AI Engine Visibility (GEO)
Find out whether ChatGPT, Gemini or Perplexity cite your shop, and fix what is missing.
Prompts are everywhere. Tested prompts, not so much.
Our pick: the most useful ones to start with.
457 results
Find out whether ChatGPT, Gemini or Perplexity cite your shop, and fix what is missing.
A web synthesis where every claim comes from a page actually read.
Reviews a git diff and writes a severity-ranked review.md in French with the tests to write.
Static HTML checks then a report ranked by RGAA and WCAG 2.2 criterion, with fixes.
Reads a crawl CSV export and outputs a prioritized report: titles, canonicals, redirects, hreflang, orphans.
Quality check, pivot table, accessible SVG chart and written summary from a spreadsheet.
Finds secrets, injections, dependencies to audit and over-broad agent permissions, and reports the fixes.
State and handoff files kept up to date, with clean resume after compaction via Claude Code hooks.
Sorts your emails, sums up each thread in three lines and drafts replies, without sending anything.
Drafts your tracking, return and refund replies based on your policy.
Computes plate cost, margin, menu ranking, target stock and price-rise simulations from CSV files and a Python script.
Sorts requests by urgency and prepares work orders, with no automated decisions.
A comparison table and positioning ideas, with the list of pages read.
Turn a long video into 5 to 10 vertical, subtitled, silence-free clips plus chapters for the long version, free and local.
Build a reference sheet, then a 6-shot sequence where the character keeps the same face, using Gemini Omni Flash or Runway references.
A ready-to-use Remotion project that turns a CSV file into an animated vertical video: bar chart race or monthly summary.
Dub a video into several languages with ElevenLabs, after first getting written permission from the person whose voice is reproduced.
A Remotion project that animates the words of an SRT file in sync with the voice, on a solid background or as transparent overlay titles.
Three prompts to adapt a lesson to several levels, comment on work against a rubric and read class results.
A free prompt that turns meeting notes into decisions, dated actions and a weekly update.
An educational prompt to build a weekly 13-week cash forecast with three scenarios.
An educational prompt to spot biases in a trade journal and data leaks in a backtest.
A read-only connector to the official Stripe MCP with a Skill that writes the weekly report.
A connector to the official HubSpot MCP and a Skill that spots stalled deals and drafts follow-ups.
A connector to a Slack MCP server and a Skill that summarizes chosen channels, read-only.
An n8n workflow that scores a lead, adds it to a spreadsheet CRM and drafts a reply.
An n8n workflow that reads invoice PDFs from email, extracts fields with an AI model and checks them before filing in a spreadsheet.
An n8n workflow that categorizes support tickets, sets priority, alerts the team and drafts a reply.
Generate 10 different openings for the same video, compare them on a contact sheet and track results to keep the best.
A read-only connector to the official Linear MCP with a Skill that writes the weekly ticket review.
A connector to the official Zapier MCP and a Skill requiring a confirmed action card before every call.
Start from one SRT and get 5 checked SRT tracks and 5 burned-in videos, with no paid service.
Turn your slide images and a synthetic voice-over into a timed MP4 with subtitles and chapters.
One shot per room with camera movement, a mandatory notice on any retouched image, and a list of flaws that must stay visible.
One source video becomes four files in 1:1, 4:5, 9:16 and 16:9, with a safe-zone preview, without cutting content.
A style reference board attached to every shot, plus a numeric check for light and colour drift, for a consistent clip series.
Writes a multi-stage Dockerfile and .dockerignore, then checks them against security rules (pinned images, non-root, no secrets).
Writes a GitHub Actions workflow (tests, build) with caching, minimal permissions and pinned actions, then lints it.
Cleans a CSV (whitespace, duplicates, numbers, dates, broken rows) and returns a clean file with a counted log.
Replies that cite the article used, and a human handoff when the base falls short.
Replies on hours and free slots for a practice, with no medical advice.
Groups your customers' requests into themes, with requester counts and a confidence level.
An n8n workflow that prepares two spaced reminder drafts, with consent checked and buyers and unsubscribed contacts excluded.
An n8n workflow that prepares a weekly reminder draft listing each open file's missing documents.
Plans an expand/contract migration, writes the files and lints them; batched backfill tested on SQLite.
Finds hard-coded strings in JSX/TSX and produces a messages file with keys, plus a report of doubtful cases.
Prepares a Shopify product CSV and validates it offline before import.
Writes tests that fail first, makes the code pass, then checks test strength with mutations.
Prepares one identical interview plan for all candidates, with a job-related scoring grid and no decisions.
Prepares the onboarding checklist and week-one plan with owners and deadlines, without sending anything or creating access.
Summarises terms and conditions in plain language with points to watch, without legal advice.
Produces and updates a lodging welcome guide in several languages from a single source file.
Writes the script and rules for a salon voice agent: AI disclosure, human handoff, no pressure selling.
An educational prompt to draft a one-page AI use charter and a three-module training plan.
An educational prompt comparing last-click, first-click and linear attribution and flagging data weaknesses.
A prompt to choose and rank three AI pilot projects with a success metric, stop criterion and 30-day plan.
A folder of product photos becomes a folder of transparent PNGs, locally, with a commercially usable model.
Virtually furnish a room with AI, with a disclosure and the original photo beside it, checked by script.
A dated model, license, threshold table with a filter by use (personal, commercial, client).
A read-only diagnostic and a step-by-step guide to run a first local language model.
A 50-question set and a scoring script: recall, rank, expected keywords and refusal when no answer exists.
A prompt to draft easy-to-read text, a script checking sentence and word length, and the reminder that validation belongs to the people concerned.
Cache the stable prefix of a prompt and compute the savings in advance with a script.
Guides Claude through a read-only site tour via the accessibility tree and checks the report with a script.
Requires running verification commands and reading their output before any 'done' claim, with a checked evidence log.
