LESS SUPERVISION. MORE CONFIDENCE
Adversarial Review
Getting Reliable Results from Your AI Harness on the first run
Getting AI to do the job right without constant supervision
10 September 2026
LESS SUPERVISION. MORE CONFIDENCE
Getting Reliable Results from Your AI Harness on the first run
Getting AI to do the job right without constant supervision
ABOUT ME
At Novo Nordisk
Helping engineers use AI effectively
DIGITOMNIA APS
Six months building Agent-Flow
THE REAL PROMISE OF AI
What I need from AI
When I ask AI to do a job, I want it to carry out the work like a reliable colleague while I focus on other things and return a result I can rely on that is exactly what I asked for.
I have more ideas than time. I want AI to move the work forward without turning my free time into another review shift.
Reliable delegation
AI · No problem. I’ll get right on it…
AII’m done.
It looks really bad on mobile.
The images are cropped.
The text is too small to read.
The tests don’t pass.
There are errors in the console.
The colours are very bad.
The images are pixelated.
And on and on…
Why am I doing checks the AI could have done?
FROM SUPERVISION TO CONFIDENCE
What I learned building
my own AI workflow
PLAN TOGETHER
Agree what you want and how to check it
Ground the approach in guidance and current sources
Challenge and refine the plan before building
SELF CHECKING CAPABILITIES
Render, inspect and exercise the result
Run checks, read failures and try again
Check the information behind the approach
A DIFFERENT PERSPECTIVE
The important detail
can be the one
you didn't notice.
Give an independent critic the job of
finding errors, issues or problems.
A FRESH PAIR OF EYES
Find issues in the code and back them with evidence
Another AI, a clean context and the original brief
Resolve supported findings and inspect the changes
THE COMPLETE WORKFLOW
From voice request to reviewed plan.
Built, tested, verified and delivered
Refine a new request from wherever you are
Speak into watch
n8n workflow
GitHub issueI really like what we have done with the MCP server that serves data and MCP UI widgets and also the sites dashboard. I basically want clone of the entire project and what we have done but now instead of financial info I want it to show health data. So basically the kind of info that is available in Apple Health like sleep, heart rate, fitness, diet, exercise.
Make up some deterministic data for the last year for "me" and then have an mcp that can server useful insights into the data so I can ask stuff like "Show me my sleep breakdown for the last month/week" and I get various metrics about sleep time, deep sleep, bed time, duration etc in a nice graph with different style like line, bar with pickers for what I want to see. Then a full health dashboard just like the existing finance one.
So basically a clone demo of the existing stuff but for health. I will want to link it as a custom dev plugin in chatgpt just like before and I will call it "Health Insights".
It might be that the existing stuff in folders now needs to be moved to a subfolder (one per plugin) so please do that so the repo files are organised easily. Once done I will test the main site dashboard locally and then I will get you to deploy to chatgpt sites location and will manually link it up. So please make appropriate installation guide for this plugin also like before.
Human + AI · The planning workflow
Health Insights starts in GitHub.
Inspect the project and sources.
Define scope and acceptance.
Challenge the proposed approach.
Resolve choices. Approve.
An agreed plan to act on.
AI execution · Human approval at the finish
Build starts · Telegram notification.
Data · MCP tools · Charts
Desktop and phone views
Check the chart behaviour.
Check the full health request.
“Geoffrey” notifies you in Telegram.
Review Health Insights.
Accept the result.
FROM REQUEST TO WORKING SOFTWARE



ON YOUR NEXT TASK
Make success explicit and checkable.
Let it see and exercise the actual result.
Challenge independently, correct and check again.
A SMALL CHANGE TO TRY
Agree on what success looks like—and how the AI can check it.
Then let review and correction happen before the work comes back to you.
Let's connect on LinkedIn