pixeltamer

pixeltamer — generate, edit, and compose images with GPT Image 2.5

Image generation that actually does what you asked for. Two backends so you don't have to pick a side: bring an OpenAI API key OR use the codex CLI signed in to your ChatGPT subscription. The API backend defaults to gpt-image-2.5-flare and accepts --model; Codex selects its image model server-side. Same skill, same prompts, your choice of how to pay.

When to use this

Trigger this skill when the user asks for any of:

A new image, poster, illustration, photograph, mockup, icon, sticker, sprite, logo, hero, infographic, ad creative, scroll-stop card, magazine cover, podcast art, product shot, banner, OG image, social card, character reference sheet, app screen, dashboard mockup
An edit of an existing image (modify, inpaint, mask, change-only-X)
A composition of multiple reference images blended into one
A batch of related images to be generated together with verification

If the request is "make me an image" or any flavor of it, this is the skill.

Prerequisites — check the install (do this BEFORE the first generation)

Run this first to see what's available:

pixeltamer doctor

If the output shows at least one backend with a ✓ marker, you're set — go generate. Auto-detect picks API if a key is set, otherwise codex. Override with --backend api|codex per-call or set PIXELTAMER_BACKEND env var.

If pixeltamer doctor fails with permission denied instead of printing backend status: the skills CLI stripped the dispatcher's execute bit on install/update. The dispatcher can't restore its own bit (it has to run to do that), so bootstrap it once via bash — it self-heals everything else from there:

# run it through bash (works even without +x):
bash ~/.claude/skills/pixeltamer/scripts/pixeltamer doctor
# or restore the bit permanently:
chmod +x ~/.claude/skills/pixeltamer/scripts/pixeltamer

Every other script now runs via its interpreter, so the dispatcher is the only file this bit can break.

If BOTH backends show ✗, do not run pixeltamer config — it's an interactive prompt that needs a real TTY, which the Bash tool doesn't have, so the call hangs forever. Use AskUserQuestion instead and offer the user this choice:

"Pixeltamer needs one of two backends. Pick: API key — fastest, full feature set (edit + compose), per-image cost. Codex CLI — uses your existing ChatGPT Plus/Team/Enterprise subscription, no key, generation only."

If the user picks API

Tell them to drop their key in pixeltamer's config dir. This is the cleanest path because the Python script auto-loads it on every invocation — no shell restart, no env-var management:

mkdir -p ~/.config/pixeltamer
echo 'OPENAI_IMAGE_API_KEY=sk-yourkeyhere' > ~/.config/pixeltamer/.env
chmod 600 ~/.config/pixeltamer/.env

If their org isn't verified for GPT Image access, the first call returns 403 — point them at https://platform.openai.com/settings/organization/general for the one-time verification.

If they prefer dotfiles instead, export OPENAI_API_KEY="sk-..." in ~/.zshrc (or .bashrc) works too — but they'll need to restart the agent so the new env is inherited. The .env file path doesn't have that requirement.

If the user picks codex

# 1. install codex if missing
npm install -g @openai/codex

# 2. log in (opens a browser, OAuth via ChatGPT)
codex login

# 3. confirm
codex login status

No restart needed. The bash backend reads codex auth fresh on every invocation.

codex updates separately from the skill — npx skills update won't touch it. pixeltamer doctor now reports the installed codex version and flags when it's behind npm latest (a stale codex is a known failure mode — an old schema version once caused indefinite hangs). If doctor says an update's available, run npm install -g @openai/codex@latest.

After the user sets things up

Re-run pixeltamer doctor to confirm. Then proceed with the original request.

Three modes

Mode 1 — One-shot generation

pixeltamer generate -p "<prompt>" --size 1024x1024 --quality high -o output.png

Or the shortcut, which infers generate:

pixeltamer "Create a pitch-deck slide titled..."

