agentleFS
DeiblerDeibler/edmund-harness/skills/iterative-design/SKILL.md

iterative-design

Skill24 starsChanged 2 months agoMITView on GitHub

Use this skill whenever the user is iterating on a visual — flat graphics, logos, badges, t-shirt prints, hockey/sports graphics, or subtle photo edits across multiple rounds. Triggers on phrases like "try it with", "move it", "make it bigger", "change the color", "not quite right", or when the user sends back a marked-up or regenerated image. Keeps an editable SVG source of truth between rounds, renders PNGs for iMessage delivery, generates parallel candidate variants for fuzzy feedback, overlays labeled grids when the user "can see it but can't describe it", and chains with the `generation` skill for AI illustration layers. Also use when starting any new design destined for screen print (t-shirts, posters, banners).

The file

We split this skill into the rules it gives an agent

12 rules. About 0 are common, 5 are distinctive, and 7 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 7

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Use this when someone is iterating on a visual.

    line 8
  2. Very specific:

    All installed and on PATH.

    line 20
  3. Very specific:

    Color trace: use trace_color.sh (vtracer).

    line 43
  4. Very specific:

    Background removal: use remove_bg.sh (rembg). Never magick -fuzz or -transparent white for anything non-trivial — fuzz-based removal eats …

    line 44
  5. Very specific:

    Before tracing a photo: run remove_bg.sh first. Tracing a raster with its white field intact gives you a 2000x2000 white polygon behind the …

    line 45
  6. Very specific:

    Write the brief to design-<slug>/brief.md so round 5 doesn't drift.

    line 56
  7. Distinctive:

    Before declaring done, run three checks:

    line 98
  8. Very specific:

    Use svgo_final.sh only for the *final* export — it merges paths and can drop the anchors you used between rounds.

    line 106
  9. Distinctive:

    Keep both the clean and the distressed version — some printers want the clean SVG, not the grunge PNG.

    line 123
  10. Distinctive:

    When the user says "vintage collegiate" or "like a Salty Dog" and the reference isn't already in the sandbox, do one short research pass …

    line 137
  11. Distinctive:

    Don't research for five rounds — do it once, up front, and move on.

    line 143
  12. Distinctive:

    If the user sends a copyrighted image as a reference expecting you to reproduce it verbatim, say so out loud: "I can do a design inspired …

    line 169

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

iterative-design

Use this when someone is iterating on a visual. One-shot generation goes to image-gen or generation. This skill owns the multi-round loop: source-of-truth between rounds, decoding fuzzy feedback, changing one thing without regenerating everything, and landing on a final that screen-prints cleanly.

Modes — pick one at round 1, stick with it

SVG mode for anything with text, geometric shapes, circular typography, borders, or destined for screen print / vinyl cut. SVG is the only sane way to do 15 rounds of "5px left, wrong color, go back to round 2." Pixel regenerations drift; SVG edits don't.

Raster mode for photograph edits and scene composition. Delegate the pixels to the generation skill; this skill owns the loop (versioning, feedback decoding, variant fans).

Mixed (illustrated badge with an AI mascot): SVG mode with the raster as an embedded <image href="..."> layer. Regenerate only the illustration when swapping the mascot; the badge structure stays stable.

CLI toolkit

All installed and on PATH. Reach for the right tool on the first try — the common failure mode is defaulting to potrace or magick -fuzz for everything and shipping washed-out output.

NeedTool / script
Vectorize a color raster (logo, badge, multi-color mark)scripts/trace_color.sh (vtracer)
Vectorize B&W line art, silhouette, hand sketchscripts/trace_sketch.sh (potrace)
Remove background from a photo, keep interior colorsscripts/remove_bg.sh (rembg / U²-Net)
Photoshop-style filter on a raster (sepia, pencil, quantize)scripts/stylize.sh (gmic)
Stylized geometric-shapes SVG from a photo (triangles, ellipses)scripts/photo_to_svg.sh (primitive)
Render SVG → PNG (fast, most cases)scripts/render_svg.sh (rsvg-convert)
Render SVG → PNG (accurate — gradients, filters, complex files)resvg in.svg out.png
Edit SVG programmatically (boolean ops, text→path, flatten)inkscape --actions='...'
Clean SVG between rounds (safe, preserves IDs)scripts/svg_optimize.py
SVG final export optimization (aggressive multipass)scripts/svgo_final.sh (svgo)
Apply distress / grungescripts/distress.py
Labeled grid overlay for "I can see it but can't describe it"scripts/annotate_grid.py
Side-by-side variant pickerscripts/variant_grid.py
Detect user-drawn markup on a returned imagescripts/detect_markup.py
Optimize final PNG for deliveryoxipng -o max --strip safe file.png
Inspect/strip metadataexiftool

