Generative Art Coach

Coach the creation of generative art from first idea to curated series. The craft insight this skill enforces: generative work succeeds on constraint, not on randomness. Randomness is seasoning; the system is the dish. So the process starts deterministic, adds chance one parameter at a time, keeps every output reproducible by seed, and ends with ruthless curation — because a generative artist's real medium is the ability to throw most outputs away.

When to use this skill

The user wants to create generative, procedural, or algorithmic art and needs direction
A creative-coding sketch produces outputs that all look samey or all look like noise
Someone asks how to turn one nice accidental output into a coherent series
A plotter drawing, album cover, print series, or animated piece needs a generative system behind it
The user says "make art with code" and needs the process, not just a script

Instructions

Start from a constraint, not a technique. Pick one system — grid subdivision, flow fields, circle packing, recursive splitting, wave interference, random walks — and one rule that will not be broken, such as "only horizontal marks" or "no more than two shapes". The rule is what gives the series a voice.
Build the system deterministically first: randomness off, every parameter fixed, black marks on white. If the bones are not interesting in monochrome with no chance involved, more randomness will not save them.
Introduce randomness one parameter at a time, and only through a seeded generator. Never call an unseeded random source: every output must be reproducible from its seed, and the seed belongs in the filename or metadata of every export.
Give each random parameter a range and a reason. "Stroke weight varies 1 to 4 to suggest depth" is a decision; "everything varies" is abdication.
Constrain color early. Choose three to five colors and assign roles — background, structure, accent — then sample from the palette with weights. Forbid per-element random hue; it is the fastest route to visual mud.
Explore parameter space methodically: vary one axis per batch, render nine to sixteen outputs, and review them as a contact sheet. Name what changed and what it did before touching the next axis.
Curate hard. Keep roughly one output in ten. Select for pieces that share a family resemblance yet differ meaningfully — a series, not a lottery. Record the seed and parameters of every keeper.
Finish properly: generous margins, consistent export size, print resolution for raster or clean paths for SVG, and an edition note listing system, seed, and date.

Craft principles

Structure carries the piece; randomness only decorates it
Constraints create style — the narrower the rule, the more recognizable the series
Edges and margins are part of the composition, not leftovers
An output you cannot reproduce is a screenshot, not a work in a system
When everything surprises, nothing does; aim for one controlled surprise per piece

Output format

For a coaching session, deliver: the system statement (one paragraph naming the technique and the unbroken rule), the parameter table with ranges and reasons, the code, and a curation note explaining which seeds made the cut and why. For code, prefer SVG or a canvas sketch with a visible seed constant at the top.

Worked example: parameter sweep plan

system:  recursive grid subdivision, rule = "split only vertically or horizontally"
seed:    runs 1000-1015
batch 1: vary split probability      0.3 -> 0.9   (find where density feels alive)
batch 2: vary minimum cell size      2% -> 12%    (find where detail becomes noise)
batch 3: vary palette weights        accent 5% -> 25%
freeze:  best values from 1-3, then render 40 seeds and curate to 6

Four disciplined batches beat four hundred random renders, and the artist can say why every keeper looks the way it does.

Common failure modes

Random hue per element — the fastest route from system to soup
Technique tourism: switching systems every session instead of exhausting one
Zero margins, so the composition bleeds meaninglessly to the canvas edge
Unseeded randomness that makes the best output of the night unrepeatable
Keeping everything, which turns a series into a folder of noise
Adding a new parameter when the honest fix is removing one

Quality bar

Every export reproducible: seed plus parameters recorded with the file
A stranger could sort the curated series from the rejects and mostly agree with you
The series is recognizable as one system across all keepers
No unseeded randomness anywhere in the code
The piece survives being printed in grayscale — structure first, always