For exploration, fire 4 in parallel:

pixeltamer generate -p "..." -n 4 --concurrency 4 -o variants/

Transparent assets

Anything that gets composited onto a background you don't control — icons, logos, mascots, product cutouts, chart layers, stickers — should be generated with real alpha, not a green screen you clean up afterwards:

pixeltamer generate -p "<isolated subject>" --background transparent -o asset.png

Works on both backends for generate. On API it's a request parameter; on codex it's translated into prompt instructions (codex has no background param) plus an automatic alpha check that warns on stderr if the PNG came back RGB. Codex edit/compose can't do it — the dispatcher says so.

The catch worth knowing before you write the prompt: the prompt outranks the flag. Any backdrop, surface, scene, or cast shadow you describe gets painted, and you get a fully-opaque PNG with no error anywhere. Load references/transparency.md for the constraint block and the verification steps.

Mode 2 — Edit / inpaint

# Modify a single existing image
pixeltamer edit -i source.png -p "Change ONLY the sky to overcast. Preserve everything else exactly." -o edited.png

# With a mask — white pixels regenerate
pixeltamer edit -i source.png --mask mask.png -p "..." -o edited.png

Edit works on both backends. The API path uses /v1/images/edits (supports --mask for region-based inpainting). The codex path uses the OAuth Responses API directly (scripts/pixeltamer_codex_oauth.py) — same ChatGPT credentials as codex exec, different transport, no API key required. Codex-OAuth doesn't support --mask; if the user needs mask-based inpainting they need the API path.

Mode 3 — Multi-reference composition

Pass 2–16 reference images to be blended into a single output via /images/edits:

pixeltamer compose \
  -p "Reference 1 is the product. Reference 2 is the kitchen scene. Compose: place the product on the counter with morning window light from the left." \
  -i product.png \
  -i kitchen.png \
  -o composed.png \
  --size 1536x1024

This is the killer feature. See references/multi-reference.md for the labeling pattern that actually works.

Reference order still matters: put the face, logo, or product detail you can least afford to lose first. GPT Image 2.5 improves subject preservation by default. --input-fidelity high remains available on direct API edits for models and hosts that expose that control.

Mode 4 — Batch (multi-image plan with verification)

For projects that need 4+ related images (a website's hero + features + footer + social), use the batch state machine. Workflow:

Survey what images are needed.
Plan a prompts.md with one entry per image (target path, format, native size, optional reference, status, prompt). Entries whose Format says PNG transparent must be generated with --background transparent — the verifier checks for a real alpha channel and fails them otherwise. When the batch is a collection (product range, icon family, sticker pack), write one shared style-and-constraint block and append it verbatim to every entry's prompt. That constant suffix is what makes N separate generations read as one set — see references/transparency.md.
Generate each one, calling the right backend per entry.
Run pixeltamer batch <path-to-prompts.md> to verify every output (file exists, dimensions match, file size sane).
Visually self-review each generated PNG using the Read tool; demote to failed:<reason> if it doesn't match the prompt.
Re-generate only the failed entries.

The prompts.md format is parsed by the verifier in scripts/verify-images.mjs. See references/multi-reference.md for the structure.

