Claude to Prepare, Prompt and Review AI-Edited Photos
Practical AI Tutorial for M&E and International Development
How to Use Claude to Prepare, Prompt and Review AI-Edited Photos
Claude cannot generate or edit the pixels of a photograph itself. It can analyse your original image, write a precise transformation prompt, and run a structured before/after review once an image tool has produced the result – protecting accuracy, dignity, privacy, accountability to affected populations (AAP) and the integrity of development evidence within your MEL/MEAL system at every step.
Quick answer
Claude does not produce images. What it does very well is the thinking around image editing: it can look at your uploaded photo, draft a tight, detail-preserving prompt, and later compare an AI-generated illustration against the original to catch inaccuracies. Upload the photo to Claude, ask it to draft a transformation brief, take that brief to an image-editing tool (ChatGPT Images, Gemini/Nano Banana, Adobe Firefly, or similar) to generate the illustration, then bring the result back to Claude for a structured review before publication.
What you will learn
Analysis, prompt-writing and review – not image generation.
Let Claude turn a vague idea into a precise, protective prompt.
Have Claude check the result against the original, point by point.
Use Claude to draft consistent disclosure and evidence-file notes.
Important: Claude is the reviewer, not the editor
Claude has vision – it can look at an image you upload and describe, reason about and compare it. It has no image-generation model behind it, so it cannot turn that photo into an illustration on its own. In this workflow, an external image tool does the pixel-level transformation. Claude’s job is everything around that step: understanding the source photo, writing the prompt, and checking the output.
Skipping Claude and going straight to an image tool still works, but you lose the structured protect-list and the independent before/after review this tutorial builds in.
The six-step workflow
Share the original photograph for analysis
Ask Claude to write the transformation prompt
Name the details that must remain unchanged
Run the brief in an image-editing tool
Bring the result back for a structured comparison
Correct inaccuracies and label substantial changes
Communication image or evidence?
An AI-transformed illustration may support a report, presentation, training manual or social media post. It should not replace the original photograph in an evidence archive, monitoring record, verification file, results framework annex or audit trail, and it is not a substitute for triangulated primary data.
Keep both versions: retain the original image, consent information, date, location, the prompt used, editing notes and the approved final version as part of your MEL plan’s data management protocol.
Upload the photo to Claude
Start with a clear image that you have permission to use, and let Claude look at it first.
Open Claude, use the attachment option and upload the photograph. You may use a programme photo, community workshop image, monitoring visit, training photo, field activity, conference image or personal vacation photo. Ask Claude to describe what it sees – rights-holders and participants, activity, setting, objects, condition of infrastructure – before asking for anything else. This description becomes your factual baseline for everything that follows and should be handled in line with your organisation’s safeguarding policy and do-no-harm principle.
Usually appropriate
- Training and workshop photos with consent
- Public events and programme activities
- Non-sensitive infrastructure or landscape images
- Personal photos you own or may edit
High-risk or unsuitable
- Identifiable children without valid consent, or anything falling under your child safeguarding policy
- Survivors, patients, or people facing protection or PSEA (protection from sexual exploitation and abuse) risks
- Confidential documents, forms, personally identifiable information (PII) or identity details
- Photos from insecure or politically sensitive locations
Development-sector check: confirm that your informed consent protocol and photo consent form cover the intended use, including public communication and substantial visual transformation, and that use aligns with your donor’s communications and visibility guidelines before you upload anywhere.
Ask Claude to draft the transformation brief
Let Claude turn your intent into a precise prompt for the image tool you will use next.
Tell Claude the style you want, the purpose of the image and where it will be used. Ask it to write a single, tight prompt you can paste into an image-editing tool, based on what it observed in step 1 – not on assumptions.
What to tell Claude
Example instruction to give Claude:
Claude will typically return a prompt similar to this, ready to paste elsewhere:
Build the protect-list with Claude
Ask Claude to list, from the photo itself, exactly what must not change.
Image transformations can unintentionally alter people, objects and environmental conditions. Ask Claude: “Looking at this photo, what specific details would be risky to lose or alter if it were turned into an illustration?” Claude will point to the concrete details in your image rather than a generic list.
Why: avoids exaggerating reach or attendance.
Add to the prompt: “Preserve the original number of people.”
Why: prevents an unfinished facility from appearing complete.
Add to the prompt: “Do not make the building newer or more complete.”
Why: protects accurate representation of disability inclusion.
Add to the prompt: “Keep the wheelchair and accessibility features visible.”
Why: reduces stereotyping and identity distortion.
Add to the prompt: “Preserve clothing, skin tones and cultural details.”
Why: keeps the programme activity accurate.
Add to the prompt: “Do not add supplies, tablets or services.”
Why: keeps the image consistent with what your results framework or logframe already reports.
Add to the prompt: “Do not imply an outcome or output level not documented in the results framework.”
Ask Claude to fold the final protect-list directly into the brief from step 2, so nothing gets lost between tools.
Generate the image in a dedicated tool
Take Claude’s finished brief to an image-editing tool and produce the first version there.
Paste the brief Claude wrote into an image-generation or image-editing tool – ChatGPT Images, Gemini/Nano Banana, Adobe Firefly, or a similar service your organisation has cleared for use – and upload the same original photo there. This is the only step in the workflow where the actual pixels change; Claude does not perform this step itself.
For a localised change, ask Claude to help you phrase a narrow, single-area instruction rather than a full-image transformation, for example:
Save the output and bring it back to Claude for review before you use it anywhere.
Have Claude compare the output with the original
Upload both images to Claude and ask for a structured, item-by-item comparison.