Tested Node.js and Python templates: HMAC signature, timestamp tolerance, idempotency table, fast 2xx response.
Queries a local CSV or Parquet file with read-only SQL using DuckDB, or sqlite3 as a CSV fallback.
Spots the clauses of an incoming confidentiality agreement and rates it green, yellow or red. Not legal advice.
Runbook template and checker script: verified steps, rollback, success criteria and contacts.
Answers questions about your files with a local Ollama model, nothing sent online, with verified citations.
What must be true for your idea to work, and how to test it cheaply.
Prepares the document checklist, timeline, itemised inventory and a draft claim letter, never exaggerating.
Finds swallowed or hidden errors in a repository and ranks them by severity, without touching the code.
Compares documentation with code and lists the pages to update, changing nothing.
Drafts a quick reply to a new sales contact within minutes, using only the facts in your offer sheet.
An educational prompt to write the notice telling visitors they are talking to an AI, and list the points to check.
A prompt that reviews a stalled sales deal and proposes dated actions.
An educational prompt that compares the bank statement and the books and ranks each gap by likely cause.
An educational prompt that keeps only material variances and proposes a cause to confirm and an action.
An educational prompt that shows day by day whether cash covers payroll in three cases, with no promise.
A prompt that turns a spoken explanation into a procedure with steps, RACI and control points.
A prompt that audits the content and search intent of a single page and prioritizes on-page fixes.
A read-only connector to the official Qonto MCP with a Skill that produces the cash position and the month's transactions.
A connector to the official data.gouv.fr MCP (MIT, no key) and a Skill that finds a dataset and cites producer, licence and date.
A Skill and script that query the free public company search API to produce a company profile and its officers.
A read-only connector to the PostHog MCP with a Skill answering on funnels, retention and active feature flags.
A ready-to-run promptfoo config with no paid calls (echo provider) and a validation script.
An n8n workflow that drafts each morning a reminder of tax and social deadlines for the next 14 days.
An n8n workflow that drafts a WhatsApp reply and sends it only after your click, within the 24-hour window.
An n8n workflow that drafts each Monday an email to yourself with sales, average basket and returns for the week.
Turn 4 to 6 screenshots into a 60-second tutorial with zoom, arrow and caption, using one free script.
Build a lyric-video MP4 whose lines appear at the exact times of an LRC or SRT file, with a progress bar, for free.
A Markdown article becomes a narrated vertical video: one card per sentence, synthetic voice, SRT and chapters, for free.
A method, a prompt and scripts to write a video's descriptions, place them in the silences and produce a WebVTT or a mixed track.
Turn a 16:9 video into 9:16 with a frame that follows the subject (simple automatic mode) or your keyframes (manual mode).
Put a video in square format with large burned-in captions readable without sound, and check it against LinkedIn rules.
From a list of timestamps, cut 5 numbered, titled 15-second teasers that end on a cliffhanger.
A library of original camera prompts to animate a photo with image-to-video, plus a free ffmpeg version to check the framing.
A script makes one vertical video per CSV row (text, colors, price) from a template, to test several messages.
A 15-second vertical clip where 3 to 5 key figures count up to their value, from a small CSV.
A vertical top 10 revealed from 10th to 1st, with growing bars, counting values and thumbnails.
A pack of sliding name banners exported as ProRes 4444 and WebM VP9 with transparency, per theme and language.
A script combines 2 or 3 videos into a split screen (side by side, stacked or trio) with labels and dividers.
A video where a slider sweeps two real images or videos to reveal the after, horizontal or vertical.
Turns site-visit notes and photo descriptions into a dated report with snags, actions and owners, inventing nothing.
Converses in the target language at a chosen CEFR level, corrects sparingly and only real mistakes, offers a session recap.
Read-only map of the files and flows a feature will touch, with verifiable file and line citations.
Diagnoses SPF, DKIM, DMARC and Google bulk-sender requirements from supplied records, offline, with no DNS queries.
Writes a skills-based job ad, flags risky terms by script and lists items to verify. Not legal advice.
Prepares the monthly accountant file: sorted documents, gaps, duplicates, draft message. Not tax advice.
Structures a quote from site-visit notes; the tradesperson enters the prices, totals are computed by script.
Drafts review replies for approval, with no guest data and no promises; escalates sensitive cases.
Drafts a welcome message and practical info per booking; access codes only go out if the host authorised them.
Writes a standardised quote request and a script-computed comparison table, without choosing or revealing the budget.
Audits a page with an 8-family grid, ranks actions by score and has a script check the report (evidence, no gain promises, no manipulation).
Decides when to split work across Claude Code subagents, writes briefs, validates the plan by script (no shared files between simultaneous tasks) and guides the merge.
Creates, lists and cleans git worktrees with a script that never forces and refuses to remove a worktree holding unsaved changes.
Reads package-lock.json, package.json and requirements.txt offline and flags install scripts, unpinned versions, risky sources and lookalike names, installing nothing.
Measures cohort retention, designs an honest cancellation flow and respectful payment-recovery sequences, with no dark patterns and the legal points to check.
Designs a draft email sequence (timing, exits, consent, unsubscribe) and has a script check it, never sending anything.
Three React components (button, dialog, tabs) following W3C APG patterns, shipped with keyboard and ARIA tests.
Converts images to AVIF and WebP at several widths, measures the real savings and outputs ready-to-paste picture tags.
Analyzes a SQL schema file and flags unindexed foreign keys, tables without a primary key, redundant indexes and discouraged types, without ever touching a database.
Reads a dependency inventory offline, understands SPDX expressions and sorts each license (permissive, copyleft, restrictive, unknown) into a table. Not legal advice.
Runs axe-core in a real browser on your pages, with a failure threshold, justified exceptions and a sample GitHub Actions workflow.
Computes WCAG 2.2 contrast with the official W3C formulas, checks a palette and proposes corrected colors that keep the hue.
