hermes-agent/optional-skills/creative/auteur/references/recon.md
teknium1 7eae49c499 fix(skills): auteur follows the modern section order and routes assets through image_generate
SKILL.md is restructured to authoring standard 5 (When to Use,
Prerequisites, How to Run, Quick Reference, Procedure, Pitfalls,
Verification) — headings only, upstream body text kept. The asset
references still carried the upstream per-CLI routing tables and command
lines for other agent products; those are replaced with the native
`image_generate` route (product names are allowed only in LICENSE and
credit lines). `WebSearch` residue in recon docs/refscout becomes
`web_search`.

source.mjs wrote its ~2.6MB Google Fonts metadata cache to
$TEMP||$TMPDIR||'.', which is the project cwd on most Linux shells; it
now uses os.tmpdir() and Pitfalls documents the location. Network-access
note now mentions that moodboard.mjs also downloads the image URLs the
search hosts return. Docs page regenerated for this skill only.
2026-09-15 04:03:43 -07:00

12 KiB
Raw Permalink Blame History

recon.md — scouting live references and building a moodboard

Taste is not invented at the desk. Every director watches films before shooting one, and the reflex table in taste.md only tells you what to avoid — recon tells you what is currently alive. Two commands do the gathering; you do the reading. Both write into design/, both are cheap, both are bounded: recon is a short pass at the top of phase 0, not a research project.

command answers
A node scripts/refscout.mjs how do the best sites in this space actually work? — live sites, their real stack, their scroll mechanics, screenshots
B node scripts/moodboard.mjs what should this feel like? — palette, light, composition, type energy, texture

Run A when the brief has a mechanic question (direct register, "wow", scroll choreography). Run B when the brief has a look question (any register with an undecided art direction). Most briefs want both, in that order, ~5 minutes total. Skip either without guilt when the brief already fixes that axis (existing brand book → skip B; "boring docs site" → skip A).

Both need playwright (npm i -D playwright && npx playwright install chromium). No API keys, no logins, no paid services.


A. refscout — take the references apart

# harvest what is winning right now, profile it
node scripts/refscout.mjs --from awwwards --limit 8

# same, filtered by the gallery's own category or a text search
node scripts/refscout.mjs --from awwwards:scrolling --limit 6
node scripts/refscout.mjs --from awwwards --search "coffee" --limit 5

# or profile sites you already know / found with web_search
node scripts/refscout.mjs https://a.com https://b.com --shots 4

Output lands in design/refs/: REFERENCES.md (read it), refs.json, shots/*.png.

Finding candidates. --from awwwards is the only harvester wired in, because it is the only gallery whose listing and detail pages render reliably headless and expose the outbound site URL. For everything else — godly.website, curated.design, minimal.gallery, land-book, siteinspire, thefwa, lapa.ninja, mobbin — use web_search to find the write-ups, then pass the site URLs to refscout positionally. Searching for the page about a site is more reliable than scraping the gallery that lists it.

Naming a technique you can see but can't name. originkit.dev is a catalogue of ~160 motion effects, each with a live preview and a name — useful when a reference does something you want to describe in the commit-sheet and have no word for. Browse it, take the vocabulary, then build the thing yourself; the components are React/framer-motion behind a signup, and half the catalogue is the exact drop-in ornament slopscan exists to keep out.

Reading a fingerprint

Each entry reports what the page actually loaded and did, not what its marketing says:

  • stack — libraries found in the JS the page really fetched. Bundlers hide globals, so this is matched against bundle text; treat it as strong evidence, not proof. GSAP + ScrollTrigger + ScrollSmoother + Lenis is the house style of the entire awwwards top tier — seeing it for the fifth time is the useful signal, not a coincidence.
  • mechanics — the derived read: pinned scenes (counted from .pin-spacer elements ScrollTrigger leaves in the DOM), WebGL driven by scroll, sticky stacks, CSS animation-timeline, mix-blend layers, custom cursor, scrubbed video. This is the column you are actually shopping in.
  • page shape — 22× viewport tall · 14 sections tells you the scroll budget the reference spends. A 22× page with 4 pinned scenes is a different film from a 3× page with one WebGL hero, and the difference is a decision you have to make too.
  • type / palette — the fonts and colours as painted, sampled from computed styles. LayGrotesk + PPNeueMontrealMono, Thunder + PP Fraktion Mono — this is where you learn that the top tier is not running Inter, which is exactly ban #8 with receipts. The display face is measured as the largest painted text (Thunder, sans-serif 118px/800), never the first h1, because a hidden or fallback-styled heading reports whatever the cascade left there and will cheerfully claim an awwwards winner ships Inter. The px figure is evidence too: this tier runs display type past 100px.
  • shots — hero plus scroll stops at 1440. Look at every hero. Look at the remaining stops only for the 2–3 sites you actually take a steal from. At --limit 8 --shots 3 that is 23 images and reading all of them is not triage, it is a context bonfire. The numbers tell you the machinery; your eyes decide whether it is beautiful — but only for the references you are actually using.

⚠ "NO CAPTURE" entries

Some sites hand a headless browser the server-rendered HTML and then never deliver their CSS or JS (edge-streamed pages, bot walls). Anything read off such a page — fonts, colours, stack — would be Chrome's defaults dressed up as findings, so refscout reports nothing for them rather than something false. Roughly 1 in 8 award sites lands here. When it does: open the URL yourself, or describe it from the gallery write-up, or drop it. Never quote a fingerprint the tool refused to give.

The steal rule

Every entry has a **steal:** line and it is not decoration — fill it in, one line, before you close the file:

"madewithgsap.com — the section title stays pinned while its cards scroll through it; we do this once, on the proof scene, with the product name instead of a title."

