I built this platform with teammates so I wouldn’t become my team’s AI bottleneck. Any researcher can write and publish their own AI research skills on it, and the quality and compliance patterns come built into the template. I owned the governance model and seeded the platform with three foundational skills. Colleagues have since shipped their own.
01 — Problem AI could speed up planning and synthesis, if we could trust it
Our team of eight spent 4–5 hours per study on planning and close to a day on synthesis. The work was manual and inconsistent, and it capped how much research the org could ship.
We were handling sensitive employee data at Apple. The open question was whether AI could save that time without inventing evidence that was never in the data, exposing employee information, or trading defensible method for speed.
Our earlier generative AI studies had already taught the team to be wary of hallucination.
The standard
For our team, saving time was only worth it if the research still held up.
02 — Who built what I shared the strategy and took the riskiest skills
I shaped the platform and its repository with a fellow researcher and a principal product designer, and the three of us set up the PR process together. I built the planning skill, then took the two riskiest ones, synthesis and synthetic usability. The researcher went on to write a review skill, and the designer and another researcher wrote a research repository skill. That’s what we’d hoped would happen.
03 — Guardrails The rules live in the template
Every skill follows one rule: be aggressive about saving time and conservative about protecting judgment, compliance, and research integrity. Building that into the template, where no author has to remember it, is what turned a personal toolkit into a platform.
I started with planning, the least flashy stage, on purpose. The process was already templated and well structured but still full of friction, and it’s where every study begins. If the model held up there, I could take on the harder stages.
HUMAN
PR GATE
Any researcher can publish, as long as the skill starts from the template and clears the gate.
How a skill reaches the team
Every skill inherits the same four guardrails and passes a PR gate, reviewed by AI and by a person, before it reaches the team. A teammate writing their own skill doesn’t have to solve trust all over again.
What every skill inherits
Testing the riskiest skills for hallucination
Hallucination and sycophancy worried me most in the two highest stakes skills, synthesis and synthetic usability. I benchmarked their output against real findings a human researcher had already produced, and kept iterating until it held up.
04 — Evidence Inside one skill
This is the documentation for the planning skill, where I started. It’s a representative rendering, with team, tool, and policy names generalized for confidentiality.
05 — Rollout Setup is where people got stuck
I got people unstuck by showing them that Claude itself could walk them through installing and troubleshooting, so they could manage on their own. Researchers took to the skills themselves because they followed our own process, cited established methods, and stayed within Apple’s legal and privacy policy.
The numbers
Taking up to 85% off planning gives researchers more time for the judgment calls, which I don’t think AI should be making.
The two time figures compare how long a study typically took to plan or synthesize before the skills with how long it takes now. They’re typical times, not instrumented measurements.
Beyond the hours, junior researchers and contractors stay within our framework and policy through the skills, and learn it by using them. And the skills are opening up to the wider UX and design team, so that team can run its own research safely.
06 — Hindsight The platform mattered more than the skills
The choice I’d make again was putting the time into the platform and standards that let my teammates build their own skills.
One thing I’d do differently: gauge the team’s AI familiarity first and start with the basics, so everyone could troubleshoot with AI on their own before I rolled the platform out. That readiness step would have taken most of the setup friction off the table.