Fact-based self-assessment template and review guide, with no rating or decision and an automatic check for forbidden topics.
Pseudonymised candidate table with evidence per recruiter-declared criterion, discriminatory criteria refused, human decision.
Crosses a PR diff with project coverage to list untested added code, prioritised, with the test cases to write.
Read-only implementation plan: sourced findings, two compared options, recommendation, verifiable steps and rollback.
Answers sourced line by line from your internal policies, marked to-complete otherwise, secrets refused, mandatory human sign-off.
List of processing activities to review field by field per CNIL guidance, with draft questions for each department. Not legal advice.
Processors per processing activity and contracts to check (Article 28, transfers, sub-processors), with a review plan. Not legal advice.
List of documents past their retention period, legal holds excluded, to be approved by a human before any deletion.
Builds the recurring invoice schedule and drafts polite payment reminders, never sending anything.
Keeps the post-handover snag list, computes delays and warranty end, and drafts polite reminders for review.
Recomputes each carrier invoice line against your contract and drafts a costed credit-note request for review.
Selects lapsed clients who may lawfully receive a text and drafts short win-back messages with opt-out, sending nothing.
Writes a band's tech rider (input list, backline, monitors, stage plot, schedule) without inventing gear.
Reads your Zotero library read-only, finds duplicates and incomplete records, and gives you a cleanup plan to apply.
Turns a funding call and your notes into an honest response plan: criteria, eligibility, page budget, gaps, inventing nothing.
Codes consultation responses, counts positions by script and writes a neutral, anonymised thematic summary with no recommendation.
Finds API keys, tokens, passwords and .env files in a repository and its git history, masks the values and prepares rotation.
Picks the value metric, builds 2 to 4 tiers, analyses a Van Westendorp survey and sizes a price test, with no revenue promise.
Splits a brief into scenes and picks the right tool for each (Remotion, FFmpeg, HeyGen, Veo, Runway, Sora), with prompts and rights to check.
Batch-adds a text layer to scanned PDFs with OCRmyPDF, as PDF/A, leaving originals untouched, with a per-file log.
Converts Markdown into Word, PDF or ODT with Pandoc, a style template and CSL citations, after checking citations and images.
Postmark API sending code with a separate transactional stream, logged MessageID and a protected bounce webhook, draft by default.
For each data.gouv.fr dataset, a license sheet (obligations, commercial use, share-alike) and a ready-to-paste attribution. Not legal advice.
Turns a brand guide into a Tailwind CSS v4 @theme block (colors, fonts, sizes, radii, shadows, breakpoints) with class documentation.
Generates a typed, accessible React form and its Zod schema, shared with server-side validation.
Generates a dependency-free typed TypeScript client with timeouts and careful retries.
One-command dev stack: safe compose.yaml, Makefile and a best-practice checker.
k6 scenarios with thresholds and a target guard, then a French report with verdict and comparison.
Playwright screenshots compared to approved baselines, dynamic areas masked, review page for differences.
Claude translates missing keys in your JSON files; a script rejects anything that breaks variables, tags or glossary.
Inventories a site's images, Claude drafts alt text, a human approves, the script writes them into a copy.
Checks your own site's security headers and gives the corrected config for your server.
A procedure and two scripts to write an "AI content" flag into image metadata and check the visible label before publishing.
An 11-point checklist scored by script and a filter that decides, action by action, whether the agent acts, asks a human or is blocked.
A dated Llama, Gemma, Qwen and Mistral table and a script that gives a verdict for your use: internal, commercial, redistributed or hosted.
A dependency-free script that chunks your documents, has each chunk given a context sentence, searches with BM25 and measures the failure rate before and after.
A criteria template, a validator that rejects vague criteria and an evaluator that tells you, with numbers, whether your LLM app meets its thresholds.
A ready-to-plug Python handler for the Claude API memory tool: six commands, one folder per user, path-traversal protection and secret filtering.
From an inventory of AI tools and roles, a script produces the annual AI literacy plan (AI Act Article 4) and the evidence register to fill in.
An offline checker that issues a documented go / no-go decision for HunyuanVideo 1.5 based on where the team, hosting and audience are.
Turn Whisper word timings into ASS karaoke captions where the spoken word lights up, then burn them in with ffmpeg.
Prepare your clips in 9:16, edit them in CapCut from a template, then check the export (format, duration, loudness) before posting.
Generate CapCut auto captions, clean the exported SRT (fillers, proper names, typography, reading speed), then re-import it.
Go from three smartphone photos to three pure-white packshots, edited with GPT Image then normalised and checked by free scripts.
Generate three genuinely different SVG logo directions with Recraft, check the files and compare them on a board (light and dark, one colour, 16 px).
Produce a set of 24 SVG icons with the same stroke using Recraft or GPT Image, then check, normalise and bundle them into a site-ready sprite.
Hand a repetitive browser task to Claude in Chrome with a checked task card, human stop points and a log reviewed after the session.
Have a second model score an assistant's answers from 1 to 5, with an anchored rubric, a strict score format and checks against bias and injection.
One storyboard sheet of 6 or 9 panels keeps the look consistent; it is cut into keyframes, upscaled, then each panel-to-panel transition is animated.
8-second clips slicing glass, lava or crystal with a cutting sound synced to the frame, checked by script and edited into a series.
A selfie-stick vlog series of an invented creature, keeping the same face, voice and prop across shots with prompt blocks repeated word for word.
A daily loop to produce a 1-2 minute episode whose plot follows the audience vote: tally, writing brief, checks, prompts and edit.
8-15 second clips where an object with a face complains in under 50 words, with a checked script, a prompt that keeps the object intact and burned-in captions.
A pro prompt that turns an idea into a Veo 3.1 prompt timed in four 2-second shots, with framing, action and sound per slice, plus a checker that blocks overloaded slices.
A vertical night security-camera style clip (infrared, grain, timestamp) of animals doing something unexpected, with a permanent AI label burned in.
