Suede Social

Suede AI Social Skills

Compatibility

Can work from supplied posts, screenshots, and analytics exports. Live account inspection needs authorized browser/platform access; publishing and scheduling require separately authorized tools.

For a concrete input, expected handoff, and failure check, read the worked handoff.

Turn one account's real evidence into posts, scripts, calendars, and tests that earn follows, leads, and sales, not only views. Every recommendation resolves to the account's own content, audience, offer, voice, and production capacity, or it is labeled an experiment. Publishing and engagement stay behind exact-content approval.

Correct these requests first

• "Post daily and use the best time." Cadence, timing, hashtag count, and format are account-level tests. Read the account's last 30 to 90 days first.
• "Browse my last 30 posts automatically." Use native analytics, an export, an authorized API, or user-supplied links. Do not scrape.
• "This went viral, so it works." Views are attention. Conversion needs follows, profile actions, DMs, leads, or sales evidence.
• "Run the daily workflow" means publish. It means refresh evidence and prepare approval-ready candidates. Publishing is a separate approved step.

Step 1: Build the Account Evidence Pack

Read .agents/product-marketing.md (or .claude/product-marketing.md, or the legacy product-marketing-context.md) plus any brand guide, voice samples, offer sheet, content ledger, and analytics export. Then fill:

textCopy code
Handle and visible identity:        Account type: personal | creator | business
Objective: awareness | qualified followers | leads | sales | community
Primary audience and their problem:
Offer, price, and conversion path:
Voice samples with dates:           Faceless preference and available media:
Cadence now and production capacity:
Timezone and follower-active windows:
Evidence source and coverage:       Attribution: none | native | UTMs | CRM | checkout
Claims or topics needing review:    Last refreshed:
Unknowns that limit confidence:

Ask one compact batch of questions only for gaps that change the work.

Label every fact in the output by evidence class:

• Observed-public: visible post, caption, date, format, views, likes, comments.
• Observed-owned: native analytics, export, DM log, link analytics, CRM, checkout.
• Computed: a formula from observed values, with the formula and denominator shown.
• Inferred: a hypothesis to test, never worded as fact.
• Unknown: the measure is unavailable; say what would resolve it.

Step 2: Pick the mode

Several modes in one request run in dependency order: audit or context refresh, strategy, asset creation, QA, approval packet.

Step 3: Audit before you prescribe

When a handle or export is supplied, audit before recommending anything:

1. Confirm the visible handle and whether evidence is public-only or owned. Never ask for a password or connect a third-party tool as a shortcut.
2. Collect the most recent 30 posts (all, if fewer). Tag boosted posts, collabs, giveaways, and launches as confounders.
3. Code each post: topic, pillar, hook family, format, CTA, offer proximity, production burden, and the search phrase it carries.
4. Compare within like formats using medians; show n for every pattern.
5. Classify each post converts, assists, attention-only, or unknown using the rules in account-audit.md.
6. Return repeatable patterns, dead weight, evidence gaps, and the next 3 to 5 controlled tests, each citing post IDs.

Pillars

Derive 3 to 5 pillars where audience problem, account expertise, observed response, business objective, and a repeatable production source overlap. For each, return:

textCopy code
Pillar | Audience job | Proof the account can own it | Supporting post IDs
Format hypothesis | Business bridge | Stop condition

Do not force equal shares. Weight pillars by evidence and the next thing the account needs to learn. Promotional content usually stays a small share; let the account's own conversion data set the number.

Platform quick reference

Format fit is a hypothesis per account. Facts verified October 2026 are in platform-limits.md, including the Instagram 5-hashtag cap, 3-minute Reels and Shorts, the 3:4 Instagram grid crop, and Google indexing of public Instagram professional accounts.

Hooks

The first line or first two seconds decide whether anyone stays. Families to draw from: a belief the account was wrong about, a specific result with its timeframe, a mistake and its fix, a how-to with the pain removed, a defensible contrarian claim. Full lists live in post-templates.md and short-form-video.md.

Reject any draft that contains:

• Engagement bait with no stake ("Thoughts?", "Agree?", "Who else?").
• Manufactured awe ("Let that sink in", "This changes everything").
• A contrarian claim the account cannot back with a result or named source.
• A hook wrapped around an unproven claim. Fix the claim, not the hook.
• Fake specificity: invented percentages, round results, unmeasured timelines.

Score candidate ideas 0 to 4 on each: audience recognition, specific payoff, evidence strength, voice fit, business bridge. Rank by total, show every component, and call the winner a test candidate, never "viral".

Asset contracts

Short video (Reels, TikTok, Shorts). Return 5 hooks (20 on request), the selected hook with its score, a second-by-second script (visual, voiceover, on-screen text, retention job per beat), a faceless shot plan when asked, a caption whose first line carries the search phrase, one CTA, alt text and caption notes, a rights and claims checklist, and a one-variable test card. With no account evidence, label 15 to 45 seconds a starting test range.

