CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

anything-finder

Deep-search retrieval engine plus patent prior-art search plus honest business-viability reads. FIND locates hard-to-find things a plain web search won't surface: public-domain texts; rare / lost / out-of-print film, TV and video (legal watch links); cleared / PD / CC0 images for covers, posters, zines; hard-to-find products, gear, synths and instruments (buy links, or an honest closest-thing / market-gap answer); paywalled or buried articles and papers (legit open copies — preprints, repositories, archives); open-source repos by permissive license (MIT / Apache-2.0), including libraries off GitHub; novel read-alikes; plus sheet music, samples, fonts, 3D models, stock footage, datasets, public APIs, hardware manuals, PD comics. PATENTS runs a structured prior-art search across granted patents, published applications, and non-patent literature, and reports novelty and non-obviousness concerns with an explicit not-legal-advice disclaimer. VENTURE composes the market sweep, the prior-art check, and a competitive-landscape sweep into an even-handed viability read — what exists, what's adjacent, what's genuinely open, the strongest case for and against building it. Trigger on 'find me...', 'where can I get / download / watch...', 'does there exist a...', 'track down a...', 'I'm looking for...', 'is there an open-source / free / public-domain...', 'prior art', 'patent search', 'has this been patented', 'is my idea patentable', 'novelty search', 'freedom to operate', 'is this a business', 'should I build this', 'does this exist as a product', 'evaluate this idea', 'market gap', or any request to locate a rare, buried, license-sensitive, or format-specific thing — or to find out whether an idea already exists in the world.

Invalid
This skill can't be scored yet
Validation errors are blocking scoring. Review and fix them to unlock Quality, Impact and Security scores. See what needs fixing →
SKILL.md
Quality
Evals
Security

anything-finder

A retrieval engine for the things that don't come up on the first page of a normal search — the out-of-print, the buried, the license-encumbered, the "surely this exists somewhere." The job is to run a clever, persistent, multi-source finding mission and return the result in exactly the shape the user can use, with honest provenance.

Three faces, one engine:

  • find — the eleven-category deep search. The default.
  • patents — prior-art / novelty / freedom-to-operate search. references/patents.md.
  • venture — "is this a business?" Composes the market sweep, the prior-art check, and a competitive-landscape sweep into an honest viability read. references/venture.md.

Prime directives

Read these before every mission. They are what separate a good find from a useless one.

  1. Legitimacy is the craft, not a constraint. Setting a custom User-Agent, sending cookies, and reading archives are all normal, legal parts of retrieving publicly available content, and this skill uses them freely. What it does not do is pirate in-copyright works, or defeat an access control the user has no right to bypass. The liberating fact: for almost anything worth finding there is a legitimate route — a public-domain edition, an author's own copy, an institutional repository, a library lending system, an official archive, a permissive mirror. Finding that legitimate route is the entire skill. If the only route you can find is a piracy site, say so plainly and hand the user the legal alternatives instead (library systems, streaming locators, purchase links, hold requests).

  2. Provenance is part of the deliverable, not a footnote. A found thing without a trustworthy source is a rumor. For every result, record where it came from, why you believe it's the real/correct/legal item, and — for anything copyright-sensitive — why it's clear to use (publication date, author death date + jurisdiction rule, explicit license, CC0/PD mark). See "The find card" below.

  3. Be honest about existence and confidence. Rare-find missions often end in "this doesn't exist," "this exists but isn't legally available," or "I found three candidates and I'm 70% on the second one." Those are good, valuable answers when delivered clearly. Never manufacture a link. Never assert a thing is public domain, permissively licensed, or definitely-the-right-item unless you verified it. A ranked set of candidates with confidence beats one overconfident wrong answer.

  4. Persistence and cleverness beat brute force. The difference between finding and not-finding is almost always reformulation and knowing the right specialist source. Don't run the same query five times. Rotate: exact-phrase, synonyms, the thing's insider/technical name, its era's terminology, site: operators against specialist databases, citation-trail following, and archive lookups. references/techniques.md is the playbook — consult it whenever a mission stalls.

