From Problem to Proof
AI-assisted verification for DSpace
Jozef Mišutka
AI can help at every step, you sign off
/ 10 The idea The tools Demo Real life Take-home
YOU + AI
Report
in plain words
AI AGENT
Reproduce
in your browser
DEVELOPER + AI
Fix
in a test system
AI AGENT
Re-test
the same check
YOU
Decide
CLOSE
AFTER RELEVANT CHANGES
REOPEN
REPORT
FIX
UPGRADE
NEXT
One check, rerun after relevant changes
2
Use AI to make handovers clear and repeatable
/ 10 The idea The tools Demo Real life Take-home
REPORTER
The request
in plain words
TECHNICAL TEAM
The evidence
screens and values
APPROVER
The sign-off
on the same record
AI
AI
ONE SHARED RECORD
The request and its evidence, the same for all three, rerun after relevant changes.
Use AI.
I will show it can be done.
No technical time?
Then this is for you.
3
A paid browser agent is enough
/ 10 The idea The tools Demo Real life Take-home
to reproduce this issue
FREE CHAT BROWSER
Talks
Drafts test steps
ChatGPT
Claude
Gemini
Copilot
Think about
using it (more)
PAID, ~$20 BROWSER AGENT
Acts
Clicks in DSpace
DESKTOP APP
TODAY'S DEMO
US ONLY
US ONLY
You can and should
use it
EDITOR TOOLS DESKTOP
Builds
In your editor
ANTIGRAVITY
You should
use it
CLI + MCP CLI AGENT
Repeats
Prompt in, proof out
ANTIGRAVITY
You probably
already do
OpenAI, Anthropic, Google, Microsoft, GitHub plan and help pages, checked September 2026
4
Write the rule, get evidence a manager can read
/ 10 The idea The tools Demo Real life Take-home
EXAMPLE OUTPUT
TESTED
The steps you asked for
RESULT
Fails our rule
EVIDENCE
Screens, saved values
STILL UNCERTAIN
What it could not check
You sign off
YOUR REQUEST
WHERE
Test/Local environment.
Live site: read-only.
DO
Each step the user took, so anyone can repeat it.
MUST (NOT) HAPPEN
The rule the issue breaks, in one sentence.
SHOW ME
Each step, compared with our rule,
and what you could not check.
5
Two agents reproduce the same failure
/ 10 The idea The tools Demo Real life Take-home
PROMPT AS TYPED
Recreate this issue in vanilla DSpace:
-
Record a video/gif proofs of this process.
-
Open demo.dspace.org in a web card.
-
Log in as administrator.
-
Add 3x same-type ("Type") options while creating a new Item (for example "Book").
-
When finished: create a .html infographics with simply describing the result & save captured proofs to my Downloads folder.
-
Text result in chat in a form of emoji "thumbs-up" / "thumbs-down" if succeeded.
ANTIGRAVITY
CLAUDE DESKTOP
6
Many tools, many prompts, checks we keep
/ 10 The idea The tools Demo Real life Take-home
OUR DSPACE SITES
TEST CODE
REWORD THE PROMPT
TOOLS WE TRIED
Many tools, no standard.
One of our workflows:
teams of CLI agents.
The prompt is a skill.
The rule is the key.
Retries are normal.
Checks we keep, as code.
Bigger tests, several sites.
AI helps write the code.
8
Three questions turn an AI pass into trust
/ 10 The idea The tools Demo Real life Take-home
AI SAYS PASS
TRUSTED
Checked against what?
Your rules.
The AI does not know
them. You write them.
Could it be wrong?
Yes, sometimes.
Not 100%, but far better
than no check. Use it.
Who signs off?
A person, never the AI.
The AI only changes
the test environment.
Save the check with the issue. Rerun it after relevant changes.
9
Thank you
Questions?
Jozef Mišutka