An edit of 4 to 6 answers from invented passers-by, same mic, street and camera, levelled sound and a permanent AI label, with a casting check that rules out real people and sexualised content.
A 6-10 shot first-person series with voice-over and period sound, where every shot is checked for anachronisms and point-of-view breaks before generation.
Create your own half-object half-animal character with a chantable name, an originality check, a free rhythm bed and output formatting, then turn it into a series.
Run a 30-60 second spot made of 15 shots kept out of 300-400 generations: quotas by difficulty, hour-by-hour schedule, scored selection and rough cut.
A vertical short where a translucent skeleton narrates a question and its escalation over time, with checked facts and automated editing.
Make the subject of a photo copy the exact moves of a video you filmed, checking reference and framing before paying.
Star in your own Sora clips with your voice, short prompts that work and strict control over who can use your face.
Fewer invented subtitles in your talking Veo clips, automatic sorting of takes and a fallback to save a good clip.
Chain reverse shots, arcs and top views without the set changing, by editing each keyframe from the previous one and animating the pairs.
Spot AI video glitches frame by frame (vanishing objects, fingers, text) and plan the number of takes per shot before publishing.
Get rotated, top-down or close-up views of a product from a single photo, open source, with a check of every invented detail.
Render Wan clips about ten times faster with LightX2V distilled LoRAs, without frozen motion or burnt colors.
Combine a person, your product and a scene into one coherent image with a commercially usable open-source model.
Bring ElevenLabs narration to life with well-placed whispers, laughs, sighs and pauses, without tags being read aloud.
A step-by-step method and an audit script to get a downloaded ComfyUI workflow running: missing nodes, versions and models.
A dependency-free Python client that runs your ComfyUI workflow with named parameters and fetches the outputs.
Consistent captions with a trigger word that describe what changes and leave out what the LoRA must learn.
A sequence of shots that keep set, light and character, each born from the previous one with Qwen-Image-Edit and the Next Scene LoRA.
Relight a product to match its scene with IC-Light or Qwen Relight, then restore its true colour and sharp logo.
A local French voice-over where a single slider sets emotion, with automatic pace checks and a built-in watermark.
One audio file per spreadsheet row, generated locally even without a GPU, with a lexicon for proper names and re-runs limited to changed rows.
Pick the right separation model, chain passes and check by calculation that nothing is missing or leaking between stems.
Turn a product photo into a textured GLB with TRELLIS.2, then lighten it and set its real scale in Blender for the web.
Two prompts to turn your own photo, mascot or product into a desk figure with box and modelling screen, with no brands and no non-consenting people.
Put each part of your prompt in a named block for more reliable, easy-to-extract answers, with a script that assembles, checks and extracts.
The right order to query a long document, with a script that spots and fixes a badly ordered prompt and an A/B kit to measure the gain yourself.
Make the model copy the relevant passages verbatim before answering, then automatically check that every sentence rests on a real quote.
Describe the task in one sentence and get a full prompt with variables, steps and output format, checked and filled by a script.
A 7-part prompt, an art-direction randomizer and a cliché detector to get distinctive, accessible web pages.
Describe the scene like a photographer, then edit with small targeted requests, with a script that builds the prompt and writes the edits.
Reproduce Anthropic's 5-step prompt improver yourself, for free, with a checker that guarantees no variable or example was lost.
An n8n workflow that asks 3 to 6 questions, then turns the request into detailed instructions without inventing anything, before you run a deep research yourself.
Keep your agents' context lean: references instead of content, verified compaction that loses no decision, and persistent notes.
A first session that sets up tests and scripts, later sessions that reload state and progress step by step, with a script that rejects deleted tests and false statuses.
Have the model propose the minimal additions and deletions that fix a behavior, and apply them with a script that rejects full rewrites.
A crisp, fact-checked infographic: research the facts, lock the exact text in the prompt, then automatically check the rendered text.
Shorts picked by an explainable virality score, cropped on the active speaker without ping-pong, subtitled and loudness-normalised.
A lively dialogue between two synthetic voices, written from your sources, checked (tics, balance, sourced figures), then synthesised locally.
From a single NotebookLM notebook, produce the five materials of a training module, with source checks before and quiz checks after.
A prompt that pits at least three explanations against each other, tracks confidence in a notes file and self-critiques against upfront criteria.
The agent re-reads the same prompt turn after turn, one plan task at a time, until a completion phrase that is accepted only if tests pass.
Write /goal conditions Claude Code's evaluator can actually judge: measurable end state, displayed proof, constraints and a turn bound.
A ready-to-use Claude Code workflow (one auditor per file, two refuters per finding, synthesis) plus an offline simulator to check it at no cost.
An orchestrator subagent that delegates each angle, has three verifiers vote on every claim and returns a cited report without refuted claims.
The agent asks one question at a time with a recommended answer, reads the code before asking, and delivers an approved brief before coding.
A fresh subagent per plan task, then a spec-compliance review and a code-quality review before moving on.
Five single-angle subagents review the PR in parallel, then a script keeps only new, in-angle findings above a confidence threshold.
A Skill that runs the scoping interview while keeping a GLOSSARY.md (one word, one meaning) and short decision records up to date.
A Skill that slices a feature into demonstrable vertical tickets, each declaring its blockers, script-checked and ordered into parallel waves.
A Skill that measures module depth across a repo, delivers an HTML report of deepening opportunities, then runs the interview that ends in a refactoring plan.
A guide to chaining Spec Kit's steps without errors, with an automatic gate and a review prompt between every step.
A method to first make an agent fail without the Skill on trap scenarios, write the minimal Skill, then check it withstands rationalisations.
A Skill that turns "never do X again" into a Markdown rule tested on examples and enforced by a Claude Code hook from the next action on.
A PostToolUse hook that formats and checks every file Claude Code edits and only feeds back the useful errors.
A Claude Code hook that parses every Bash command before it runs and allows it, asks you to confirm it, or blocks it with an explanation.
