Keep this card open next to your Claude conversation while you commission images or video through Higgsfield. It doesn't explain how the connection works. It walks you through one batch of assets: setup, the brief and corrections, sign-off, logging in the sheet, and the call on whether the process is ready to automate. The blocks follow the order of the work. Each one has boxes to tick and one line to remember.
Two terms the card relies on. A description (the prompt) is the detailed instruction Claude sends to the generation model. A reference image is a real photo of your product that the model has to reproduce rather than invent.
1. Before you generate anything
Most wasted credits come from a setup nobody checked. You go through this block once per project.
- [ ] You have a Higgsfield account on an active plan and you know how many credits are left.
- [ ] You've chosen how to connect: the chat connector for a quick trial, or the CLI in Claude Code for repeatable work.
- [ ] Connector: you've checked the permissions in its configuration and kept manual approval of generations until you know how the assistant behaves.
- [ ] Claude Code: you have a dedicated project folder with the CLI installed, your account signed in and Higgsfield's ready-made agent skills added.
- [ ] The project holds a file of advertising knowledge for your channels (what grabs attention, what sells, how the platforms differ), and the assistant knows to consult it.
- [ ] The assets folder holds a real photo of every product that will appear in the ads.
You pay credits for every generation, including the ones built on a bad setup. Checking it takes less time than one failed batch.
2. The brief for one asset
Copy this template into the conversation and fill it in before every new asset. An empty field is a place where the model will guess.
Product: [name] — reference image: [file in the assets folder]
Do NOT change: [label, colours, lettering, packaging shape]
Type: [product photo / static ad / video / UGC / launch clip]
Channel and ratio: [e.g. vertical story, square post, 16:9 video]
Angle: [curiosity / contrarian / surprise / question / hard fact]
Headline and line: [the copy, or a request for options to approve]
Style and pace: [e.g. calm, cinematic, energetic, fast cuts, close-ups]
Person in frame: [yes, who / no, product only]
Banned words: [phrases that got generations rejected before]
Result goes to: [sheet, tab, row]- [ ] Every field is filled in, or deliberately marked "assistant's call".
- [ ] The "Do NOT change" field says outright that the product must look exactly as it does in the reference image.
- [ ] Every promise in the headline (a number, a guarantee, an outcome) is something you actually offer.
- [ ] You've asked the assistant to ask questions before generating if anything is missing.
The words you think in ("energetic", "up close") are enough. Turning them into a model description is the assistant's job, not yours.
3. Generation and corrections
The first result is a draft. This block keeps corrections quick and makes sure they aren't lost.
- [ ] You correct in a plain sentence that points at a specific asset ("remove the second headline on this ad") instead of regenerating everything.
- [ ] When a result is too quiet or generic, you add specifics: format variant, pace, type of shots, aspect ratio.
- [ ] When the platform rejects a generation as sensitive content, you don't resend it. You ask: "show me the description you sent and work out why it was rejected".
- [ ] The words the assistant identified as the cause go into the brief's "Banned words" field.
- [ ] When one variant clearly wins, you ask for a series of variations on it that each change one element (headline, shot, angle).
Resending a rejected description gets the same result. Diagnose first, then try again.
4. Sign-off: can this asset go out
Tick these for each asset separately. A single "no" sends you back to block 3.
- [ ] The product looks like your product: colours, shape, label, small print.
- [ ] The text is legible and free of typos. In video, check it frame by frame wherever the label is visible.
- [ ] If the video garbles the lettering but the photo is clean, consider a simplified label for video (logo and name) rather than more retries.
- [ ] No odd scenes: unnatural gestures, warped elements, doubled objects.
- [ ] Aspect ratio, pace and room for the headline suit the channel in the brief.
- [ ] Every line with a number, a guarantee or a promised outcome has been checked against your offer. The assistant will happily invent lines like that on its own.
- [ ] A person approves the asset before it reaches publishing or the ads manager.
The most common defect is a product that almost looks like yours. In an ad, "almost" means someone else's product.
5. The sheet: a record that lets you repeat a win
A good result without a saved description is a one-off. This block checks that the sheet works as production memory.
- [ ] The sheet has columns for product, style, type (image or video), model, the full description sent to the model, job ID, result URL and status.
- [ ] Every generation in this batch has a row with the complete description, not a summary.
- [ ] Status has a fixed set of values (for example: to do, in review, done, rejected) and the assistant sticks to it.
- [ ] The plan for the next variants went into the sheet rather than staying in the chat window. If the assistant didn't do that on its own, you asked for it explicitly.
- [ ] For Marketing Studio assets, you noted that the module picked the model, not you.
- [ ] Once you have channel data, you record each asset's performance next to its row.
An empty status is an agreement: nobody has touched this row yet. Without that column, automation will start duplicating work.
6. Skill: when to turn a good result into a recipe
A skill is a text file in the project's .claude/skills folder that tells the assistant how to make a given type of asset. Write one only once you have a result you want to repeat.
- [ ] You generated several versions and picked the one or two you genuinely want to repeat.
- [ ] You pasted the winning description into Claude Code and named what works about it (pace, shots, detail, light).
- [ ] The skill includes the banned words from block 3 and the reference-image rule from block 2.
- [ ] The skill tells the assistant to ask questions before generating (for example, a person in frame or product only).
- [ ] The new skill shows up in the app. If it doesn't, you restarted the app.
- [ ] On first use you checked which skill the assistant actually ran. If it reached for a general generation skill, you stopped it and named the right one.
- [ ] After every batch you said what to keep and what to change, and had the skill updated.
A skill that isn't corrected after each batch freezes the first, accidental version of your taste.
7. Routine: when to hand the schedule to the assistant
A routine is an instruction Claude runs by itself on a schedule. You switch it on last, not first.
- [ ] Several batches in a row passed block 4 with no fixes to the product or the text.
- [ ] You've set up two separate routines: planning (review the sheet and data, add a new batch of ideas) and production (rows with an empty status → description → generation → marked in the sheet).
- [ ] Production only takes rows with an empty status.
- [ ] You start with a small batch size and low frequency, and raise them only when results are stable.
- [ ] You've agreed who reviews routine output, and when, before anything gets published.
- [ ] Automatic publishing, or a link to the ads manager, waits until you trust the skills enough not to review every result.
A routine multiplies the process you already have. If the process gives random results, the routine just gives you more of them.
After the batch
An empty box on this card rarely hurts straight away. It comes back in the next batch as a product with a different label, a generation rejected for the same word, or a great result nobody can repeat because the description was never saved. When that happens, go back to the block the box came from.
The principle I take from this: quality doesn't come from one better description, but from a process that remembers what worked. If you tick only blocks 2 and 5 today, your next batch starts from knowledge instead of from zero.