Find the open-source solution that fits.

open-book-is-good teaches an agent to turn officially published exam papers into an offline practice system by inspecting each source, parsing its layout and answer format, extracting figures, and building a question bank with explanations. It is mainly tested with Taiwanese national exams; scripts need adaptation, AI explanations are not official answers, and some outputs still merit spot checks.

unifusion runs an OpenCode orchestrator that sends one difficult brief to several panelists in parallel and keeps the synthesizer separate from the panel. It captures the final analysis, panel reports, session context, task events, and saved provenance, making the harness useful for comparing independent model views before a recommendation.

Throughline turns interview transcripts, notes, usability sessions, or survey exports into a prioritized product backlog. It segments evidence, clusters themes, ranks actionable problems or opportunities by evidence weight, and links every item to verbatim participant quotes; three or more participants form a Finding, while one or two remain a Hunch.

play-console-draft inspects an Android, Kotlin Multiplatform, or Flutter project, prepares a signed AAB and policy pages, and drafts store listing assets and declarations from the detected data practices. It can drive Play Console through a Chrome extension, but agreements, identity and payment steps, the first upload, and production rollout remain human gates.

Lead Magnet Architect guides a lead-magnet build instead of generating a one-shot asset. It tests whether a stranger would pay $50 for the standalone resource and whether the small win exposes the larger paid problem, then routes the idea to a format checklist, reveals steps incrementally, and runs a final quality pass.

MoneyPun collects bank and card alerts, merchant confirmations, receipts, e-wallet notices, and utility bills from Gmail, matches them into transactions, and asks about ambiguous items before producing JSON and HTML reports. It stores account and category preferences for later runs, and requires Claude Desktop Cowork mode, a Gmail connection, and notification emails from the relevant banks or cards.

luu closes a Claude session by extracting decisions, current state, daily notes, and AI-ecosystem traces into a configured vault. It backs up files before overwriting, avoids duplicate appends, checks paths and modification times, and reports disk-backed results when a step fails, so it can be rerun without silently erasing older notes.

im-designer translates subjective design language into requests that specify the user problem, target component, state, visual property, scope, and accessibility considerations. When a phrase has multiple plausible meanings, it can render HTML options for comparison before producing versions for designers, developers, AI design tools, or Figma and Penpot, while keeping color choices token-based.

news-for-beginners collects recent articles from a user-selected site or feed, organizes them into a narrative for beginners, and writes three artifacts: a source ledger, a standalone HTML news page with topic filtering, and a preview PNG. It runs as an independent Claude Code skill and keeps collection inside the selected site boundary. Network access is needed for the source, and the preview image requires Playwright with Chromium or another real browser renderer.

Rameez Taste Skill helps AI coding assistants give landing pages, portfolios, and redesigns a more deliberate visual direction. It maps the brief to a design system, tunes design variance, motion, and visual density, and checks for AI tells, dark mode, and accessibility. It excludes dashboards, data tables, and multi-step product UI, and this version is manual-invoke only through the taste keyword.

fk_marketing’s achete-pas skill separates a real need from a product story before comparing new items. It considers doing nothing, repairing, borrowing, or buying used first, then examines specifications, observed prices, total cost, and footprint only if a new purchase remains justified. The current implementation assumes France and French sources, runs on claude.ai rather than Claude Code, and requires manual re-upload when updated.

Deep Stock Analysis provides a Buffett-style framework that treats a stock as a whole business rather than a short-term price forecast. It asks the analyst to separate facts, assumptions, inferences, and opinions while examining the business, moat, management, financial quality, intrinsic value, margin of safety, and reasons not to buy. The templates work as plain prompts or a Claude Skill across three markets. They are for research and method practice, not investment advice, and current data must be checked independently.

RGS Skill helps Dutch sole traders, BV owners, bookkeepers, and bookkeeping-tool builders map ledger entries to the national Referentie GrootboekSchema. The same mapping can support annual accounts, tax filings, statistics, and SBR reporting, with guidance for entity and sector subsets and tools such as MoneyBird. Its scope is Dutch-GAAP and SME bookkeeping: it does not supply IFRS mapping or consolidation logic for multinational groups.

visual-plan helps teams turn a text plan into a self-contained interactive HTML document for collaborative review instead of a wall of terminal Markdown. It supports diagrams, file trees, annotated code and diffs, anchored comments, and open-question decisions; reviewers can print or export PDF and send section-specific feedback back for iteration. No separate build step is required, and the page can be deployed to Cloudflare.

bdev-copyguard audits a URL or local codebase across hotlinking, bots, AI crawlers, canary fingerprints, watermarking, source exposure, framing, ownership evidence, and open data endpoints. It returns a prioritized score and hardening plan while separating real server-side access control from measures that only raise the cost of copying. The skill explicitly rejects right-click disabling as security: rendered content can still be saved, while protected data must stay behind authenticated server endpoints.

claude-dependency-mapper turns static imports into two complementary views: an Obsidian vault for people to inspect architecture and change impact, and a compact PROJECT_MAP.md that Claude can read once to navigate code. A hook refreshes the views after source edits; small projects stay with direct exploration, while larger ones get the map to reduce repeated searches. Resolution is static-import and regex based, so unusual constructs may be incomplete.

HQ Report renders Markdown and frontmatter into a branded PDF with a title page, dark cover, accent headings, severity tables, code blocks, callouts, and page-number footers. A small JSON brand file controls the organization name, badge letter, and color tokens without editing the template, and the same renderer works as a Claude Code skill or standalone Python CLI. WeasyPrint and its native dependencies are required, while the output remains a standard PDF.

ai-footprint reads local Claude Code and OpenAI Codex transcripts, removes duplicated resumed messages, and estimates water use, energy, and CO2e with model- and datacenter-specific factors. It reports combined totals, ranges, per-agent shares, and offset-cost scenarios without sending data away or requiring a paid API. The skill is read-only and portable across Claude Code and Codex. Its numbers are modeled estimates, not direct measurements from a provider’s meter or invoice.