Start with the job, then find an open-source answer

Shopify Custom App Builder focuses on single-store custom apps running on Cloudflare Workers. It guides credential setup, storage choices, App Proxy and webhook signature verification, merge-on-write synchronization, and secret-safe packaging. The scope is intentionally narrow: examples assume Workers and custom apps, while Node, Deno, and multi-store App Store development are outside the primary focus.

Ganhuo Travel Lock helps plan family trips, graduation travel, and city routes by comparing tickets, hotels, transport, and budgets. It records the source and query conditions, presents trade-offs, and keeps a status table for locked, suggested, pending, or backup choices. It is a decision aid rather than a ticket bot or payment robot, and it does not promise the lowest price.

Ganhuo Textbook Downloader parses natural-language requests for Chinese school or university textbooks, using province, city, grade, subject, volume, and edition clues to rank public PDF candidates. Cover or page OCR can add matching signals, and split PDF parts are reassembled automatically. It returns local paths, upstream locations, source URLs, and ranking reasons, but does not claim official edition status.

Transcrever Reel downloads audio from an Instagram Reel and transcribes it locally with yt-dlp, ffmpeg, and Whisper. It returns readable paragraphs plus timestamped segments and can detect Portuguese, English, and other languages. Audio stays on the computer, but macOS may need a logged-in browser session, and the first run installs tools and downloads a speech model.

HTML to Presentation converts an index.html into a self-contained 16:9 deck with keyboard navigation, deep-linked slides, an overview grid, speaker notes, a dark-mode toggle, and print layout. Its extractor and linter have no dependencies, while the Playwright harness screenshots light and dark slides and exercises keyboard navigation end to end. Playwright is an extra requirement for the visual checks.

D2 Diagrams for Claude Code turns an architecture request into D2 code, renders an SVG, and checks the visual result before delivery. Its render loop looks for tangled edges, drifting labels, imbalance, and clipped text, using ELK by default and TALA when a valid license is available. The plugin works without TALA; PNG self-inspection may download a headless browser on first use.

ClaudeCodeTidy audits project instruction files such as CLAUDE.md, rules, skills, and memory notes, then relocates misplaced content, compresses verbose guidance, and removes only confirmed-dead material. Its companion skill learns from real tidy runs using evidence rather than guesses, while every run records a durable log. The plugin is scoped to instruction hygiene, not application code.

Interview Mapper builds qualitative analysis from atomic, quoted evidence instead of summaries of summaries. It verifies that quotes are verbatim and checks whether each quote actually supports the attached claim, then assembles insight cards, personas, journey maps, opportunity prioritization, or decision memos. English and Russian packages include sixteen analytical lenses, five output types, and reliability checks.

Agent Production Readiness reviews an AI agent across evaluation, cost, observability, guardrails, human oversight, reliability, and governance. It follows a four-phase process, assesses each dimension against available evidence, and returns gaps in priority order alongside a score out of 28. The score is a conversation starter rather than a certificate, so the skill is best used for operational review before launch.

Loop Engineering turns an ambiguous coding request into five stages: frame, model, gate, build, and verify. Each slice starts with a check that should fail, moves to green through small changes, and ends with tests, type checks, builds, and an independent verifier. Frontend or deployment work also needs a real runtime probe, while high-risk slices cannot skip independent verification.

seele-s-html injects a presenter view into an HTML deck that can be paged in a browser. Pressing S opens current and next-slide previews, speaker notes, and a timer, with navigation synchronized to the audience window. The zero-dependency Python postprocessor leaves the original generator alone and produces a portable HTML file; PNG and PPTX inputs are outside its scope.

Layerdex uses hfviewer to render a public Hugging Face model at a chosen granularity and save a cropped architecture graph plus normalized JSON metadata. The JSON can include nodes, operation types, vocabulary, and attributes, and is checked against a published schema. It needs Playwright and Chromium, does not download weights, and does not support gated or private models.

secret-hygiene teaches Claude Code to read credentials from the environment instead of hardcoding them, then uses a zero-dependency Python scanner on directories or staged changes. It detects JWTs, PEM keys, common cloud credentials, connection strings, and high-entropy secret-like variables, with narrowly scoped ignore rules. For public or production repositories, it recommends adding gitleaks and trufflehog.

snakeoil-radar watches a user-maintained list of TikTok handles, ranks fresh videos by reach and engagement, extracts transcripts, and checks concrete claims against PubMed literature. A traffic-light verdict is issued only when a PMID and link can be retrieved; otherwise the finding stays UNVERIFIED. The free setup uses yt-dlp, curl, and optionally local Whisper for missing captions.

agent-or-workflow reads a PRD, ticket, or code sample, extracts seven design signals, and passes them to a deterministic verdict engine. It chooses among a single model call, a workflow, a model-assisted workflow, one tool-using agent, or multiple agents, then writes a decision record with oversight, risks, and the cheaper alternative. The model interprets the design; the script owns the mapping.

fable-setup-optimizer reads Claude Code settings, plugins, agents, and external CLI availability, then presents a done/partial/not-done assessment. With approval before each edit, it can route exploration to cheaper models, load plugins per project, reduce context overhead, tier effort, and delegate typing. Changes to model, plugins, or CLAUDE.md take effect in a later session.

content-audit scans user-facing text across sites and apps for AI-like filler, TODOs, internal implementation details, placeholders, and weak UX copy. It reports each finding with the original line, location, and a drop-in revision, then gives separate AI-likeness and release-readiness scores. The skill evaluates wording and visible leaks, not the truth of external facts.

video-reader drives ffmpeg and ffprobe to extract frames, metadata, spectrograms, and waveforms from local media, with optional Whisper transcription and yt-dlp URL downloads. It exposes both standalone Python scripts and a Claude Code skill, while the core features require only ffmpeg. Frames are downscaled and capped by default so visual inspection does not overwhelm the context.