Tool-selection rules

  • Color trace: use trace_color.sh (vtracer). Never potrace — it is B&W only and will collapse all color to a single fill. This is the Warriors/Bears-Cup-logo failure mode.
  • Background removal: use remove_bg.sh (rembg). Never magick -fuzz or -transparent white for anything non-trivial — fuzz-based removal eats interior regions that happen to match the background (red logo on white → red shadows inside the logo also vanish).
  • Before tracing a photo: run remove_bg.sh first. Tracing a raster with its white field intact gives you a 2000x2000 white polygon behind the subject.
  • Render before delivery: SVG → PNG always. iMessage can't preview SVGs, and .svg attachments arrive as .txt.

The four phases

1. Brief (round 0, ~30 seconds)

Before touching the canvas, ground the design. Read references/brief-template.md for the question set. For a repeat user like Casey, most of the brief is already in his person file — palette preferences (vintage navy + cream), format (flat on white, no mockup), and style (distressed collegiate). Still confirm anything ambiguous with one multi-choice question rather than a freeform ask:

"Going full distress like the Salty Dog, or cleaner like the pirate one?"

Write the brief to design-<slug>/brief.md so round 5 doesn't drift.

2. Explore (first delivery, parallel candidates)

For a fresh design, generate three candidates in parallel with distinct interpretations, not three variations on the same interpretation. Compose with scripts/variant_grid.py into one image and deliver. One round of "which direction?" beats three rounds of guessing at intent.

For an iteration on an existing design, skip this phase and go straight to Refine.

3. Refine (rounds 2..N)

The loop:

receive tweak → interpret → edit source → render → deliver

Interpret — the tweak arrives as text, a marked-up image, a silent new photo (a visual target), or nothing at all (they just sent a new reference). For text tweaks see references/feedback-playbook.md. For marked-up images, run scripts/detect_markup.py to locate the region.

Edit, don't regenerate. In SVG mode: change the one attribute — transform, fill, stroke-width, a textPath content string. In raster mode: pass the previous round (not the original) as reference_images so the AI is editing what they saw, not starting over.

When the user sends a raster and asks for an SVG. Two steps, always in this order:

bash scripts/remove_bg.sh input.png design-<slug>/trace-src.png    # if bg needs to go
bash scripts/trace_color.sh design-<slug>/trace-src.png design-<slug>/source.svg

Then render_svg.sh and deliver. If the user asks to "clean up" or "make it editable" after that, run svg_optimize.py — svgo_final.sh is for the final export only.

When the user asks for a style variation on a photo (oil paint, sepia, pencil), use stylize.sh rather than regenerating — it preserves the subject exactly and just treats the pixels.

When genuinely stuck — two-round rule. If you've asked "do you mean X or Y" once and missed, switch modes:

  • scripts/annotate_grid.py overlays labeled cells (A1–D4) on the last round. Send back and ask "which cell is off?" Users can nearly always point, even when they can't describe.
  • scripts/variant_grid.py composes 3 focused candidates. Use when you have a guess but want to cover adjacent options.

Render. SVG → PNG is not optional before sending; iMessage can't preview SVGs.

bash scripts/render_svg.sh source.svg rounds/NN.png --width 2400

View before you send. Always Read the rendered PNG yourself before delivering. This is the step that catches "potrace flattened the Warriors red to black," "rembg took a chunk out of the beard," or "the SVG has a 2000px white rectangle behind the logo." One visual check per round is cheaper than one round of user confusion.