Three checks hooked into Claude Code: instant reminders on write, diff review at end of turn, mandatory multi-file review before every commit.
A Skill that scores your CLAUDE.md against the actual repo and only adds, at session end, what the code cannot show.
A Claude Code subagent that reads a page as a compact ref tree, acts in one command, waits for a precise signal and re-reads, with domain limits and confirmed clicks.
How to make Claude Code explore read-only, produce a code-anchored plan, check it with a script, and only then write code.
Three phases separated by /clear, each closed by a short script-checked document that alone starts the next phase.
A Skill that enforces a written art direction before coding a UI, then checks by script and screenshot that the code follows it and avoids defaults.
The free n8n-mcp server in documentation mode by default, with a validated build method and a guard before any deployment to a test instance.
Converts a folder of Word, Excel, PowerPoint and PDF files into structured Markdown, flags doubtful conversions, and shows how to run the MCP server without exposing the whole disk.
The reference MCP memory server with a precise instruction on what to remember and when to recall, plus an audit for duplicates, sensitive data and lost relations.
The reference filesystem MCP server limited to one folder, read-only by default, with a checked and reversible bulk-rename method.
Serena connected to Claude Code over MCP to find, reference, replace and rename symbols via language servers, with a checked setup and instructions so the agent actually uses it.
Claude builds objects, materials, lights and camera in your open Blender, pass by pass, with a numeric size check.
A gitmcp.io address gives your agent the current docs and code of a public repo, with a probe that measures doc size first.
Install browser-use and run it safely: allowed domains, small steps, no submissions, post-run checks.
Write durable web scripts: hard-coded where the page is stable, AI only where it varies, with a local selector cache.
Prefix a URL with r.jina.ai to get a page's clean text, with a script and an n8n workflow that handle rate limits, errors and hard pages.
A Telegram bot that transcribes your voice notes, reads your calendar and proposes events, email drafts and tasks, run only after your YES.
A key-protected HTTP endpoint that receives text, has an AI process it and returns clean JSON, with validation, rate limiting and error handling.
From the doc's cURL example to a reliable n8n workflow: key in a credential, right pagination, no silent partial data.
"Hey Siri, ask my assistant": an Apple Shortcut sends the dictated question to an AI agent in n8n and reads the answer.
Build and check VACE's three inputs to switch between video tasks without changing models.
Pick the right GGUF level for your GPU, match text encoder and VAE, and check your files before launching ComfyUI.
Train a LoRA that reproduces your mascot or product in any scene, without overfitting.
Describe your brand voice in seven elements and get an original voice without cloning anyone.
Give each line its target duration, rewrite what overflows and fit the generated voice to the millisecond.
Original tagged lyrics and a clean Style field to get verses, chorus, bridge and a clean ending in Suno.
Film, sort, calibrate and publish a photorealistic 3D scene viewable in the browser, with commercially usable tools.
Paint with AI inside Krita, layer by layer, choosing where it acts and how strongly.
Keep your whole landscape shot and let VACE invent the top and bottom, with a seamless, flicker-free join.
Render visuals in seconds locally with 4-bit Qwen-Image and Lightning, without install errors or burnt settings.
Turn a product photo into a trendy effect clip, locally and for free, with a Wan 2.1 motion LoRA.
Edit an image with a single sentence (scene, text, outfit) on your own machine, without the face or framing drifting.
Make the generated image follow an exact pose, depth map or sketch, tuning strength and step range to avoid stiff images or ignored guidance.
Get faithful stylised portraits from a single photo of yourself, with written consent and settings balanced between likeness and style.
Automatically redraw blurry faces and broken hands of a wide shot in high definition, then paste them back without visible seams.
Generate from one word ("bottle", "sky") the mask of the same object across hundreds of photos, set aside doubtful masks, then edit the rest automatically.
Isolate any object named by a word in a video, repair tracking where it breaks, then get a clean matte down to the hair, exportable as ProRes 4444.
Compute a still photo's depth with Depth Anything V2, then animate it with real relief (truck, dolly) without stretched edges.
Generate locally a natural two-voice dialogue with laughs, sighs and coughs, preparing the script and keeping the same voices across chunks.
Generate full songs locally from tagged lyrics, then rewrite a single passage (repaint) without touching the rest of the track.
Generate locally the sound effects of a silent AI clip synced to the picture, split long clips into passes, measure and fix the offset, then mix with the voice.
Bring the same product, mascot or character back in new Midjourney V7 scenes, with a tuned reference weight and a test that limits GPU time.
A prompt that sorts your shot into JSON fields (camera, subject, action, light, sound) and compares it with the plain-text version through a scored A/B test.
Compute the switch point between Wan 2.2's two experts, check each one's CFG and LoRA, and diagnose a blurry or frozen test clip.
Prepare a Qwen-Image poster whose titles and notices stay legible, accents included, with a size estimate and a letter-by-letter proofreading check.
Tie each brand hex code to an object in the FLUX.2 prompt, then measure with CIEDE2000 whether each visual respects the palette.
Generate and check the trajectory file (pan, orbit, zoom, combinations) that forces Wan 2.2 Fun to apply an exact camera move to your image.
Prepare clips, captions and config without mistakes to train a Wan LoRA with musubi-tuner that reproduces a precise motion.
Erase a passer-by or unwanted object from a video shot with ProPainter: dilated masks, overlapping segments and cross-faded stitching.
Make the filmed person's mouth follow the dubbed voice with LatentSync, after checking consent, duration and front-facing shots.
Generate a Phantom video where your product and a fictional character keep their look, thanks to normalized, described references.
Have 3 to 5 Claude Code teammates investigate different hypotheses and refute each other, with a checked findings file and control hooks.
The agent builds a minimal isolated prototype to settle one risky question, measures, writes a quantified verdict, then throws the code away before the final spec.