Workflow — what to do when invoked

Read references/index.md first. Find the closest match for the request in the routing map. Load the smallest useful slice — one recipe or one playbook file or one reference, not all of them. Never write a prompt from scratch when a similar pattern exists in our recipes or playbooks. If no match, fall back to references/prompting.md plus the closest recipe.
Understand the request. Identify the mode (generate / edit / compose / batch). Confirm the recipe / playbook / reference loaded in Step 0 actually fits — if not, route again.
Read references when relevant. Beyond the file index.md pointed you to: load references/prompting.md for any non-trivial prompt. Load references/multi-reference.md for compose mode. Load references/transparency.md for any transparent-alpha asset. Load references/prompt-patterns.md when you're composing the prompt programmatically, hitting one of the doctrine categories (UI / infographic / brand identity / e-commerce hero / architectural render / scientific atlas / typography poster), or need brand consistency across many images. Don't dump every reference into context — pull only what's needed.
Resolve unspecified params. Pick a sensible size based on the use case (see the size table in references/prompting.md). Default --quality high. Pick a backend if the user didn't specify. If the asset will be composited onto something else — icon, logo, mascot, cutout, chart layer — reach for --background transparent and load references/transparency.md before writing the prompt.
Build the prompt. Apply the canonical structure: Intent → Scene → Subject → Details → Text → Style → Constraints. Drop magic words ("8K, ultra detailed, masterpiece, professional"). Quote any text that should appear in the image. Specify what to preserve on edits. For UI screens, infographics, brand renders, scientific atlases, or any request where you're composing the prompt programmatically, escalate to the JSON-config schema and other advanced patterns in references/prompt-patterns.md.
Generate. Call the right pixeltamer subcommand. Print the resulting absolute path.
Visually self-verify. Use the Read tool to view the generated PNG. Judge it against the prompt:
- Does the subject match?
- Is the text rendered correctly and spelled exactly as quoted?
- Are there obvious artifacts (warped anatomy, misspellings, glitched typography)?
- Does the composition match what was asked?
- If you asked for transparency: is the alpha actually transparent? An opaque alpha channel is the standard silent failure — check it, don't assume it. If it fails — say so honestly. Don't paper over a bad result with "looks great!"
Iterate or hand off. If the result is wrong, change ONE dimension and regenerate (see iteration table in references/prompting.md). If the result is good, surface the file path to the user.

Flags control properties, the prompt controls content

The cleanest mental model in this whole skill, and the one agents get wrong most often. Output properties are flag-controlled. Describing them in the prompt is unreliable and backend-dependent.

Property	Use the flag	Not the prompt
Dimensions	`--size 1536x1024`, `2K`, `4K`	"a 1536x1024 image of…"
Output container	`--output-format png\|jpeg\|webp`	"save it as a webp"
Compression	`--output-compression 0-100`	"compress it a bit"
Render quality	`--quality low\|medium\|high`	"high quality, detailed"
Variant count	`-n 4`	"give me four versions"
Edit region	`--mask region.png`	"only change the top left"
Reference fidelity	GPT Image 2.5: critical reference first; older models: `--input-fidelity high`	"keep the face exactly"

The prompt is for what is in the picture. Size, format, count and masking are not in the picture.

Background is the exception, and only on codex. On the API backend --background is a real parameter and the rule above holds. On codex there is no such parameter, so transparency genuinely is prompt-driven — which is why --background transparent there rewrites your prompt rather than setting a flag. Both facts live in references/transparency.md.

Visual self-verification — non-negotiable

Every generated image gets Read-loaded back and visually judged before claiming success. Image generation is stochastic; "API call succeeded" doesn't mean "the image is what was asked for". This step catches:

Misspelled or paraphrased text overlays
Wrong subject (a coffee shop instead of a coffee cup)
Visual artifacts (warped hands, glitched typography, extra limbs)
Wrong aspect ratio framing (subject too small or cropped)
Off-brand color or tone

When something is wrong, the right move is usually to change ONE thing in the prompt and regenerate, not to bolt three new clauses on hoping one helps.

Output format — what to tell the user

After a successful generation, surface:

Absolute file path (the script prints this on stdout)
One sentence on what was generated (e.g. "Generated a 1024×1536 portrait cover with the headline rendered exactly")
If multiple variants (-n N), list all paths

After a self-verified failure:

Explain what went wrong (text was paraphrased, subject was generic, etc.)
Suggest the single most likely fix and offer to regenerate
Don't auto-regenerate without the user's nod — they may want to tweak the prompt themselves

Recipes available

recipes/infographic.md · recipes/meta-ad.md · recipes/viral-linkedin.md · recipes/ui-mockup.md · recipes/editorial-cover.md · recipes/product-photo.md · recipes/mascot.md

Load the matching one when its trigger fits. They contain prompt skeletons, worked examples, sizing defaults, and common-failure tables.

