-students), 3 topic hashtags.
• Language-aware: Hebrew hashtags for Hebrew content, etc.

Caption SEO

• Open with the hook (no preamble).
• Embed primary keyword + 2-3 long-tail variants in the first 125 characters (the visible-before-"more" range).
• End with the CTA.
• For Instagram alt-text (not the caption but the IG accessibility field): emit personalization_slots: ["{alt_text}"] with a draft alt-text in image_description.

Carousel rules

• 7-10 slides.
• Slide 1 = hook cover (visual + 1-line hook).
• Slides 2-N-1 = body, one idea per slide.
• Slide N = CTA / save-prompt.
• Every slide gets its own image_description + model-specific prompts.

Reel rules

• Duration 15-60 sec target (specify duration_seconds).
• Shot list with 4-8 cuts, timestamped.
• On-screen text mirrors the hook in slide 1 / shot 1.
• No music/audio specification (explicit user requirement).

Story rules (3 per day)

• 3 stories form a mini-arc: hook → reveal → CTA (or hook → poll → CTA).
• Each story may have a sticker_prompt (poll/question/quiz/slider/none).
• Use stickers to drive engagement-loop reach.

Forbidden in output

Strictly enforce:

• No audio / music_suggestion fields anywhere.
• No posting_time / scheduled_at fields.
• No AI-tell phrases (see virality.md § Anti-patterns).
• No Duolingo/neon/bro-marketing language UNLESS voice.md explicitly authorizes (it won't, per typical taboo lists).

Flags

• --auto — unattended mode. Voice ingestion uses auto-extract only (no sample-paste, no taboo prompt, no confirmation). Applies default conservative taboo set. Writes voice.md with auto_mode: true + confirmation: pending headers. No language prompt (uses inherited --lang).
• --days N — campaign length (default 30).
• --lang he,ar,en — output language(s). Multi-lang = parallel sets per language.
• --reset-voice — redo Phase A voice ingestion.
• --reset-calendar — redo Phase B skeleton.
• --target-users N (default 2500) — used to tune CTA aggressiveness.
• --target-impressions N (default 1000000) — used to tune virality-score floor.

Verification

After running:

• calendar.json lists --days × 6 pieces, each with virality_score >= 6.
• All per-day JSON files validate against schema.
• Bucket ratio is within ±5% of 60/30/10.
• No piece has audio or posting_time fields.
• dashboard.html renders without external network requests.
• Spot-check 3 random pieces: hook + caption + hashtags + CTA + 2 image fields all present and non-empty.