Every change becomes an OpenSpec proposal (added, modified, removed requirements), gated, applied, then archived into the project's living spec.
The project is split into phases and atomic tasks; each task runs in a fresh subagent, and state flows through short files checked by script.
After each unit of work, the agent writes the lesson (symptom, cause, fix, verifiable prevention) to docs/solutions and promotes only general rules to CLAUDE.md, within a budget.
Claude Code hooks capture the session, summarize it, index it in SQLite and inject only a short index at startup; details are searched on demand.
Long work goes into Beads, a CLI issue tracker with dependencies and claimable ready work; a checker spots cycles, duplicates and abandoned claims.
Two dependency-free Node scripts continuously show context fill, 5-hour and 7-day limits, git branch and each subagent's state in Claude Code.
A read-only script reads Claude Code session logs and reports usage by day, session, project and model, with a list-price cost equivalent.
Claude Code's official LSP plugins provide definitions, references and post-edit diagnostics; a checker and a probe prove the language server answers.
Repomix produces a single file with tree, filtered files and token count, then our checker blocks leftover secrets and oversized packs.
Claude has the change reviewed by another model (Codex, advisor or subagent), compares both reviews by script and lets the human decide within two rounds.
An @claude mention in an issue or PR runs headless Claude Code in GitHub Actions, bounded by minimal permissions, a cost cap and a hook that denies risky actions.
A library of six Claude Code subagents where each definition sets its role, tools and model, with a checker that flags tiering mistakes and duplicates.
Two French Claude Code output styles: one explains its choices, the other leaves the learner a 3-10 line formative piece with a failing test and graded hints.
A command-line tool where each card creates a branch and worktree for one agent, then helps review, order and merge without surprise conflicts.
A single AGENTS.md imported by CLAUDE.md and wired into Gemini CLI and Cursor, with a checker and a canary word to prove each tool loaded it.
A Skill that searches the skills.sh directory, compares candidates, audits the best one in quarantine without running anything, and installs it into the project only after approval.
A prompt that has the agent explore the repo and write an implementation-ready request: measurable goal, scope, context, pitfalls, plan and three executable validation levels.
Map a legacy system, extract its rules, build a test bench on the old code, rewrite module by module and prove old and new give the same outputs on the same inputs.
A Skill that builds a self-contained HTML page with controls, live preview and a generated prompt listing only the changed values, to paste into the agent.
Six agent roles (analyst, product manager, architect, scrum master, developer, QA) hand over brief, PRD, architecture and stories in fresh context, with a script that checks every handover.
A Claude Code hook reads real pytest results and blocks production code without a failing test, functions no test asks for, and Bash workarounds.
A Skill that turns an instruction you retype daily into a safe /name command, with a guardrail checker, a rendering simulator and four ready-made commands.
An autonomous and safe Claude Code routine prompt (tests, tickets, draft PR on a claude/ branch), a prompt checker and a ticket triage script that masks personal data.
A Claude Code agent that explores training ideas alone overnight, with a script enforcing the time budget, noise threshold, evaluation protection and clean rollback.
A Skill that writes, when the context is saturated, a handoff document under 80 lines for a fresh agent, checked by script before /clear.
What /rewind really restores, its blind spots, and an original git safety net (side-effect-free snapshot, dry-run then real restore, automatic hook before risky shell commands).
Design and test Claude Code Router rules offline (order, conditions, subagents) and estimate savings on your own prices before routing anything.
A custom one-way, protected Claude Code channel (secret, localhost, size and rate limits) that pushes CI or monitoring alerts into the open session as data, never as orders.
An original Archon flow (plan, approval, bounded coding until green tests, fresh review, approval, draft PR), a guardrail checker and a simulator that replays the path identically.
Checked configuration for the claude-context MCP server, a preview of what gets indexed and a measured evaluation (recall, tokens saved) against a local BM25 baseline.
The reference sequential-thinking MCP server, configured without thought logging, with a prompt that forces revision or branching and a script that tells whether the tool actually helped.
Desktop Commander set up with a hardened configuration, an audit script and the security limits actually observed, to give Claude Desktop a terminal knowingly.
Write a precise PRD, have Task Master split it into dependent tasks, check and re-split before starting the agent, then have it execute one task at a time.
Have another model review, debate or settle a decision without leaving Claude Code, sending a prepared excerpt instead of whole files, and settling disagreements with a test rather than a vote.
The community Framelink server gives the agent a simplified layout of a Figma frame; this resource adds an audit that scores the design before coding and a component-by-component generation method.
Turn an in-house tool into an MCP server with a few decorated Python functions, model-friendly descriptions, bounded parameters, clean errors and automated tests.
A 4-step diagnostic method with MCP Inspector (web UI and CLI mode) and a script that recognises the 6 most common MCP server failures and explains the fix.
Give Claude pre-approved, read-only, parameterised SQL queries instead of free SQL access, with a checker that rejects dangerous configurations.
Two ready-to-use GitHub Actions workflows (@claude mention, label-triggered review) with minimal permissions, capped cost and repository guidelines, plus a checker that rejects risky setups.
Run several coding agents at once, each in its own worktree, with batches that share no file and a check of each branch (conflict, overflow, tests) before merging.
The excel-mcp-server MCP server limited to one folder, so Claude creates real xlsx files (formulas, formatting, charts, pivot tables), with a formula checker before use.
Replace an n8n agent's giant system prompt with an index of Claude-format skills stored in a GitHub repository: the agent loads only the useful method, through a tool that rejects any name outside the index.
An n8n workflow that sends invoices and scans to the Mistral OCR API, gets text and tables back as Markdown, then checks number, dates and amounts (net + VAT = gross, French rates) before human approval.
Classifies each new Outlook email with a locally running Ollama model and moves it to the right folder, sending any doubtful case to a review folder.
Every night, exports all n8n workflows, strips secrets and volatile fields, and commits only real changes to GitHub.
