turn-the-crank - SKILL.md Agent Skill

name: turn-the-crank description: Generate a brand-new song by reading the whole skill rack and composing in one artist's style — chasing TIMELESS, catchy, instant-classic tunes, not pastiche. Use when the user says "turn the crank", "/turn-the-crank", "crank one out", "make a song in someone's style", or wants a fresh generated track for the album. Claude picks the artist. Bonobo-style layering crescendo; drums only if the song wants them, with a kit that FITS (Travis-Barker busy style is one option, not a default). Sparse→bloom arc (track 01 "dawn" is one example, not a required template — size the arc to the song).

You're turning the crank. One turn = one complete, finished song, composed in the style of one artist pulled from the rack, chasing a single north star: a timeless, catchy, instant classic — a tune someone hums on first listen and still loves in twenty years. Not a genre demo. Not a style pastiche. A song.

The north star (read it twice)

Every choice serves memorable + timeless:

The hook is the song. A melody you can hum with no backing. Write it FIRST, before any arrangement. If it isn't catchy a cappella it won't be catchy produced — make it VERY hooky or start over.
Simple, strong harmony. Timeless tunes ride simple loops (the Bonobo i–♭VI–iv–♭VII, a clean modal vamp). They resolve. Don't flex.
Restraint = timelessness. Space beats clutter; every voice earns its slot ([[style-bonobo]]). A track that's busy to sound clever dates fast — one that says one thing clearly lasts.
Emotional clarity. It should feel like one thing — yearning, triumph, calm — not a tour of techniques.

If a choice doesn't make it catchier or more timeless, don't make it.

Turn the crank (the process)