Human-plus-Claude review
Primary source record
Communication output from the image tool
Flags each protect-list item that changed
Approve, send for revision, or reject
Ask Claude directly: “Compare this generated image with the original photo I shared earlier. Go through the protect-list and tell me exactly what changed.” Claude will check for differences in:
Treat Claude’s comparison as a second opinion that speeds up review, not as a replacement for a human sign-off – especially for anything involving children, protection risks or donor-facing evidence.
Refine, approve and disclose
Use Claude’s findings to write a focused correction request for the image tool, then document the decision.
Feed Claude’s flagged issues back into short, focused instructions for the image tool – address one or two problems at a time rather than restarting with a vague prompt:
Once approved, ask Claude to draft the disclosure line and a short evidence-file note recording the original file name, consent status, date, the prompt used and the reviewer’s decision. Where the transformation is substantial, label the image clearly so viewers do not mistake it for an unedited documentary photograph.
Briefs to ask Claude for, by M&E use case
Give Claude the use case below along with your uploaded photo, and ask it to produce the paste-ready prompt for your image tool, protect-list included.
Community consultation
Useful for learning briefs, stakeholder presentations and participatory evaluation materials.
Enumerator training
Useful for survey manuals, data-quality training and fieldwork presentations.
Infrastructure monitoring
Useful for monitoring reports, donor updates and inspection guidance.
Focus group discussion
Useful for qualitative research training and methodology sections.
Humanitarian distribution
Useful for situation reports, process manuals and staff learning.
Agriculture and livelihoods
Useful for extension guides, programme reports and climate adaptation materials.
Public health activity
Useful for awareness materials, training and health programme communication.
Education programme
Useful for learning reports, teacher training and school programme materials.
Playful vacation-photo example
The same three-step Claude workflow – analyse, draft the brief, then review the result – works for a personal travel photo too. The transformation can be more creative, but Claude should still be asked to protect recognisable people, and you should get permission from anyone else shown before uploading.
Privacy-preserving transformations
Turning a photograph into an illustration may reduce facial detail, but it does not automatically anonymise the people shown or remove personally identifiable information (PII). Identity may still be inferred from clothing, location, assistive devices, body position, name tags or distinctive background details. Ask Claude to look at the photo specifically for these identity cues before you draft the brief, and treat this step as part of your data protection impact assessment (DPIA) or equivalent data protection review.
Remove unnecessary identifying areas.
Reduce identifiable facial characteristics.
Hide names, plates and confidential forms.
Reduce distinctive details when necessary.
What to avoid
- Treating Claude’s description of an image as if Claude produced that image.
- Skipping the review step because the illustration looks convincing.
- Adding participants who were not present.
- Making a poorly attended event appear full.
- Adding equipment, supplies, services or branding that did not exist.
- Making unfinished infrastructure appear completed.
- Removing visible problems that are relevant to monitoring or evaluation.
- Changing demographic, cultural, gender or disability characteristics.
- Exaggerating emotions, vulnerability, poverty or programme success.
- Breaching your donor’s communications and visibility guidelines or branding requirements.
- Bypassing your safeguarding or PSEA reporting obligations because “it’s just an illustration.”
- Publishing a transformed image as if it were an untouched documentary photograph.
Master instruction to give Claude
Paste this into Claude along with your uploaded photo. Replace the bracketed text, and Claude will return a paste-ready prompt for your image tool.
Responsible-use checklist
Suggested disclosure wording
Light transformation:
Substantial transformation:
Frequently asked questions
Can Claude edit or generate an image directly?
No. Claude can view and analyse an uploaded photo and produce text – descriptions, prompts, comparisons, disclosure wording – but it has no image-generation model and cannot output a new picture. The actual illustration must be produced in a separate image tool.
Then what is Claude actually useful for in this workflow?
Three things: turning a rough idea into a precise, protective prompt; comparing the generated image against the original point by point; and drafting consistent disclosure and documentation text. These are the steps where accuracy and consistency matter most, and where a language model is genuinely strong.
How can I keep people recognisable?
Ask Claude to build the protect-list from the actual photo and fold it into the brief before you take it to the image tool. After generation, ask Claude to compare the result with the original and flag anything that changed in faces, clothing, poses or positions.
Can an AI-edited image be used as M&E evidence?
A transformed illustration should normally be treated as a communication product, not primary evidence, results-framework data or triangulated verification material. Retain the original photograph, the prompt Claude drafted and supporting documentation in the evidence archive as part of your MEL plan’s data management protocol.
Does turning a photo into an illustration anonymise participants?
Not necessarily. People may remain identifiable through their clothing, location, body position, assistive devices or surrounding details, which still counts as personally identifiable information (PII). Ask Claude to check for these cues specifically, and conduct a separate privacy and protection review.
Does this workflow satisfy safeguarding and donor visibility requirements?
This workflow supports those requirements but does not replace them. Your organisation’s child safeguarding policy, PSEA reporting obligations, informed consent protocol and the donor’s communications and visibility guidelines still apply in full, and any escalation they require takes precedence over publishing the image.
Which image tool should I pair with Claude?
Any image-editing tool your organisation has cleared for use and that meets your data-handling requirements works with this workflow. Claude’s role – drafting the brief and reviewing the output – stays the same regardless of which tool generates the illustration.
Official product guidance
For current details on what Claude can and cannot do with images, consult Anthropic’s documentation and Claude support.
Continue learning with EvalCommunity Academy
Build practical and responsible AI skills for evaluation, programme management, development and humanitarian work.