4. Export (final round)

Before declaring done, run three checks:

bash scripts/svg_validate.sh source.svg                    # well-formed XML / SVG structure
bash scripts/svgo_final.sh source.svg source.min.svg       # aggressive multipass optimization
bash scripts/size_check.sh source.svg rounds/size-strip.png # legibility at 2400/800/400/128

Use svgo_final.sh only for the final export — it merges paths and can drop the anchors you used between rounds. For in-loop cleanup stick to svg_optimize.py.

Optimize the PNG delivery too:

oxipng -o max --strip safe rounds/final.png

For screen print, also run:

python3 scripts/printability_check.py rounds/NN.png

Surface any warnings to the user before they send to the printer. Then apply distress if that's the style:

python3 scripts/distress.py rounds/NN.png rounds/NN-distressed.png --intensity 0.4

Keep both the clean and the distressed version — some printers want the clean SVG, not the grunge PNG.

Chaining with generation

For the raster illustration inside an SVG badge, or for pure photo edits, delegate to the generation skill. Write the prompt with three pieces — see references/prompt-templates.md for the full structure:

  1. Dimensional: "for a 12-inch t-shirt print at 300 DPI" or "fills a 1040x1040px circle"
  2. Aesthetic: "flat 2-color vector style, navy on cream, no gradients, distressed collegiate"
  3. Technical: "transparent background, no text, no border — the SVG will add those"

A prompt that skips the technical piece is how you end up with an AI-rendered rope border fighting your real rope border.

Research when you need reference

When the user says "vintage collegiate" or "like a Salty Dog" and the reference isn't already in the sandbox, do one short research pass before generating:

  1. WebSearch for the style/era ("vintage collegiate varsity badge", "1970s surf shop logo").
  2. WebFetch 2–3 images into design-<slug>/refs/.
  3. Pass them to generation as reference_images.

Don't research for five rounds — do it once, up front, and move on.

Sandbox layout

sandbox/<thread>/design-<slug>/
  brief.md              # palette, format, style, any user constraints
  source.svg            # SVG mode: edit this between rounds
  illustration.png      # AI raster layer referenced by source.svg
  refs/                 # inspiration images, from web research or user uploads
  rounds/
    01.png              # every delivered output — no exceptions
    02.png
    02-distressed.png
    03.png
  notes.md              # "round 2: rope thicker, liked crab, rejected pirate"

notes.md pays for itself on round 5 when the user says "go back to the one with the rope" and you need to remember which round that was.

IP safety

Sports-team and institutional graphics (Warwick hockey, Bears Cup, Ocean City tourism logos) often ride on registered trademarks. Rules:

  • Reproducing an official team mark, logo, or wordmark is a no — even as a "reference" that the user will "just use personally." Flag and decline.
  • "In the style of" a well-known brand (Patagonia, Stüssy, college varsity marks) is fine as long as the output is a fresh composition, not a trace.
  • If the user sends a copyrighted image as a reference expecting you to reproduce it verbatim, say so out loud: "I can do a design inspired by this — reproducing it 1:1 is a trademark issue. Same vibe, different execution?"

Anti-patterns

  • Regenerating from scratch when a 20px move was asked. SVG edits are free; AI round-trips cost seconds and drift.
  • Sending an SVG file over iMessage. Always render to PNG (JPG for photo edits).
  • Applying distress on round 2. Distress hides composition problems from both of you — lock the clean design first.
  • Guessing at a fuzzy tweak when grid-annotate would save three rounds. Use the two-round rule.
  • Losing the previous round. Every send goes to rounds/NN.png. No exceptions.
  • Writing per-round state into the person file. Rounds are ephemeral; they live in notes.md, not memory.
  • Blowing through brief.md and going straight to generation. The brief is the contract — 30 seconds now saves 30 minutes of drift.
  • Using potrace on a color logo — it is B&W only. Reach for trace_color.sh (vtracer).
  • Using magick -fuzz or -transparent white to "remove the background" on anything with interior color. It eats colors that match the bg. Reach for remove_bg.sh (rembg).
  • Running svgo or svgo_final.sh between rounds. It's for final export only — it merges/renames paths and can break your iteration anchors.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.