Rules, in order of how often they are broken:

  1. One idea per reference, named as a mechanic — "pinned title, content scrolls through", not "cool scroll effect", and never "the vibe".
  2. Adapted, or it is theft. A reference is a starting camera position, not a set. Changing the colours of a copied layout is not adaptation.
  3. Two references maximum feeding one scene. Three is a collage, and collages read as slop.
  4. Never cite a reference you did not look at. A fingerprint without eyes on the shots is half a reference.
  5. Named brand assets — logos, custom typefaces, photography, illustration — are theirs. Mechanics and structure are free; identity is not.

Feeding the commit-sheet

Recon changes what you write in the commit-sheet, especially field 6:

  • The stacks and mechanics you saw five times ARE the first-order reflex (6a) for this category. If your peak is the fifth pinned-WebGL-hero of the day, you found the mode, not an idea.
  • The type and palette columns give you a real, dated map of what the category currently looks like — much sharper than reasoning about it from memory.
  • Any mechanic you take goes into the motion budget (field 5) as one of the ≤3 families, not on top of it.

In the direct register, put the 2–3 references and their one-line steals in the STORYBOARD header (References taken) so the decision survives into the build.


B. moodboard — decide what it should feel like

node scripts/moodboard.mjs "editorial brutalist layout dark" "high contrast type poster" --limit 24
node scripts/moodboard.mjs "terracotta ceramic studio light" --source bing,arena --limit 16

Output lands in design/moodboard/: contact-sheet*.png (look at these), numbered img/NN.jpg, and MOODBOARD.md mapping every tile number to its origin page. Twenty tiles arrive as one image, so reading a moodboard costs one look, not twenty.

Sources (all no-auth, tried in this order and interleaved so none owns the sheet):

  • bing — the workhorse. Bing's image index reaches into Pinterest, Dribbble and Behance CDNs and serves originals that hotlink fine through a browser context.
  • pinterest — logged-out search renders a partial grid behind the login wall: sometimes ~15 pins, sometimes zero, day to day. Genuinely useful when it answers, never load-bearing; when it returns nothing the script says so and bing has already covered it.
  • arena — the public are.na search API. Lower volume, highest curation: these are blocks designers saved for themselves.

Query craft

The quality of a moodboard is decided entirely by the queries. Three rules:

  1. Two or three queries on different axes, never one. A single query returns near-duplicates of one template. Pick axes: subject/medium ("ceramic studio photography"), treatment ("hard rim light, deep shadow"), graphic language ("swiss grid poster, red accent").
  2. Search the feeling, not the category. "coffee website" returns other coffee websites — the category reflex, delivered to your desk. "steel and steam, industrial macro, warm shadow" returns material you can actually direct from.
  3. Check the anchor noun is not half of a fixed compound. "polished brass instrument macro" returns trumpets and saxophones, because brass instrument is one word to a search index. Same trap: hard surface, light bulb, glass ceiling, sound board, steel drum. The noun is supposed to ground the query, and an idiom hijacks it instead.
  4. Anchor a treatment query with a concrete noun. Image search collapses an abstract phrase to its most commercially indexed substring: "hard rim light on dark glass, wet stone, near-black macro" came back as hard surface — hi-vis workers, gravel, granite pavers. "wet black stone, single hard light" does not. When one query of three drifts, that is the mechanism.
  5. Never search a brand you intend to resemble. That is the shortest path to a page that looks like a competitor with the logo swapped.

Reading the sheet

Fill the four lines at the bottom of MOODBOARD.md — the moodboard exists to produce these, and an unfilled index means the sheet was decoration:

  • dominant palette across the tiles you liked → commit-sheet field 2, converted to OKLCH. Take the relationship (drenched single hue / near-mono with one accent / earthy midtones), not a literal eyedropper.
  • light and contrast character → the lighting: line of every scene-sheet and the literal lighting parameter in generated-asset prompts (assets.md §1).
  • one composition move worth stealing → commit-sheet field 4 (grid break).
  • what to explicitly avoid from these → a real answer, because a sheet always contains the category reflex too, and naming it is how you stop drifting into it.

Expect 10–20% junk per sheet. It usually is not spread evenly: one query drifts wholesale while the others stay clean, so read the junk as a signal about that query, not about the sheet. Ignore the strays; re-run only if a whole axis came back wrong, and re-run that axis with a concrete noun.

Reference images are not assets

Downloaded images are direction only. They are other people's work, they are not licensed to you, and none of them ships:

  • never place a downloaded image in the page, not even as a placeholder that "we'll swap later";
  • never reproduce one composition — the moodboard sets palette, light and energy, and the actual frames get generated per assets.md;
  • when the art direction is locked, the moodboard's job is done: it informs the generation prompts and then stays in design/.

design/moodboard/ and design/refs/ are working artifacts. Ship them to no one and add them to .gitignore if the repo is public.


When recon fails

symptom what it means do this
awwwards listing returned no cards the gallery markup moved fall back to web_search + positional URLs
most entries are NO CAPTURE network or a broad bot wall profile fewer, better-known sites; lean on the shots and your own eyes
pinterest returned 0 login wall, normal ignore, bing covered it
the sheet is 20 variants of one image single narrow query re-run with 2–3 queries on different axes
one query's tiles are a different subject entirely search collapsed an abstract phrase to a common substring re-run that axis anchored to a concrete noun
a script looks hung for minutes you piped it through tail/head both stream progress to stderr; drop the pipe to watch it
references all look alike you found the category mode that IS the finding — write it into reflex-check 6a and deviate

Recon is bounded. If it has not produced a usable read in one pass, stop and decide from taste.md — a director who cannot find a reference still has to shoot the film.