Research Operations · 2026–Present

Building a governed AI research platform

01 — ProblemSpeed the team could still trust

An 8-person team spent 4–5 hours per study on planning and roughly a full day on synthesis. Manual, inconsistent, and a ceiling on how much research the org could ship.

We were handling sensitive employee data at Apple. AI could clearly save time; the question was whether it could do that without three failure modes:

Most AI tooling optimizes for speed and ignores that second half. Prior generative-AI studies had already earned the team an informed skepticism of hallucination.

02 — RoleCo-strategist, and owner of the hardest parts

I co-strategized the platform and repository with a fellow researcher and a principal product designer, and owned the two hardest, highest-impact parts myself. My collaborators stood up the repository and PR process with me. The researcher later contributed a review skill; the designer and another researcher authored a research-repository skill, which became proof the platform worked without me in the loop.

03 — ApproachStart where the payoff was real

I started with the least flashy stage on purpose: our planning process was already templatized, high-friction, and well-structured, the natural front of the research lifecycle. It let me prove the model before tackling harder stages.

Every skill then followed this principle:

That discipline living in the template, not in any one author's diligence, is what turned a personal toolkit into a platform.

The four guardrails every skill inherits

I validated the risky skills empirically

Hallucination and sycophancy were my biggest worry in the two highest-stakes skills, synthesis and synthetic usability. I benchmarked their output against real findings a human researcher had already produced, iterating until it held up, holding my own tool to the same evidentiary bar as our studies.

04 — AdoptionGovernance drove adoption

The barrier was setup friction, not skepticism.

The breakthrough was teaching the team that Claude itself could help them install and troubleshoot, so they became self-sufficient. Researchers leaned in precisely because the skills respected their process, cited established methodology, and honored Apple's legal and privacy policy.

05 — ImpactFaster work with protected judgment

The up-to-85% planning reduction freed researchers for the judgment work AI shouldn't touch. That time moved from mechanical prep to thinking.

Three ways it compounded

06 — EvidenceInside one skill

A representative rendering of a single skill's documentation: what it does, the privacy and compliance guardrails it inherits from the shared template, and the cited methodology it is grounded in. Team, tool, and policy names are generalized for confidentiality.

A skill in depth

The planning skill

The full documentation for one skill: the guardrails it inherits from the template and the cited method behind its output.

View skill →

07 — ReflectionEmpowering a team through skills

The instinct that shaped this work was refusing a false choice between speed and rigor. For a research team, rigor is what makes speed usable at all.

The other choice I'd make again was building enablement. The platform and standards that let my teammates build their own skills mattered more than any skill I wrote myself. That is the difference between being productive and multiplying a team.

One thing I'd do differently: gauge the team's AI familiarity first and start with 101-level guidance, 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.