Before you start
- An existing Stash artifact with a version you want to preserve.
- A connected AI with edit access and a clearly bounded experiment.
Connect your AI to the work
Open Stash and choose Connect your AI tools. Authorize your client, then check that it can read the intended work before asking it to edit. The connection uses MCP, a way for AI tools to read and update work in another service.
For a new publication, explicitly ask for private access. New MCP artifacts otherwise allow public editing. Your selected space is the default publishing destination, not a limit on everything the connection can access.
See the connection and handoff workflow. Client availability and workspace permissions can vary.
The workflow
- 01
Name the baseline and what must stay
Open the preferred version and record its identifier. Write down the protected elements: logo placement, booking flow, prices and content. Keep a source download as an additional handoff copy if useful. “Make it better” gives the model too much room; “try an editorial hero while preserving the booking section” defines a comparable experiment.
- 02
Fork the experiment
Ask the AI to read the selected version and fork from that baseline. Name the direction in the brief so you can distinguish it from the original. Ask it to preserve the work’s access setting and shared pin. Verify the returned branch and parent relationship rather than assuming the model chose the right starting point.
- 03
Change one meaningful variable
Try a compact hero versus a spacious hero, or a warm palette versus a neutral palette. Keep the brief, content and evaluation criteria fixed. When comparing two AI models, start each from the same baseline and record the prompt used. This produces a useful project comparison, not a universal ranking of models.
- 04
Compare the actual outputs
Open both versions at the same viewport size. Check the protected content, keyboard navigation and the main action. Ask a reviewer which direction better serves the stated audience and why. Keep subjective preference separate from whether a required link or layout works.
- 05
Choose and communicate
If the experiment wins, deliberately select the version you want people to review and create the appropriate link. Change the shared pin only if that is your intention. If it loses, keep the baseline and record one sentence explaining why. Do not describe this process as an automatic branch merge or undo of every external action.
WORKED EXAMPLE
Try a quieter workshop hero
The existing workshop page has a purple hero and a visible dates section. The experiment is a cream hero with the same text and navigation.
Starting point
A broad request such as “redesign the landing page” risks changing prices, wording and layout together.
A clearer result
The fork changes only hero treatment. Reviewers compare two directions against the same brief, then deliberately choose one for the next review.
Check the result
- The baseline still renders unchanged.
- Dates, price and CTA destination match.
- The shared pin has not moved unintentionally.
A prompt you can use
Replace the bracketed fields with your project details. Check what the AI can access before requesting changes.
Read [work URL], baseline [version ID]. Fork a separate experiment from this version. Change only [specific visual direction]. Preserve [protected elements], access and the shared pin. Return the original and experimental version links plus a concise list of differences. Do not merge or delete anything.Download the brief, prompt and checklist ↓
Questions & answers
How do I keep a good AI version while trying another direction?
Record the preferred version, then fork an experiment from it. Stash keeps immutable versions, so the earlier source remains available. Inspect the branch and deliberately choose which version to share.
How do I compare two AI models on the same design brief?
Use the same source, brief, constraints and review checklist for both models. Compare the actual outputs on the task you care about. A single design exercise is not evidence that one model is generally better.
How do I ask AI to change one section without redesigning everything?
Name the exact version and section, list what must remain unchanged, and ask for one bounded revision. Review the output against those constraints. Versioning helps recover the baseline but cannot guarantee model obedience.
Does publishing a new version move the shared pin?
No. Publishing a version does not implicitly change the shared pin. Check the version and sharing scope you intend to send to reviewers.
Keep in mind
- Forking preserves a starting point; it is not an automatic merge system.
Put it into practice.
Start with the example or bring a small project of your own.
Open your versions in Stash ↗