ChatGPT Projects or Gemini Gems · Practical walkthrough

Build a reusable portrait-editing task kit

The reusable part is the method, not last week's face. A portrait brief you save should ask for a fresh photo every time, and never remember the previous person's grey hair.

Task by Task7 min read + practice

What you'll make

one saved portrait-editing brief, plus a short list of what each new job must supply, that you can reuse for every person on a team without carrying anyone’s appearance over.

What you'll need

an image editor your organisation lets you use with staff photos (ChatGPT and Gemini both edit uploaded photos), one recurring job such as “office-light team portraits”, and a current photo for each person who has agreed. Allow about 20 minutes to set up and test.

Example result.

Saved brief: office-light team portraits (stable rules, step 1)
Each job supplies: fresh photo · intended use · destination size · who approves
Stop rules: clipped hair or a face the photo can't support → retake, don't repair
Never stored: anyone's photo, or a description of what they look like

1. Write the brief for one recurring job

Pick one job you do repeatedly, such as neutral team portraits or window-light headshots, not “every portrait style”. The brief holds the method; each new job brings its own facts:

Follow this portrait-editing brief for the current job. Use the photo
attached in this chat as the only reference for the person. If the
intended use, the change wanted, the destination size or the approver
is missing, ask before editing. Keep facial proportions, expression,
age detail, hair, accessories, clothes and posture as in the photo.
Change only the light, background or layout I ask for, and invent no
body or hidden detail. Make one candidate, list the differences from
the photo, then suggest one repair or a restart. Don't use anything
about people from earlier chats or files.

Why this works: the fixed rules live in the brief, and everything about the person comes from the photo attached to this job. That split is what stops one person’s details turning up in the next.

What you should see: a brief with no names, appearance details or photos in it.

2. Save it where your editor will reuse it

ChatGPT: create a project, open the project’s ⋯ › Project settings and paste the brief as the project instructions. Instructions apply only inside that project.
Gemini: at gemini.google.com, choose Gems › New Gem and paste the brief as its instructions.
Anything else: keep it in a shared document, or as a saved prompt (in Microsoft 365 Copilot, hover over a prompt you’ve run and choose Save prompt).

Attach the fresh photo in each new chat. Don’t upload team photos as project files or Gem knowledge; they’d become a reference for everyone. OpenAI is retiring custom GPTs on 11 December 2026, so don’t build a new one for this.

Why this works: a saved instruction is applied to every chat in the project or Gem, so you paste the method once and supply only the job’s facts each time.

What you should see: a new chat in the project or Gem that asks for whatever the job hasn’t told it.

3. Test it on an awkward case before anyone relies on it

Run the kit on a practice job with something missing or wrong: no destination size, a photo with clipped hair, or a request that needs hands the photo doesn’t show. A good kit asks a question or says what it can’t do. A weak one fills the gap.

Why this works: a kit is only useful if it behaves well when the input is imperfect, which is most of the time.

What you should see: a question about the missing destination, and a plain “the photo doesn’t show hands” rather than a generated pair.

Check the result

The saved brief contains no names, appearance details or photos.
A new job without a destination size gets a question, not a guess.
The first candidate lists its differences from the photo.
Nothing from a previous person (hair, clothes, pose) appears in a new job.

The usual failure: the kit carries a previous person’s details into a new job, usually because a photo or a description was left in the project. Spot it by comparing the candidate’s hair, clothes and pose with the new photo, not with the last result. Fix the rule, then retest:

In this job the kit reused [previous detail] or assumed [missing fact].
Rewrite the rules so everything about the person comes only from the
photo and brief in this chat, and missing facts become questions. Show
me the changed rule and the failure it prevents.

Worked example: an office-light kit for a team of six

Fictional example. A comms officer saves the step 1 brief in a ChatGPT project for six staff portraits with office light and a square profile crop. For the second person, she attaches a fresh photo and writes “office light, square, for the staff directory”.

A good first reply:

“I have the change and the use. What size is the square crop, and who approves the final image?”

What actually came back:

Weak: “Here’s your office-light portrait. I’ve kept the navy jacket and slight head tilt for consistency with the team.”

The failure. The second person wasn’t wearing a jacket. The first person’s photo had been uploaded as a project file “for reference”, so the kit treated it as a style for everyone.
The repair. She deleted the photo from the project and ran the repair prompt with “the first person’s navy jacket and head tilt”. The amended rule now reads “Don’t use anything about people from earlier chats or files”, which is the last line of the step 1 brief.
The retest. The same job, rerun, asked for the crop size and kept the second person’s own green cardigan.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Keep approval out of the kit

The kit can list who approves and leave the answer blank, but it can’t decide that a person consented because the image looks good. Record each person’s decision separately, against the file they saw. The team portrait approval sheet covers it.

Change one rule at a time

When a test fails, record the failure, the rule you changed and the retest. Changing three rules at once makes it impossible to tell which one helped. The stable rules guide applies the same habit to other repeated tasks. A saved brief is just instructions; it doesn’t train the model, so it needs the same testing every time the tool changes.