Grill first — capture the target ([[grill-me]]). Before reading the rack, run a focused grill (not a survey): what's the feeling/vibe, the energy (chill↔hype), any reference track or artist, and — if they want a video — the visual theme/style/reference image (a look from [[blender-styles]], or a picture to match). Resolve the real intent, then still pick the influence in step 2. The grill is what stops a blind pick: you compose toward a stated target, not a guess. If the invocation already gave a clear brief, grill only the gaps. Write the answers down (incl. any reference image, and — if a video is wanted — the reference→shapes/camera/palette/motion decomposition from [[blender-3d-concepts]]) so the compose and visual phases aim at them. Default energy when the grill doesn't say otherwise: dancy / upbeat (Adam, 2026-06-12: "in general I want more dancy upbeat kinda music") — treat chill/ambient/downtempo as the briefed exception, not the comfortable default.
Read the whole rack. ls skills/skills/, then read every style-* skill (your artist palette). Two craft skills are your menus — know them cold: [[strudel-sample-library]] is every instrument you can reach (the 128 General MIDI voices — real strings/winds/brass/choir/organ — plus VCSL mallets, world percussion, the amen breaks, ~50 drum kits, synths + FM), and [[strudel-modifiers]] + [[strudel-effects]] are every method (combinators, granular, filters, modulation). Also [[strudel-compose]], [[strudel-conduct]], [[strudel-weird]], [[strudel-pro-tips]] (the high-leverage shortlist), and the craft foundation — [[strudel-arrangement]] · [[strudel-harmony]] · [[strudel-melody]] · [[strudel-groove]] · [[strudel-mixing]] · [[strudel-sound-design]] · [[strudel-genres]] · [[strudel-sampling]] · [[strudel-automation]] · [[strudel-bass]] · [[strudel-texture]] · [[strudel-emotion]] (the music-theory + production fundamentals behind a track that lasts). You can't pick well without seeing the whole rack.
Pick — your call, but check the log first. Read skills/skills/turn-the-crank/crank-log.md. Hard rule: don't reuse any artist from the most recent run — never the same pick back-to-back. Soft rule: skim the whole log and lean toward the artists you've cranked least, or never — that's how all sixteen get their turn. From what's left, choose one artist as the lead voice AND one specific song of theirs as the reference — pull it from that artist's skill (its analyzed-tracks table or Adam's hearts). Base the track on that ONE song's DNA — its key, tempo, structure, groove, register, dynamics — don't aggregate the artist's catalog into a generic style-average (that's what made flux muddy: it blended four FP tracks). One song = a focused, authentic target. The DNA is a fingerprint, not a melody — it gives you the frame; you still have to write the tune (step 4). A flavor artist, if used, is also anchored to one of their songs — one song each, never a catalog blend. State in one line why this artist + song, now, and what's timeless about it. Energy rule ([[feedback-station-energy]], standing 2026-06-12): pick danceable/upbeat references by default — the station should be nearly all dancy (Adam: "i want nearly all of these songs to be dancy"). Downtempo is allowed only when it GROOVES and CLICKS (Adam: "if it's downtempo it needs to groove and click, not be boring sad downtempo"): a real pocket — swung head-nod drums, a bass riff locked/syncopated against the kick, momentum you can nod to. Think Kong-pocket (thrum), dub 2-step (tarn), boogie (sprite), boom-bap — NOT beatless melancholy drift (the cinder/wraith/glade lane is now a briefed-exception only, never a dealer's-choice pick). Say explicitly in the pick line how the groove clicks.
Ground it in the data — and the instruments. Each style-* skill carries measured DNA (key, BPM, scale-degree emphasis, chord loops, contour) AND the voices that define the artist. Use the data for the frame — key, tempo, register, dynamics, the chord loop, the rhythmic feel. But the measured contour/seed is a statistical fingerprint, NOT a melody — do NOT transcribe it into your hook (that's how cobalt turned Labyrinth's +octave ×216 into a gimmicky a4–a5 tremolo instead of a tune). Let the data set the frame; compose a real, original, hummable line inside it. And reach for the real instruments from [[strudel-sample-library]] — gm_string_ensemble_1/gm_cello for Kiasmos & Rone strings, gm_flute for David Wise, kalimba for Bonobo, gm_sitar/mridangam/gm_alto_sax for DJRUM. Don't reuse the last crank's signature voices (the log remembers), and don't keep defaulting to "sawtooth + one mallet." The point isn't strange — it's the right, varied voice for a good song.
Find the hook first. Before a single drum, write the central melody in the chosen artist's melodic DNA. Hum-test it. Make it feel inevitable. This is the part you cannot fake later.
Build it in the station's own genre — "toaster house" — not the reference's. ([[feedback-station-energy]], Adam 2026-06-12: "take inspiration … then make our own genre instead of trying to make sad drawn out bloom.") The reference supplies the frame (tempo, groove feel, energy); the genre is OURS: groove-first (the pocket lands by section 2 — kick + octave-displaced bass locked together — not after a minute of mist), hook-forward (a hummable riff that RIDES the groove), warm-dark (capped lpf ceilings, no shrill highs), ≤3 min. Default architecture = club arc: in → pocket locks → hook → lift → breakdown that still moves (strip the kick, the bass/arp keep grooving — never dead air) → drop/peak → DJ-out. Develop by adding/subtracting around a constant pocket; call-and-response between voices still applies ([[style-bonobo]]). The old sparse→bloom layering crescendo (dawn's spine) is now the briefed-exception architecture for explicitly-requested ambient pieces — not the dealer's-choice default.