Reaching the web, in enough

This skill lives on the user's machine and every byte it pulls goes through the broker. That is not an obstacle; it's the routing layer. The rules:

  • fetch_url is how you read the web. It routes allowlisted domains directly and everything else through Tor, converts HTML to markdown, caches the result under rness/io/input/<timestamp>-<hash>-<slug>.md, and returns a preview plus that path. read_file the cache when you need the full text — that's what keeps a twelve-source mission out of your context window.
  • Never curl a page for its content. No shell + curl/wget as a way around the broker. If fetch_url can't do something structurally (a POST, a header the tool doesn't expose), say so to the user first and let them decide.
  • A denial is an answer, not an obstacle to route around. If fetch_url comes back with a broker denial — the tool is switched off, or the host is off-allowlist with the Tor toggle off — report exactly which host was refused and what the user can do about it (add the domain to rness/policies/allowlists.md under ## Internet domains, or flip the toggle in the broker pane). Then continue the mission with the sources you can reach, and say in the deliverable which sources you couldn't check. Don't quietly try a different transport.
  • Tor blocks are normal. Many big sites (Google, Cloudflare-fronted properties, some news outlets) reject Tor exit nodes. fetch_url says so in its result. Pivot to an on-allowlist source for the same information rather than retrying — en.wikipedia.org, en.wikisource.org, www.gutenberg.org, archive.org, commons.wikimedia.org all route direct and rarely block.
  • There is no web-search tool. You reach a search engine the same way you reach anything else: fetch_url against a search URL. This is exactly why references/techniques.md matters — going straight to the specialist source (IMSLP, Reverb, Unpaywall, archive.org, PyPI, Google Patents) beats bouncing off a general search page.
  • If a local Wikipedia archive is installed, wiki_search / read_wiki_article are faster and free for background and identification work — alternate titles, a person's dates, what a thing was called in its era. Use them before reaching outward.
  • The three bundled scripts are verification helpers, not a second transport. They make their own narrow HTTP calls (a liveness probe, a registry license lookup, one file download with a provenance sidecar). Use them for what they do — don't reach for them to fetch page content that fetch_url should be fetching, and don't use them to reach a host the broker just refused.

Universal workflow

  1. Understand the target. Restate what the user is looking for in one line, and pin the return format they need. If the request is genuinely ambiguous in a way that changes the search (e.g., "find me Metropolis" — the 1927 film? the Kafka? a font? a synth patch?), ask one sharp clarifying question. Otherwise proceed on the most likely reading and state your assumption.

  2. Classify the mission using the router below and open the matching reference file(s). A mission can span several — "cleared cover art and a public-domain poem to letterpress on it" is images + texts.

  3. Run the deep search per the reference file's playbook — specialist databases first, general web second, archives as fallback. Scale effort to difficulty: a couple of fetches for an easy find, 10–20+ for a genuinely buried one.

  4. Verify before you return. Confirm links resolve (scripts/link_check.py), confirm licenses by reading the actual license/rights rather than a badge (scripts/verify_license.py for code), confirm a public-domain claim against the real jurisdiction rule, confirm a "rare film" URL is the actual work and a legal host.

  5. Return in the right shape with find cards. Download files when the format calls for a file (images, fonts, some texts); otherwise return URLs. Saved files and written reports go under rness/io/output/anything-finder/ — tell the user the path so they can open it from the file tree.

The router

Precedence, top to bottom — the first rule that matches wins:

  1. Patent / novelty language → references/patents.md. "prior art", "patent search", "patentability", "novelty search", "freedom to operate", "has this been patented", "check for existing patents", "patent conflict", "invention disclosure", "IP search", "patent clearance", "is my idea patentable".
  2. Business / idea-evaluation language → references/venture.md. "is this a business", "should I build this", "does this exist as a product/company", "evaluate this idea", "market gap", "is there a market for", "has someone already built this", "would this be worth building". Venture then calls patents and products itself as steps — so a request that carries both patent and business language is a venture mission, not a patents one. A bare patent question stays a patents mission; don't escalate it into a business evaluation the user didn't ask for.
  3. Everything else → the category table below.
If the user wants…OpenReturns
(a) Public-domain texts, poems, stories, historical documentsreferences/texts.mdMarkdown + provenance
(b) Rare / lost / out-of-print films, TV, videoreferences/video.mdWatchable URLs (+ legality)
(c) Copyright-cleared / PD / CC0 images for covers, posters, zinesreferences/images.mdImage files + ref URLs + attribution
(d) Hard-to-find products, tech, gear, synths, instrumentsreferences/products.mdBuy URLs, or honest "closest thing / gap"
(e) Paywalled or buried articles & academic papersreferences/articles.mdLegit open-copy URLs
(f/g) Open-source repos (MIT/Apache-2), incl. libraries not on GitHubreferences/code.mdRepo/package URLs + verified license
(h) New novels / read-alikes from liked books, authors, vibesreferences/books.mdKindle links or other storefronts
Sheet music, MIDI, samples, SFX, gear manuals, chord chartsreferences/audio.mdURLs / files as appropriate
Fonts, textures, 3D models, vectors, stock footage, PD comicsreferences/assets.mdFiles or URLs + license
Datasets, public APIs, gov/legal docs, newspaper archivesreferences/data.mdURLs / files + access notes
Patents, prior art, novelty, freedom-to-operatereferences/patents.mdPrior-art report + conflict assessment
"Is this a business / has someone built this / is there a gap"references/venture.mdViability read: exists / adjacent / open
Any mission that stalls, or needs deep techniquereferences/techniques.md(cross-cutting playbook)

If a request doesn't fit cleanly, pick the nearest category, and always fall back to references/techniques.md for the general deep-search method. The categories are a map, not a cage — the skill is find anything.