Carousel or document post. Pick the narrative first (value stack, problem-proof, numbered techniques, critique, demo walkthrough). Slide count equals the beats the idea needs. Slide 1 names the audience tension or payoff; each middle slide does one job; the last slide pays off and gives one action.

Text post or thread. First line stands alone before truncation. One idea. One CTA. Use the skeletons in post-templates.md.

Story sequence. Use Stories for a job (research, proof, objection handling, launch, reply, conversion), never a daily quota. Frame table in content-production.md.

Write captions and voiceover from the account's real voice samples: sentence length, vocabulary, humor, punctuation, taboo list. Run final copy through suede-deslop before it goes for approval.

Repurposing

Extract self-contained atoms from one approved source: a quotable claim, a complete mini-story, a tactical tip, a data point, a behind-the-scenes moment. Map each atom to the platform where its format fits, and give each asset its own point. When the source cannot support the requested count without repeating itself, say how many it supports and what extra input would support more.

When the source is a third-party video, or the clip must lead to an article or guide, route to suede-clip-to-guide, which owns rights routes and the short-to-long bridge.

Calendar row schema

Every calendar row carries these fields. A row missing evidence source, test variable, or approval status is not ready:

textCopy code
Date + timezone | platform | format | pillar | audience problem | hook | payoff | CTA
Search phrase | evidence source | one variable under test | primary metric
Asset owner | production status | approval status | readback field

Use the lowest cadence that preserves evidence, voice, rights, and review quality. Daily candidate generation is fine; daily publishing is a hypothesis.

Listening and engagement

Listening surfaces posts worth a reply draft: ICP fit, buying intent, a real chance to add something. Score, draft, and return for approval; every comment is a separate external action. Method, rubric, and source recipes: listening.md and listening-sources-template.md.

Measurement

Pick one primary metric from the objective and 2 to 4 diagnostics. With denominators available, compute rates explicitly:

textCopy code
save rate = saves / accounts reached        share rate = shares / accounts reached
follow rate = follows from post / reached   profile-action rate = profile actions / reached
lead rate = attributed leads / reached      search share = search-sourced views / views

Compare each post with its format's trailing median, not a universal benchmark. Change one variable per test. Diagnosis rules and the experiment card are in measurement.md.

Suede AI account mode

When the account belongs to Suede AI, Jason Colapietro, or a Suede AI product:

• Lead with what Suede AI sells first: AEO, GEO, and SEO services (https://suedeai.agency) and Agent Studio multi-agent automation for businesses (https://suedeai.app). Product lanes such as Suede Sing, StoryBeam, GuitarHub, and the creator skills come after.
• Build proof-led lanes: live demos, before-and-after search visibility with a named source, agent workflows shown working, founder lessons with specifics.
• Bylines and bios use Jason Colapietro. "Johnny Suede" is for social handles and casual creator content only.
• Never claim rankings, AI citations, revenue, or client results the account cannot show.
• Use only docs/assets/suede-ai-logo-transparent.png as the Suede S mark (SHA-256 83a7ee0317e4debe2e7b076c20ba067feb76a587f9e829dc6310ae4be4b44dfa). Never redraw, recolor, or regenerate it. If the file is missing or the checksum differs, omit the mark and halt.

Halt contract

Use this format when access, asset rights, visible identity, or exact-content approval blocks the result:

textCopy code
HALT: <one-line blocker>
Why it blocks: <specific missing authority or evidence>
Resolve with:
1. <option>
2. <option>
3. <option, when useful>
Waiting for: <the exact item or approval>

Keep producing safe drafts and worksheets when they stay useful and do not imply the blocker is resolved.

Boundaries

• Do not publish, schedule, reply, follow, like, repost, message, or edit an account without approval of the exact content and the visible identity.
• Do not scrape, bypass access controls, evade rate limits, or put a consumer password into an automation.
• Do not invent private insights, trends, quotes, sentiment, competitor conversion, testimonials, or results. Public competitor data cannot show saves, shares, retention, or sales.
• Do not run engagement pods, buy followers, auto-like or mass-comment, or post undisclosed synthetic endorsements.
• Do not use music, footage, likenesses, logos, or reposted creator content without a verified lawful basis and any required credit.
• A draft, scheduled item, or composer preview is not a published post. Require a live permalink readback.

Routing

• Paid social campaigns, budgets, and targeting: suede-ads. Paid creative variants: suede-ad-creative.
• One idea that must carry several surfaces with distinct arguments, or an agent-run editorial pipeline: suede-newsroom. This skill decides where and when; that one decides what each asset argues.
• A clip, interview moment, or third-party video bridged to an article, newsletter, or guide: suede-clip-to-guide.
• Rendering, editing, captions tooling, and exports: suede-video. Still images and graphics: suede-image.
• Editorial portfolio and topic clusters: suede-content-strategy.
• UTMs, events, and attribution repair: suede-analytics.
• Site-level SEO and AI answer visibility: suede-seo-audit and suede-ai-seo.
• Final anti-slop pass on anything public: suede-deslop.