I Let AI Prepare Me for Every Meeting With My Manager
Measuring whether AI can autonomously prepare accurate, manager-ready summaries for recurring 1:1 meetings using a week's worth of work activity.
The Question
Can AI become a reliable record of my professional impact?
Microsoft recently made WorkIQ generally available, giving employees access to an AI system that can retrieve information across their work data. It’s marketed as a way to surface relevant information, understand context, and reflect how you work. I wanted to know whether it could solve a problem I’ve had throughout my career.
Why This Matters
Engaging in 1:1s has never been my strong suit. I’ve tried countless approaches to documenting my work week in hopes that I’d reconcile all activities into a comprehensive summarization of the progress made for the prior work week. For one reason or another, I’ve failed to stay the course for the entirety of a quarter and resulted to ad-hoc preparation.
While this may have been sufficient in prior roles, my current role would benefit most from a more concrete solution. For starters, my current manager has more than 10 direct reports. Of those direct reports, 3 of them also have teams that they manage. That means that my manager has more than 20 individuals that they’re responsible for retaining performance information about. My manager is also very active in their role which requires a healthy amount of meetings plus travel. Rather than leave the fate of my performance reviews to the mercy of ad-hoc 1:1s, I felt it necessary to take control of my career and implement a solution that would make my progress over the semester memorable…or at least retrievable.
Methodology
Workflow
To test whether AI could reliably prepare me for recurring 1:1s, I built two custom GitHub Copilot skills using Claude Haiku 4.5 as the underlying model.
The first, daily-debrief, uses WorkIQ alongside the GitHub MCP Server to retrieve and synthesize my work activity into a daily report. The second, weekly-brief, aggregates those daily reports over a specified time period and generates a manager-ready briefing for my weekly 1:1.
Experiment
The experiment spans four weeks.
At the end of each workday, I execute the daily-debrief skill and evaluate its output across four dimensions: recall, accuracy, relevance, and efficiency. The results are logged after each execution to establish a day-by-day record of the agent’s performance.
On the following Monday, I execute the weekly-brief skill to generate a summary of the previous week’s activity. The weekly briefing is evaluated using the same methodology to determine whether the workflow can accurately prepare me for an upcoming 1:1 with my manager.
Iteration
This experiment evaluates an AI workflow rather than a single prompt.
Whenever the output exposes shortcomings in retrieval, summarization, or overall quality, I revise the underlying GitHub Copilot skills before the next evaluation cycle. Every revision is documented alongside the corresponding results so that improvements in performance can be attributed to specific changes in the workflow rather than chance.
Experiment Timeline
Experiment opened. Hypothesis: With iterative refinement, an AI agent can reliably prepare manager-ready 1:1 summaries that accurately capture the majority of my work.
First daily brief run. Recall 100%, accuracy 89%, relevance 100%, cost $0.10.
Findings So Far
- For ~$0.10, AI produced a workday summary that was 89% accurate. This was for a day with no meetings but with calendar blocks for other tasks. I learned most of the token cost was due to cache read tokens (571.5K). Without caching, the 571.5K cached tokens would have been billed at the normal input rate of 571.5K × 100 credits/M = 57.15 AI credits. Instead, they were billed as cache reads: 571.5K × 10 credits/M = 5.715 AI credits. That’s roughly a 90% reduction on that portion of the input, which is why nearly 590K input tokens only ended up costing about $0.10 overall.
How to Follow This Experiment
This experiment publishes findings as they emerge, not on a schedule. Subscribe to the Pretty Wired Labs YouTube channel to be notified when new results are posted, or check back — the timeline updates with each new development.