"Toaster house" is a FAMILY of grooves, not ONE groove (Adam 2026-06-14, after krypton v1 came out a near-clone of tungsten: "they're too similar"). The genre identity lives in the values above (groove-first, hook-rides-pocket, warm-dark, ≤3 min) — NOT in a fixed template. Do not reach for the same machinery every time. The tungsten skeleton (straight bd*4 + octave-displaced triangle bass + dub-delayed saw arp + warm-triangle hook) is one groove; having nailed it is exactly why it's tempting to reuse — resist. Each crank must vary the actual machinery, anchored to its own artist's signature: the kit (4×4 / syncopated-funk / half-time / broken 2-step / garage shuffle / boom-bap), the bass character (octave-triangle / FM type-bass slap / acid-resonant / sub-only / plucked), the lead timbre (warm triangle / brassy FM square / supersaw / bell), and whether there's an arp at all. Worked example: krypton (Sonic) had to be rebuilt off tungsten onto real Genesis machinery — FM type-bass as the lead, syncopated kit, PSG chip sparkle — to stop sounding like tungsten in a different key. Pull the groove from the style-* card, not from the last dance cut. Keep the whole track under ~3 minutes (Adam's standing preference — a tight, repayable tune beats a sprawl; it also serves the timeless/catchy north star). Size the section cycle-counts so the full arc lands ≲3:00 — at ~~130 BPM/4/4 (~~1.85 s/cycle) that's ≈96 cycles total; fewer at slower tempos. Mind the seams — gain-stage the arc so the peak has somewhere to arrive from, and bridge each section change ([[strudel-transitions]]).
Drums only if the song wants them — and only the KIND of kit that fits. Many of these artists are drumless/ambient; honor that. When a song does want a busy, driving kit, the Travis-Barker recipe below is one strong option — but it is not a default. A mid-tempo, mysterious, spacious, or delicate tune usually wants a simpler kit, a washed/clean one, or none. Match the kit to the song (the log's repeated "NOT Travis-Barker" departures are the norm, not the exception).
Write + sound + test. Compose with [[strudel-compose]] / [[strudel-conduct]], sound it with [[strudel-sample-library]] + [[strudel-effects]], shape it through the chosen style-* lens, and test it ([[strudel-test]]). A track that doesn't parse is worse than none.
Iterate until it is actually satisfying — metrics are not enough. See references/quality-gate.md. A crank is not "done" just because it parses, renders, or lands near the reference card. After every render/measure pass, make a musical judgment against the north star: is the hook memorable, does the groove feel good, does the arc create desire, would someone ask to replay it? If the answer is no, keep iterating. If metric-driven patches only make the numbers better while the song still feels bad, stop polishing and rewrite the musical idea — hook, groove, arrangement, or even artist/reference — rather than shipping a bad song with good-looking readouts. Do not present a compromise as finished.
Close the proof loop before micro-tweaks. See references/shipping-discipline.md. A crank is not shipped unless the exact final files have been parse-validated, rendered to WAV, measured, logged, and judged musically satisfying. If tool/context budget is getting tight, stop aesthetic tweaks and ship the current verified version only if it is also satisfying; otherwise call it a draft or scrap/restart it. Don't make a last patch that leaves the final state newer than the last render/metric pass.
Ship the song + log it. Fresh generated cranks for this repo belong in the active generated station by default: tracks/v2-gen/<NN>-<name>.strudel plus tracks/v2-gen/<NN>-<name>/ section arc (~6–8 sections, sized to the song — dawn's layout is a reference, not a template to copy). The root .strudel file is intentional: it is the full playable/live working copy discovered by /tracks, while the matching directory holds NN.strudel section files discovered by /sections?track=<id>. Do not delete or move the root file just because a same-named folder exists. If the user asks for a different group, move the pair together (e.g. tracks/ep/<name>.strudel and tracks/ep/<name>/); use loose top-level tracks/<name>.strudel only when explicitly requested. Before declaring the folder structure correct, verify the pair exists in the intended group, the section directory has only expected section files plus _changelog.md/optional manifest or assets, and compare against existing paired tracks such as tracks/v2-gen/crank-halo.strudel + tracks/v2-gen/crank-halo/ or tracks/ep/01-dawn.strudel + tracks/ep/01-dawn/. Name it for the vibe. Append one row to skills/skills/turn-the-crank/crank-log.md (date · track · lead artist · the reference song · flavor · the hook) so the next crank can steer clear. Also write a per-song change log tracks/v2-gen/<NN>-<name>/_changelog.md — one entry per ears-verified iteration capturing intent → what/where → why → measured before/after (the render→measure readout), plus a target row from the reference card and any layout/move proof if paths changed. This is the eval corpus: the (intent → edit → measured outcome) trail that lets us evaluate the compose/iterate process later. Keep appending to it on every future edit pass to that song. Then tell the user: who you cranked, why, what the hook is, the iteration count, why the final pass is musically satisfying, and the final render/metric proof.
(If the grill asked for a video) Turn the visual crank — same discipline, visual. After the song ships, generate its music video: (a) pick a look from [[blender-styles]] to fit the song + the grilled theme (PS1 / toon / watercolor / generative / abstract — don't default to PS1); (b) read [[blender-3d-concepts]] FIRST and write the reference→shapes/camera/palette/motion decomposition into renders/<id>/README.md — getting the SHAPES + frame-fill + projection right up front is what turns a day of art-direction thrash into minutes (it's why dawn took a day: the shaders were fine, the shapes/scale/projection weren't); (c) build on tools/blender_style_kit.py with features from tools/audio_features_for_blender.py; (d) probe → look → fix via [[blender-video-iteration]] (render a still, actually look at it, frame_check the projection, fix the SHAPE before any detail), then full render + audio mux per [[blender-music-video]]. Analyze/log the video like the song (artifacts + a README post-mortem in renders/<id>/). Most cranks are audio-only — only do this when the grill asked for visuals.

