Prompts and agent skills I actually use, with the rules that make them work.
-
Trendwatch — a self-bootstrapping agent skill for organic TikTok/Reels growth: mines competitor accounts (including shadow accounts and partnered influencers), decomposes viral videos on 7 axes (thumb-stop ≠ hook), filters out celebrity-tier placements that don't transfer, generates adapted hook ideas mapped to your production stack, and compounds learnings in an experiments log — winners get variant-iterated (message locked, creator/setting varied), losers get killed fast. Drop the folder into your agent's skills directory; on first run it interviews you for 5 strategic questions and auto-discovers the rest.
-
Viral Content Factory — end-to-end generation of TikTok/IG carousels and mannequin outfit-change videos with AI characters: character archetypes that work (unlikely expert, age-defying, pet judge), Gemini prompt templates with anti-artifact constraints, the anchor-first technique for consistent backgrounds across frames, Kling 3 animation via fal.ai, and a 50+ entry viral hook library organized by mechanic (bait & switch, revenge energy, transformation, save-worthy).
-
Voice Realism — voice is the #1 tell that a video is AI-generated; this skill rewrites any video/TTS prompt with spoken lines so the voice survives scrutiny. It bundles every line as emotion + action + dialogue, auto-picks delivery from scene context (whisper is the cheat code — synthesis artifacts vanish in whispered speech), fixes per-model dialogue syntax (Veo colon, Seedance quotes, ElevenLabs audio tags), checks words-per-second against clip length, bakes acoustics and lipsync-friendly framing into the prompt, and outputs a scene-matched post-processing checklist (cut the 2–4 kHz AI ring, 5–10% room reverb, −50 dB noise bed, barely-there voice morphing).
-
Photo Realism — the image counterpart to voice-realism: rewrite any image prompt (Nano Banana, GPT Image, Flux, Seedream, Midjourney, Soul) so the result looks like a real phone photo a friend took, and still looks good. Includes two before/after examples, both generated with nano-banana-pro. The rules that make it work: drop the words that ask for a render (photorealistic, 8k, cinematic lighting, bokeh) and describe how the photo was taken instead, with the friend's phone at their eye level tilted slightly down (light foreshortening, lots of floor in frame); swap stock-cliché ideas for a specific ordinary moment; fill the place with real details and people living their own lives; give the person an action or an attitude instead of a model stare; use one named light; ask for the iPhone look (crisp everywhere, punchy HDR, a little over-sharpened) instead of film grain; state what must stay visible (whole outfit, product label); end with a hard negative clause. It aims for real and beautiful, because "too real" (grey, messy, awkward) fails as surely as "too clean". Ships a prompt template, a drop/use word table, a QA table of AI tells with fixes, reference and edit rules, and per-model notes.
-
Almost Paid — find the users who got within one tap of paying (abandoned checkout, declined card, hit the free limit mid-task, lapsed trial, repeat paywall), price what they're worth, and turn them into a one-screen dashboard with a matching offer per segment. The rules that make it work: validate the trigger events before building on them (paywall events lie), six segments not one blast, friction fixes before discounts, price the discount's cost on people who'd have paid anyway, a 24h cool-off so organic conversions don't get credited to the campaign, and a hold-out on every send. It never sends — it hands a CSV + drafts to your own sender behind a --dry-run / --confirm <count> gate.
-
Carousel Conveyor — the production half of the content system (trendwatch decides WHAT, this produces and schedules it): stand up a recurring AI persona for ~$0.60, then output unlimited carousels at $0. Key ideas: one locked reference selfie → ~6 cover photos (never edit an edit — one hop only), all text composited in PIL not generated, the battle-tested UGC-realism prompt block that kills the "AI gloss" (amateur camera artifacts + unretouched skin + hard negative clause), and a mandatory human gate before anything auto-schedules.
-
Carousel Benchmark — find the TikTok photo-mode carousels that actually go viral in a niche and the accounts behind them, ranked by organic quality instead of views. The rules that make it work: only TikTok's Top search returns slideshows (the /video section and hashtags return zero), profile-scan candidates for medians instead of trusting one search hit, never guess a handle from an app name, and treat a like-rate under 1% on a 50k+ post as bought reach (organic median in our sample was 5%). Ships two tools — carousel_scan.py (Apify, marks is_slideshow, downloads slides + contact sheets) and the offline carousel_leaderboard.py (medians, save/like-rate, boost flag, cadence, markdown table) — plus the cross-niche format table ("help me choose", numbered lists with the product on the last slide, Notes-card app recs, icon roundups).
-
UGC Hooks — the video counterpart to carousel-benchmark, research-to-posting in two days: find the organic app-promo UGC formats going viral right now (1–3 s emotional hook of a person → plain screen recording), read the hook line off the first frames with OCR, sort ~250 winners into 15 hook formats, then turn your own takes into a batch of hook + demo videos. The rules that make it work: search by app/niche phrases and extract the hook afterwards, last 6 months only (hooks get cloned across apps within weeks), take the sound from the viral original but skip voiceovers and put its beat drop on the cut, step captions for multi-screen demos, a frame-by-frame privacy gate on anything recorded from your own phone, and a posting kit with the 2026 rules (Your-brand disclosure, AI label for try-on avatars, 5-hashtag cap) and anti-spam checks (unique captions and pinned comments, rotating hashtags, no links in comments).
-
Influencer Outreach — the paid half of creator growth: scrape creators by niche hashtags, verify their real reach (pinned videos inflate scraped medians up to 28×, and the views/followers ratio does not catch it), score who to write first by expected views per dollar × will they reply × will they accept, cold-email in the creator's language asking for their rate in the first message, then contract and pay for views that actually happened (base upfront + per-25k bonus, counted on a locked day). Includes the two-species model that decides everything — production sellers reply 7× more often but sell off a $150–1000 rate card, reach creators barely answer but find view-based pricing fair — the CRM schema, and a human-only send gate enforced in code.
-
Stop Slop — rewrite any English text so it stops reading like an LLM wrote it: posts, emails, cover letters, product copy, READMEs, abstracts, essays, fiction, toasts. Instead of a banned-word list it works from the eight causes behind the patterns (most-probable phrasing, rater-pleasing structure, hedging to avoid being wrong, no stake in what matters, absorbed genre templates, sentence-by-sentence generation, repetition avoidance, the assistant persona), because deleting a surface pattern just grows a sibling: ban "delve" and you get "dig into", ban em dashes and you get a colon in every line. The rules that make it work: rewrite from the meaning instead of line-editing (line edits keep the skeleton), never invent facts to sound human (vague claims become [placeholders] the text survives without), density is the tell rather than presence, and a don't-overcorrect list against staccato fragments and fake typos. Ships a 45-pattern catalog with cause, fix and before/after pairs, a word list with plain replacements, 16 whole-text rewrites across genres, and a voice-matching mode: give it a sample of someone's writing and it builds a voice profile and writes to it without borrowing the sample's content.
Universal, composable CLIs the skills call on demand — collect, analyze, generate, distribute. Every tool is single-purpose, takes JSON/flags, and emits a uniform JSON envelope on stdout, so they chain together (e.g. tiktok_account.py → account_stats.py → engagement.py). The offline analytics core (engagement, account-stats) needs no keys or network. Full registry with flags and canonical chains: tools/README.md.
