Changed by Aya Taleb from Replicate's prompt-images and run-models skills (Apache 2.0): added a cost check, a spend cap, key setup, error handling, more providers and a working run tool.

generate-images

Make or edit images with the user's own provider account. The user pays per run, so every run starts with a price, waits for a yes, and stays inside one spend cap shared by all providers. The tool does the money and error handling: scripts/generate.py (standard library Python). scripts/replicate_run.py still works for Replicate alone and shares the same ledger.

What this skill reads and writes

Always look for the business file first, even for a quick ask. It holds the brand look and the off-limits topics, and skipping it is how off-brand images get made.

• Reads: the active business file from founder-context (~/.founder/active and ~/.founder/businesses/<slug>.md, or a project-local founder/context.md) for brand look, tone and off-limits topics. If there is none, ask once for the look you need. Never invent brand facts.
• Writes: images to media/images/, one .json record per run (provider, model, prompt, id, estimate). The spend ledger lives at ~/.founder/media/spend-ledger.json. A saved provider order lives at ~/.founder/media/providers.json.
• Next: instagram-carousel or launch-video if the images feed a post or a promo, generate-video to animate one, talking-head-cards for card art.

1. Find out which account the user has (once)

Run python3 <skill>/scripts/generate.py providers. It reads environment variables only, makes no network call, and says which keys are set and which provider is the default.

Then run generate.py check. It makes a free call for each provider with a key and says what works. If no key is set, it prints the setup steps for every provider. Ask which account the user already has, and walk them through that one only. Why so careful: a key can spend their money, so it must live in their shell environment, never in chat or in a project file.

If keys are set for more than one image provider and no order is saved, the tool says "ask the user". Ask which they prefer, then save it: generate.py providers --order cloudflare,openai.

If they paste a key into chat anyway, do not repeat it back, and do not pass it on: not in a command, a file, or the arguments of another skill or tool. Tell them to revoke it in that provider's dashboard and set a new one in their shell, since chat logs are not a safe place.

2. Pick the model from live data and the dated table

Never pick from memory. Image models change every few weeks, and a remembered name may be gone or overpriced (OpenAI shut down gpt-image-1 on 2026-10-23). Use the find-model skill, or:

python3 generate.py models --kind image
python3 generate.py schema <model>
python3 generate.py price <model> -i prompt="..."

models lists live models where the provider has a listing call, joined with prices from scripts/prices.json. Each price there has its source link and checked date. For Replicate, models --provider replicate --query "product photo" searches live, and prices come from the model page. Prefer Replicate's official models (no cold boot, fixed prices).

When the user says cheap, price two or three candidates on the providers they have and pick the cheapest one that fits. Say why if you pick a pricier one. Read schema before writing inputs, because names like size, aspect_ratio or num_outputs differ by model.

The provider facts, with links and dates, are in references/providers.md. The Replicate API notes are in references/replicate-api.md.

3. Write the prompt

Write full sentences, not keyword lists. Say the subject, setting, light, lens and framing. Name things directly ("the woman in the navy blazer") instead of using pronouns. When editing, say what must stay the same. Put any words that should appear in the image in double quotes.

The full guide, with editing, text, product shots and character consistency, is in references/prompting-images.md. Read it when the job is more than a plain scene.

Keep facts out of images that you cannot back up. Do not put prices, stats, logos of other companies, or customer quotes into a prompt unless the user supplied them.

4. Show the cost and wait for yes

Do a dry run first. It prints the estimate, how it was worked out, the session total and what is left under the cap. It spends nothing, and for table providers it needs no key at all.

python3 generate.py run <model> -i prompt="..." --label hero

Show the user, in plain words: the provider, the model, the prompt, how many images, the estimate, and the session total. Then ask. A request like "make me an image" is not a yes to a price they have not seen. Why: providers bill the moment a run starts, and it cannot be undone.

One exception: if the user already named a dollar amount they agree to ("go ahead, up to $1"), and the full estimate fits under it and under the cap, that is the yes. Say the estimate, then run. If it does not fit, stop and ask.

Batches: one run per call. If the user wants 4 options, say the total for all 4 and get one yes for that total. Then run them one after another, not in parallel, so a failure or a cap hit stops the rest. Cloudflare makes one image per call. OpenAI takes -i n=4 in one call.

Big asks ("run it 50 times overnight", "don't ask me about cost"): do not agree to the 50, now, later, or "in this session". Unattended loops and big batches are how people wake up to a surprise bill. Instead:

1. Price it. price works without a key, so there is always a real number. Multiply it out.
2. Say the session cap ($2 unless the user changed it) and how many runs fit under it. Offer a small first batch (for example 4 runs), with its total, to look at before doing more.
3. Wait for a yes to that number. "Don't ask me about cost" is not a yes to a number.

Never quote a price from memory. If you have not run price in this session, you do not know it. If the tool warns the table is more than 60 days old, read the source link first.

5. Run it

After a clear yes, run the same command with --yes:

python3 generate.py run <model> -i prompt="..." --label hero --yes --out media/images

The tool then: reserves the estimate on the ledger, sends one request, waits or polls with backoff, saves every image at once, and writes the record. Local files go in as -i image=@path/to/photo.jpg (the input name comes from schema; on Cloudflare FLUX.2 it is input_image_0).

Then show the saved file paths and the spend so far (generate.py ledger).

6. When something goes wrong

Read the exit code and the message. Do not just rerun.

Rate limits (429) that mean "slow down" are handled inside the tool: it waits and says so.

7. Close every answer the same way

Whether you ran, did a dry run, or declined, end with one line that names the next step: instagram-carousel for a post, generate-video to animate the image, launch-video for a promo, or "say yes and I'll run it for $X". Why: the user should never wonder what happens next.

Describe results from the files and the prompt, not from what you imagine. You have not seen the image unless you opened it.

Rules

• No run without a price shown and a yes. The first run call is a dry run, even when the user pre-approved a budget, because the estimate has to exist before you can check it fits.
• Stay under the cap (default $2 a session, one cap for every provider). Never raise it on your own.
• One request at a time. No loops that fire runs in the background.
• Save outputs right away. Never hand the user a provider link as the result.
• Never put a key in a file, a command you show, or your reply.
• Do not make images of real people without their consent, fake screenshots, fake reviews, or anything that passes for a real document. Say no and offer a version that is clearly an illustration.
• Prompt-only requests ("just write me a prompt") need no key and cost nothing. Do not push a paid run on someone who did not ask for one.

Tool reference

generate.py commands: providers [--order a,b], check [--provider P], models [--provider P] [--kind image|video] [--query Q], schema MODEL, price MODEL, run MODEL, ledger. Flags for run: -i key=value (repeatable), --input '{json}', --input-file, --provider, --yes, --out, --label, --max-cost, --assume-cost, --timeout (seconds, default 600), --deadline (Replicate auto-cancel, default 15m). The provider is worked out from the model name; pass --provider when it is not clear.