The crank log

skills/skills/turn-the-crank/crank-log.md is the memory between turns. Read it before you pick (step 2), append to it when you ship (step 8). Two rules:

Never reuse the most recent run's artists — no back-to-back repeats. (Hard.)
Lean toward the least-used artists across the whole log, so all sixteen cycle through over time. (Soft.)

The log is local (gitignored). If it's missing, this is run #1 — pick freely.

Travis Barker drums (one option — only if the song wants this energy)

Reach for this only when a busy, driving kit genuinely fits the song (not by default — see step 6). When it does fit, it's busy, syncopated, unconventional — but locked to a half-time backbeat so it grooves. Ghost notes everywhere, fills that turn the corner. Always .bank() — bare s("bd") is silent ([[strudel-sample-library]]).

stack(
  s("bd ~ ~ bd ~ ~ ~ bd ~ ~ bd ~ ~ ~ bd ~").bank("AkaiLinn").gain(0.7).lpf(2200),   // syncopated kick, off the grid
  s("~ ~ ~ ~ ~ ~ ~ ~ sd ~ ~ ~ ~ ~ ~ ~").bank("AkaiLinn").gain(0.6),                  // half-time backbeat on 3 — the anchor
  s("~ ~ sd ~ ~ sd ~ ~ ~ sd ~ ~ sd ~ ~ sd").bank("AkaiLinn").gain(0.16).lpf(3500),    // GHOST snares — the signature
  s("hh*16").bank("AkaiLinn")
    .gain("0.3 0.12 0.18 0.12 0.26 0.12 0.18 0.14 0.3 0.12 0.18 0.12 0.26 0.14 0.2 0.16").hpf(3000),  // relentless 16ths, accented
  s("~ ~ ~ ~ ~ ~ oh ~ ~ ~ ~ ~ ~ ~ oh ~").bank("AkaiLinn").gain(0.26).hpf(2000),       // open-hat off the beat
  s("<~ ~ ~ [ht ht mt lt]>/4").bank("AkaiLinn").gain(0.5),                            // tom fill every 4th bar
)

The point isn't more notes — it's notes in surprising places that still land on the groove. Locked, never erratic. Scale gain/density to the section (sparse intro → full payoff).

Style reframe / rescue passes

If a crank is close but the style frame is wrong, do a reframe, not an automatic rewrite. Preserve any hook the user likes, then change the harmony, instrumentation, texture, and section arc around it. For wintery/snowy VGM themes, see references/winter-vgm-wise.md: David Wise-style coldness often comes from an icy arpeggio engine, restrained snow/noise, chromatic maj7/dominant color, and a continuous hummable long-line hook rather than bracket-spliced phrase tiles.

What "timeless" is NOT

A pile of layers with no tune. Layering serves the hook; it never replaces it.
Technical flexing — odd meters, dense fills, rare scales — for its own sake.
A faithful genre pastiche with no song underneath. The artist's style is the lens; the catchy tune is the subject.

When in doubt, ask: would someone hum this in the shower? If not, go back to the hook.