Connects your everyday Chrome to Claude through the open-source mcp-chrome extension, with a dedicated profile and a hook that restricts sites and blocks passwords, payments and history.
Politely crawls a website locally with Crawl4AI and produces clean Markdown per page, honouring robots.txt, with a report of refused or empty pages.
Converts PDFs with tables and columns using Docling, then chunks them for RAG without ever breaking a table or losing the heading path.
Adds per-user memory to an assistant with mem0, with consent, rejection of sensitive data, export and complete erasure, history included.
Connects an agent to Gmail, GitHub, Slack or a CRM through Composio, with a policy that enables only needed tools and blocks sending, deleting and payments.
Connects a Langflow project to Claude as an MCP server, names its tools well, and checks after each update that no flow changed silently.
Connects Claude to Perplexity's official MCP server, with a prompt that picks the cheapest sufficient mode and a hook that caps expensive searches.
Connects Claude to arXiv through a free MCP server, with a monitoring prompt using precise queries and a checker that catches invented figures and fake quotes.
Connects Claude to an Obsidian vault through mcp-obsidian, with a hook that restricts writing to one folder and blocks deleting or overwriting notes.
Install n8n, Ollama models and the Qdrant vector database on your machine with Docker Compose, secured and ready for a first workflow on your documents.
A question becomes sub-queries, read pages and a long sourced report saved as a Notion draft, through a bounded loop running on your own machines.
Every morning a checked idea becomes a vertical AI clip with its caption; every evening the videos you approved are sent as private drafts.
Claude creates objects, C# scripts and settings in your open Unity editor, waiting for each compilation, with an automatic review of risky changes.
Claude creates tracks, MIDI clips and instruments in Ableton Live, pass by pass, with pattern checks before writing and a fallback MIDI export.
The model decompiles, renames and summarises an executable in Ghidra, and each of its claims is checked against the program's real inventory.
A local bridge that relays a remote MCP server (OAuth or token) to a client that only runs local servers, with token-safe cache troubleshooting.
An MCP tool that returns an interactive table with buttons in compatible clients, and complete text everywhere else, with no injection risk.
An open source agent fills your forms on never-seen portals, stops before the final button, and a sheet compares each entry with your data before you submit.
Set up the free UI-TARS Desktop app with a self-hosted model in a safe sandbox, and check every click with a guard and reliable coordinate conversion.
OmniParser detects every button and icon in a screenshot; our scripts clean, number and locate them, flagging ambiguity instead of guessing.
Microsoft's Magentic-UI set up in cautious mode, with a mission prompt enforcing an approved plan and stop points, plus two checking scripts.
How to give an agent a memory that dates every fact and knows what was true when, with a tested reference model and a ready-to-adapt Graphiti script.
Activepieces self-hosted with a generated and checked .env, a minimal PostgreSQL + Redis deployment, and Claude connected over MCP with a cautious instruction.
A document assistant built in Dify, published as an API, with a solid system prompt and an evaluator that checks answers, sources and refusals before going live.
A Python launcher that runs agent-written code in a disposable E2B micro-VM, with short lifetime, no network, checked files and guaranteed teardown.
Agent-to-agent messaging, file reservations and written rules so several coding agents share one repository without overwriting each other, with a pre-commit check.
An n8n workflow that reads product pages from a screenshot with Gemini, falls back to text when unsure and checks every price before writing it to Google Sheets.
An n8n chat that answers questions about your figures: the model writes the query, code checks it and does the maths, and every number in the answer is verified.
An n8n workflow where AI picks, page by page, which links to follow to find a specific piece of information on a site, within a strict budget and honouring robots.txt.
The 21st.dev MCP server connected to Claude Code or Cursor, a prompt that enforces your design tokens and a checker that blocks hard-coded colours, arbitrary values and accessibility gaps.
Official AWS MCP servers (documentation and account access) wired with four read-only safeguards and an IAM policy checker.
An n8n workflow that drafts a text, has it reviewed against a rubric with mandatory evidence, then fixes only what failed, logging every step.
An n8n workflow where a lead agent splits the question into scoped tasks, subagents each work on their own documents with evidence, then the lead merges a fully referenced synthesis.
Settings and three instruction templates so an agent neither over- nor under-searches, plus a script that flags contradictory instructions and checks an execution trace.
Improve a prompt in short rounds: fixed test cases, old and new outputs compared blind, per-case voting and a decision rule that protects critical cases.
The official rules that change with Gemini 3 (delimiters, constraints first, question last, verbosity, temperature), two templates and a script that converts and checks your prompts.
A prompt that turns "act as a…" into a complete role (audience, goal, tone, constraints, success criteria, limits), 12 ready-made roles and a checker.
A fixed-section template for GPTs, Gems and Projects instructions, with a checker that flags missing sections, contradictory rules and excessive rigidity.
A prompt that extracts verbatim values, and a script that finds each value's exact position in the text, highlights it and isolates invented values.
Convert PDF, Word and Excel to Markdown, strip conversion noise, flag doubtful tables and assemble a prompt with tagged documents and the question last.
Describe a comic page panel by panel (style, characters, framing, background, speech bubble, mood) and produce a full prompt per panel to keep characters and style consistent.
A prompt that turns a one-line idea into a complete video prompt (camera, subject, action, context, style, sound), positively phrased, plus a pre-generation checker.
From topic to 9:16 video with MoneyPrinterTurbo, plus a useful-script prompt, an automatic anti-spam check and an ffmpeg edit with normalized loudness.
Simulated experts with different viewpoints question your sources, a plan emerges, then a long article where every sentence is cited and checked.
A 10-page story with the same hero in every picture thanks to word-for-word character sheets, script-checked, with local audio narration.
Replace the prefill trick, rejected by recent Claude models, with a JSON schema, a no-preamble instruction and code-side checks.
One model drafts, a second scores it against a rubric and says what to change, and the loop stops at the threshold, the turn limit or when progress stalls.
