Instagram content

Produce Instagram content that a person would actually stop for, check it with the bundled scorers, and publish only after the user approves the exact post. Story comes first: platform tactics amplify a specific, true story and never replace one.

Before writing

1. Read the product/brand context if present: .agents/product-marketing.md, .claude/product-marketing.md, or product-marketing-context.md. Ask only for what is missing.
2. Read a voice file if present (.agents/instagram-voice.md or ~/.claude/instagram/voice.md). If none exists and the account is the user's own, pull their recent captions: npx -y bun --no-env-file ../instagram/scripts/main.ts posts --limit 12 --json. Infer the voice: sentence length, vocabulary, emoji use, the CTA they repeat. Offer to save it as .agents/instagram-voice.md.
3. If the idea is thin, ask one batched question: what happened, to whom, and what it cost or returned. Do not pad a thin idea.
4. If a swipe file from instagram-research exists, use its winning hook formulas over the defaults.

Choose the format

One idea per post. If a draft holds two ideas, split it and say so. Format playbooks, safe zones and slide/beat structures are in references/formats.md.

Scripts (offline, no network)

Run from this skill directory with npx -y bun:

npx -y bun scripts/caption-lint.ts caption.txt [--keywords "term one,term two"] [--json]
npx -y bun scripts/hook-score.ts hooks.txt            # rank hook options, one per line
npx -y bun scripts/hook-score.ts --hook "one line"    # score a single hook
npx -y bun scripts/beats.ts script.txt --target 30    # timed beat sheet for a Reel script
npx -y bun scripts/ai-tells.ts detect draft.txt       # AI-tell score and flagged phrases
npx -y bun scripts/ai-tells.ts humanize draft.txt     # mechanical cleanup (add --report for what changed)

• caption-lint prints what the feed shows before "... more" (~125 chars) and checks length, first line, concreteness, hashtags (cap 5), tag placement, dead links, one ask, emoji density and search terms. Exit 0 means READY.
• hook-score scores a hook on mechanical features (length, specificity, stakes, front-loading, direct address). Use it to kill weak hooks, not to crown winners. Its author found it weak at predicting hits. The 26 hook formulas to write from are in scripts/hooks.json (reference data; the scorer does not read it).
• beats times a script at 165 wpm and flags slow hooks, long beats, abstract runs and a missing loop.
• ai-tells uses the phrase and structure lists in scripts/slop.json. humanize deletes and replaces mechanically (invisible characters, em dashes, curly quotes, stock phrases), which can leave broken punctuation behind. Always re-read its output and fix it by hand. Structural tells are only flagged, never rewritten.

Workflow per post

1. Draft 3 hooks from different formulas. Score them and show all three, word for word, with their scores, so the user can pick a different one than the scorer did. Score a hook only once its numbers are filled in; a {{your number}} placeholder inflates the score.
2. Write the body in the user's voice: contractions, short lines, numbers over adjectives, no invented facts. A missing number becomes {{your number}}, and say so.
3. Reels: write the on-screen text as a separate script, then run beats. Carousels: give a numbered slide list the user can edit in ten seconds before rendering.
4. Write the caption, for every post including a Reel script, because the caption is what ships with it. The hook goes in line one, there is one ask (save, share, comment or follow, never "none"), at most 5 topical hashtags below the fold, and no links in the body. Run caption-lint and ai-tells detect. Fix and re-run until clean, or explain the remaining warnings.
5. Show the final package: media plan, on-screen text, caption, lint verdicts.
6. Publishing: once media files exist, run the instagram skill's dry run, for example npx -y bun --no-env-file ../instagram/scripts/main.ts post carousel s1.jpg s2.jpg ... --caption-file caption.txt. Show its output, wait for an explicit yes for that exact post, then re-run with --confirm --account <username> (add --experimental for a write type not yet verified live; status lists them). Never publish without that yes, and never batch-approve several posts.

Writing rules (no AI tells)

Never write:

• Contrast reveals ("It's not X, it's Y"), negation lists ("No X, no Y, no Z"), trailing pile-ons.
• Self-answered questions and colon reveals ("The result? 3x faster.").
• Stock phrases: "Say goodbye to", "Unlock the power of", "Take it to the next level", "Here's the thing".
• Em dashes in posts. Use a period or a line break.
• Broetry (one sentence per line for the whole post), engagement bait ("Agree?", "Read that again"), manufactured vulnerability, emoji as bullets.

Keep fragments and lists of three to one per post. ai-tells detect catches most of these mechanically.

Hooks and post templates

Hook families: curiosity ("I was wrong about [belief]"), story ("Last week, [unexpected thing]"), value ("How to [outcome] without [pain]"), contrarian ("[Common advice] cost us [result]"). Fill each with a specific fact. A formula is only a shape. More templates are in references/post-templates.md. Reel hook and script patterns are in references/short-form-video.md. Carousel narrative frameworks are in references/carousel-frameworks.md.

Planning and repurposing

• Pillars: 3 to 5 pillars tied to expertise and audience questions (for example insights 30%, behind the scenes 25%, education 25%, personal 15%, promotion 5%).
• Cadence: 3 to 5 feed posts a week (Reels and carousels first) plus Stories most days. Consistency beats volume. For the account's real best formats and slots, use instagram-audit.
• Weekly plan: a table of day, format, pillar, hook, and status (draft, approved, posted). Save it as instagram-plan.md in the project if the user wants it kept.
• Repurposing: extract content atoms from long-form (a quotable line becomes a Reel, a story arc a Reel, a framework a carousel, a stat a photo or carousel, behind the scenes a Story). Every post must stand alone without the source.

Engagement on own posts

To triage comments, read them with npx -y bun --no-env-file ../instagram/scripts/main.ts comments <post> --json. Sort them into questions, objections, praise and spam, and draft a reply for each worth answering. Replies go out one at a time, each text approved by the user, with npx -y bun --no-env-file ../instagram/scripts/main.ts reply <post> <commentId> --text "..." --confirm --account <username>. No bulk replies, likes, follows or DMs.

Additional resources

• references/formats.md: Reels, carousel and Stories structures, safe zones, specs.
• references/carousel-frameworks.md: five slide-by-slide carousel architectures.
• references/short-form-video.md: video hook formulas and scripting templates.
• references/post-templates.md: post templates and hook bank.
• scripts/hooks.json: 26 hook formulas (reference). scripts/slop.json: the AI-tell lexicon used by ai-tells.ts.

Credits: adapted from marketingskills social (MIT, Corey Haines) and instagram-agent-skill (MIT, Jake Schincariol); see the plugin's THIRD_PARTY_NOTICES.md.