# Reference Hunt > Use whenever the deliverable is *found* material rather than *generated* material — real examples, reference images, prior art, competitor work, source collections. Visual references first (GFX, UI, thumbnails, brand, packaging, web layout), then code/implementation examples and written sources. Gives the multi-angle query fan-out, per-source-type phrasing, the reject list, the verification rule, and where findings land. Use it on "find examples of X", "get me references for X", "who does X well", "show me how others do it", "busca referencias de X" — and use it even when the request looks like one quick web search, because firing a single query and reporting whatever ranks first is exactly the failure this prevents. Do not answer these from memory. - Skill: `presidenteog/reference-hunt` (Agent Skill) - Install (CLI): `npx skillmds@latest add presidenteog/reference-hunt` - Raw SKILL.md: https://api.skillmd.com/api/skills/presidenteog/reference-hunt/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: PresidenteOG (https://skillmd.com/u/presidenteog) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/presidenteog/reference-hunt --- # Reference Hunt The output is found material with provenance. Every finding carries a live URL, a date checked, and one line saying why it survived. A claim without a source is not a finding — it is memory wearing a citation's clothes. ## 0. Pin the frame before the first query Three lines, written down. They are the accept/reject rule for everything after. | Pin | Question | Example | |---|---|---| | Type | visual / code / written / mixed | visual — Roblox GFX character renders | | Acceptance | what makes one finding good, testably | "a named in-frame emitter lighting the character, not studio 3-point" | | Volume + stop | how many, and what ends the hunt | "6 refs; stop when 2 consecutive angles return only already-seen URLs" | If X is a broad category ("good UI", "nice thumbnails"), narrow it to a testable phrase first. A broad hunt returns aggregator sludge, and no amount of ranking rescues it. ## 1. Vault first — one cheap pass Skip only when the topic is plainly external and new. - `mem-search` — did an earlier session already hunt this? - `/vault-search ""` — semantic index over the vault. - Visual: check `ClaudeFX/Knowledge/Design//` — a profile and a `refs/` folder may already exist for that domain (`gfx`, `thumbnails`, `web-layout`, `logo`, `vfx`, …). Name what already exists before adding to it. A duplicate reference is worse than no reference: it inflates a profile's apparent evidence without adding any. ## 2. Fan out — multi-search, never one query One query is one hypothesis about vocabulary. Run 3–5 angles, in parallel, in a single tool block. | Angle | Phrasing | |---|---| | Plain | the thing as the requester named it | | Practitioner | the jargon a professional in that field uses instead | | Container | the artifact type that holds it — "breakdown", "case study", "portfolio", "showcase", "post-mortem", "gallery" | | Negative space | the adjacent thing it is often confused with, to map the boundary | | Named authority | a studio, artist, repo or author known for it, searched by name | Tool per angle: - `mcp__exa__web_search_exa` — default. Describe the ideal page, not keywords: *"blog post breaking down a Roblox GFX render pipeline in Blender"* beats *"roblox gfx blender"*. - `mcp__exa__web_search_advanced_exa` — when a filter is doing real work: `includeDomains` for gallery/authority sites, `category` (`github`, `pdf`, `publication`, `personal site`, `news`), `startPublishedDate` for fast-moving domains, `excludeDomains` to kill a farm that keeps ranking. - `WebSearch` — a different index, US-only. Run it alongside Exa on at least one angle so an Exa blind spot becomes visible instead of silent. - `mcp__exa__web_fetch_exa` — batch every confirmed candidate URL into one call, not one call per URL. Stop at saturation, not at a target count. If angle 4 and angle 5 both return only URLs already seen, the vocabulary is exhausted and more queries buy nothing. ## 3. Visual references take a different path This is the common case and the one most often done wrong. Search tools return text. They cannot see an image. Four consequences: **Search for pages that host image sets, not for images.** Named galleries beat generic queries — target them with `includeDomains`: `artstation.com`, `behance.net`, `dribbble.com`, `cara.app` for art and GFX; `mobbin.com`, `land-book.com`, `godly.website`, `awwwards.com`, `siteinspire.com` for UI and web layout; `devforum.roblox.com` and `create.roblox.com` for Roblox specifically. **Look at what you found.** Download candidates to disk, then `Read` the image file — Claude sees images through Read. A reference that was never viewed is a link, not a reference. Never describe a style from a page's alt text, its title, or your own prior: that produces a plausible profile of a picture nobody looked at. **File them with provenance.** Images go in `ClaudeFX/Knowledge/Design//refs/`, each with its origin URL recorded. Then hand off to `image-dna` (extract mode) to turn refs into a style profile. That skill writes the rules; this one only supplies vetted inputs. **Handing refs to an external image model:** batch into labelled contact sheets — Gemini web caps at 10 attachments. See `image-handoff`. Roblox avatars come from the public thumbnail API, not from a search engine. ## 4. Reject before ranking Kill on sight: - Listicle content farms ("Top 15 …") that only link elsewhere — chase the outbound link, drop the page itself. - Undated pages in a domain that moves. - Anything whose claims trace to no primary source and no inspectable artifact. - AI-generated example galleries when the acceptance criterion was "how professionals actually do it". They match the description and miss the craft. Rank survivors: primary over secondary; an artifact you can open over a description of one; recent over old *only* where recency changes correctness. ## 5. Verify each survivor Fetch it. A search snippet is a claim about a page, not the page. Every named claim — a number, an API, a version, a technique, an attribution — is confirmed in the fetched text or dropped. A finding that survives on snippet alone ships labelled **unverified**, never silently. Same discipline as `verifying-agent-reports`. ## 6. Deliver — choose, don't ask | Situation | Output | |---|---| | ≤8 findings, one-off, no reuse | chat only: name, one-line why-it-survived, link | | Reusable link collection | `ClaudeFX/Sources//.md` — entries with per-entry verdict | | Visual refs meant to shape later work | files in `Knowledge/Design//refs/` + origin URLs, then `image-dna` | | Findings that harden into a rule | `ClaudeFX/Knowledge//`, tagged derived (traced to a ref) vs general (asserted) | | >15 findings, or feeds a decision | research note in `Planning//` | Any written form carries, per finding: what it is, source URL, date checked, why it survived, verified or unverified. Assets downloaded alongside a note live next to it, and the note names each file's origin URL. ## 7. Delegation Bulk fan-out goes to `crew-researcher` (has Exa, WebSearch, WebFetch). Hand it the step-0 frame verbatim — type, acceptance rule, stop condition — or it will return volume instead of matches. Verify every count and claim it reports before repeating them. Visual judgment does not delegate cleanly: a subagent can return links, but a report saying a reference "matches the style" is worth nothing unless something actually opened the image. Use `crew-vision` when many downloaded images need describing, and look yourself at the ones the decision rests on. ## Hands off to, doesn't replace `image-dna` (refs → style profile) · `gfx-prompt` / `image-handoff` / `image-request` (generating from the refs) · `/graphify` and `/vault-search` (questions answerable inside the vault — don't go to the web for those) · `customer-research` (evidence about people, not artifacts) · context7 MCP (library and framework docs — query it instead of web-searching them).