Mascot / brand character / app icon character requests — always load recipes/mascot.md BEFORE writing the prompt. The model's default for "cute character" is a generic Slack-blob; the recipe codifies the context-to-anchor decision tree, the silhouette/scale rules, and the explicit anti-patterns to suppress. Skipping it is the single most common cause of bland mascot results.

References available

references/prompting.md — the doctrine. Canonical structure, what NOT to say, style/composition/lighting vocabulary. references/api-backend.md — API specifics, env vars, troubleshooting. references/codex-backend.md — codex CLI specifics, both invocation patterns, tradeoffs. references/multi-reference.md — compose mode mastery, labeling patterns, reference set sizing. references/post-process.md — compress, resize, convert, alpha-extract, combine. references/ui-mockup-prompting.md — analogy vs inventory style, real-data over placeholders, asset codification.

Pull only what's needed for the current job. Don't dump them all.

Common mistakes — drop these patterns

Don't	Why
"8K, ultra detailed, masterpiece, trending on artstation"	Magic words don't specify a visual result. Describe materials, light, framing, and texture instead.
"professional, beautiful, premium, stunning"	Praise language with zero instructional content.
Generating before checking which backend is available	Run `pixeltamer doctor` first if it's the first call this session.
Edit / compose on codex backend	Works since 0.3.0 via the OAuth Responses API (`pixeltamer_codex_oauth.py`). Mask-based inpainting still requires the API path — the Responses API doesn't take a mask parameter. So does any edit needing a specific output size: codex returns the input's aspect and ignores `--size`.
Chroma-key green screens for cutout assets	Stale advice. `--background transparent` gives real alpha on both backends for `generate`. Key only when the edges come out dirty, or for codex `edit`/`compose`.
Describing a backdrop while asking for transparency	The prompt outranks the flag. "Clean white studio" beats `--background transparent` and you get a white rectangle.
Skipping visual self-verification	Image gen is stochastic. "API succeeded" ≠ "image is correct".
Stacking three new clauses when one isn't working	Change one dimension at a time. You won't know what helped otherwise.
Running examples folder PNGs as ground truth	They're demonstrations, not specs. Composition will vary on regen.

Backend cheat sheet

Need	Use
Fastest single image	API with the default `gpt-image-2.5-flare`
Don't have / don't want an API key	codex
Edit an existing image (no mask)	Either backend — API or codex-OAuth
Mask-based inpainting	API only — Responses API doesn't take a mask parameter
Preserving a face / logo / label type across an edit	API; use Sunburst and keep the critical reference first
A guaranteed output size when editing	API only — codex ignores `--size` for `edit`/`compose` and returns the input's aspect
Compose 2–16 references into one	Either backend — API or codex-OAuth
Run on a teammate's machine without sharing credentials	codex (each user signs in separately)
Transparent background / real alpha channel	Both, for `generate` — API takes it as a param, codex via prompt + verification. Codex `edit`/`compose`: API only
A specific output file format (png / jpeg / webp)	API only via `--output-format` — both codex transports always return PNG
Custom OpenAI-compatible host (jmrai, ZenMux, OpenRouter)	API with `OPENAI_IMAGE_BASE_URL` set
Maximum quality / edit precision	API with `OPENAI_IMAGE_MODEL=gpt-image-2.5-sunburst`
Largest sizes (4K) at high quality	API; on codex, `--size` is honoured for `generate` only and its reasoning loop slows on large outputs

Persona note

Output (file path, success message, failure explanation) should be terse and useful. No hype, no "here's your stunning generated masterpiece" framing — that's the kind of slop pixeltamer is supposed to help kill, not produce. Just: what was generated, where it lives, and what to do next.