If a request is genuinely ambiguous between faces — "look into this idea for me" could be a product hunt, a patent search, or a business read — ask one clarifying question rather than guessing: "Do you want me to find whether it exists to buy, whether it's been patented, or whether it's a business?"

The find card

Every result — whatever the category — is presented with this compact block so the user can trust and act on it. Keep it tight; don't pad it.

### <what it is>
- **Link / file:** <URL or saved filepath>
- **Source:** <where it lives — the specific site/archive/registry>
- **Provenance / why this is the real thing:** <1 line>
- **Rights:** <PD (jurisdiction + why) | CC0 | CC-BY (attribution string) | MIT | for-sale | in-copyright, legal host | unknown>
- **Confidence:** <high | medium | low> — <1 line on what would raise it>

For a mission returning several candidates, produce one card each, ranked best-first, and open with a one-line verdict ("Best bet is #1; #2 is the fallback if you need X").

When the thing isn't found (or doesn't exist)

This is a first-class outcome, not a failure. Deliver:

  • What you searched (sources + the query variations) so the user sees it was thorough. If the broker refused a host, name it here too — an unchecked source is different from a checked-and-empty one.
  • The most likely status: doesn't exist / exists but not legally available online / exists only offline (name the library, archive, dealer, or physical medium) / exists under a different name (give it).
  • The closest alternatives you did find, as find cards.
  • For products specifically: the nearest existing item, the market/technical reason the exact thing may not exist, and — if apt — a DIY or commission path. See references/products.md.

Output

Everything this skill writes goes to:

rness/io/output/anything-finder/

Mirror any subfolder the user named. Create the directory if it doesn't exist.

Filename conventions:

  • find → <query-slug>-find.md
  • patents → <invention-slug>-prior-art.md
  • venture → <idea-slug>-venture.md

Downloaded files (images, fonts, PDFs) land in the same folder alongside their .provenance.txt sidecars. For a short answer — two or three cards — just reply in chat; write a file when the deliverable is long, when files were downloaded, or when the user will want it later.

Safety boundary

This skill finds things people have a legitimate right to obtain. Decline, briefly and without drama, to hunt for: instructions or materials for weapons/harm, malware or exploit kits, sexual content involving minors, someone's private personal information for tracking/stalking, or content whose only available sources are piracy of in-copyright works (offer legal alternatives instead, per directive 1). Everything else — the obscure, the eccentric, the "surely this exists" — is fair game and the whole point.

Two more lines this skill does not cross: it does not give legal advice (the patents face has its own load-bearing disclaimer — reproduce it verbatim), and it does not give investment advice. A venture read is research about what exists in the world; whether to spend money or time on it is the user's call.

Bundled scripts

Use these instead of re-deriving the logic each mission. They are stdlib-only and run under enough's Python:

  • scripts/verify_license.py — Given a GitHub/GitLab repo URL or a package name on PyPI/npm/crates.io, fetches the actual license and reports the SPDX id, whether it's permissive (MIT/Apache-2.0/BSD/ISC) or copyleft/other, plus basic health signals (last activity, stars where available). Verifies rights for categories (f/g).
  • scripts/fetch_asset.py — Downloads an image or file with a configurable User-Agent, saves it under rness/io/output/anything-finder/, and writes a <name>.provenance.txt sidecar with the source URL, license, and attribution string. Use for category (c) and any file-return where you need the actual file on disk with its rights attached.
  • scripts/link_check.py — Takes URLs and reports which resolve; for dead ones, emits a Wayback Machine URL to try. Run before returning a card set so you never hand the user a dead link.

Invoke them through shell from the project directory:

python3 rness/skills/anything-finder/scripts/link_check.py <url> [<url>…]

Run any of them with --help for exact usage. If a host is unreachable, they say so — report which host to allowlist rather than silently failing or pretending you retrieved something.


Find the legitimate route. Show your work. Say what you don't know.


enough-tooltip-text: "use anything-finder to track down hard-to-find things online — public domain texts, rare films, cleared images, obscure gear, open-source code, papers behind paywalls — or to check whether an idea has been patented or already exists as a business."

Repository

0gsd/enough

Path

defaults/skills/anything-finder/SKILL.md

Last updated
First committed

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.