Describe the task, provide examples and a metric: DSPy picks the prompt demonstrations itself, with a complete, tested and saveable pipeline.
Three complete prompts that make the chatbot step back, compare several answers and explore options before concluding, with a voting script.
Five original French patterns for Fabric, a maintainable layout checked by a script and a fallback runner for a local model.
A French guide to POML with a complete template, data kept in separate JSON files and a script that renders, sanitizes and checks the prompt before sending.
Explore styles through random codes on fixed subjects, log them, then reuse the chosen code across a whole series without switching model versions.
Meaning-based segmentation, enforced glossary, three-pass translation and automatic checks for readable translated subtitles, then assembly with ffmpeg.
Install open-notebook at home, locked to the local machine and wired to a local model, with sourced prompts for questions, notes and podcasts.
After each research round, a reviewer lists what is missing and runs targeted queries, with four guaranteed stop rules, before writing a cited report.
A document becomes a faithful outline, then editable PPTX slides in your brand with self-hosted Presenton, checked automatically.
Classify a request to send it to the right specialised prompt, and ask a closed question to several parallel voters to keep the majority.
Several candidate prompts run on test cases, judged pairwise in both orders, then ranked by Elo score with reliability warnings.
Shorten a long context two to five times before sending it to the model, and measure lost facts and quality before adopting the compression.
Run GPT Researcher locally or on your own documents, give it a well-scoped question, then check and review the cited report it produces.
A research run that digs into each lead over several levels with two settings, breadth and depth, and states its cost in calls before starting.
Install Perplexica / Vane at home with SearxNG and the model of your choice, query it from the command line and check that every fact cites its source.
An excerpt from your own podcast spoken by a fictional baby in a studio, lip-synced, as a subtitled vertical clip labelled as AI.
Physics prompts to get an animal diver's rotation and water entry right, then full TV-style packaging: lower third, slow motion, judges' scores and commentary.
Write, check and edit a 1-3 minute mockumentary where invented characters talk about their own prompt to camera.
Tell a moving 30-second story in 8-12 images of the same cat, cut on the beats of an original song.
Generate a locked-off shot where your furniture box opens and parts assemble into a furnished room, then auto-check camera and light.
Get a continuous pull-back from a person to Earth from space, then reverse it to dive seamlessly into your own video.
Have tiny construction workers prepare a giant dish while keeping the same scale across every shot.
Animate your collectible figurine image in Gemini without losing its painted plastic look, with an automatic clip check.
Turn a scene into a JSON video prompt (characters declared once, timed shots, precise fields), validate it and compare with its prose version.
Find on your own machine the fastest quantization, distilled LoRA, sampler and step combination without quality loss, with a measured test bench.
Shoot a clean driving video, check and crop it, then animate a portrait or a pet with LivePortrait locally.
Turn a horizontal video vertical by generating top and bottom with VACE instead of cropping, with a measured plan, canvas, mask and band check.
Train a small LoRA on 20 to 40 images of your character to keep it identical across every Wan 2.2 shot.
Give your character one authorised voice across all AI clips without losing emotion, then re-sync the lips to the new audio.
A Claude Skill that turns your idea into ready-to-paste Veo 3.1 prompts with camera, sound, timed shots and a consistent character, then checks them.
Bring your child self and your current self together in an instant-photo style picture, then animate the hug for a few seconds without merged faces.
Test a film concept with a 60 to 90 second trailer that follows film grammar and YouTube's anti-deception rules.
An n8n workflow that writes, breaks down and checks dozens of 1-2 minute vertical episodes with the same virtual actors and a controlled cost per minute.
Turn 480-720p AI clips into restored high definition with SeedVR2, with the right memory settings and no seams or flicker.
Give silent AI clips a 48 kHz sound-effects track synced to the picture, using an engine whose licence allows your use.
Fix one detail of a generated shot without regenerating it all using VACE, with clean masks and source videos checked before rendering.
An n8n workflow that takes a 3 to 10 minute short film, mixing live action and AI shots, all the way to a festival-compliant submission pack.
A 30-second 9:16 line-art video: the crowd runs the same loop, one breaks out, falls, starts again and succeeds. No video AI, no credits.
A 35-65 s 9:16 video: shears cut balls that come back more numerous and in a new colour, until the circle bursts. Computed locally, no video AI, no credits.
A 9:16 low-poly 3D video of about 1:30: two everyday people trade blow for blow, each blow gets a gold title, and the "Well done, you created…" ending restarts the loop.
A 15-25 s 9:16 video: by a lake at night, one stick figure keeps skipping stones while the other looks for the perfect one. Scoreboard, speech bubbles, punchline and your name at the end. Computed locally, no video AI, no credits.
A ~30 s 9:16 video: a volcano drops countdown mini-volcanoes that erupt, a carnivorous plant bites it and steals its HP. Health bars, red damage, green heals, a close fight picked by simulation, and your name at the end. Computed locally, no video AI, no credits.
A 30-50 s 9:16 low-poly 3D video: a boy, an orange box and a small robot act out a three-voice sketch, with yellow holograms and coloured subtitles. Rendered locally in Blender, ElevenLabs voices.
Reads a product page and writes three testimonial-style ad scripts.
Drafts replies in your tone; nothing is sent without your approval.
An object presented as an animated spec sheet, in 16:9.
Sends your script to a text-to-speech service and files the audio.
Pulls rushes from a Drive folder and files them for editing.
Prepares captions, keywords and a schedule for each network, for you to approve.
Every finished video lands in your Notion calendar, with its status.
An agent that spots rising formats and suggests five video ideas.
Assembles clips, music and captions into 9:16, 4:5 and 16:9 in one command.
Droplets, reflections, slow focus pull: the thumb-stopping shot.
A Skill that turns an idea into a shot-by-shot storyboard, prompts included.
A realistic studio walk for a fashion brand, one consistent model across 6 shots.