Researcher control
The researcher retains authority over substantive decisions throughout the work.
A practical guide to my AI Operating System
A system for organizing AI-assisted work, from the first instruction to the final handoff.
v2.0 — September 15, 2026
The AI OS is a library of plain-text instructions that tells an AI agent how to work on your research.
appendix/.The researcher retains authority over substantive decisions throughout the work.
The AI-assisted process is documented so the user and/or others can inspect its inputs, intermediate work, decisions, and outputs.
For the fuller argument:With Great Powers ↗Read the chapter draft ↗
Across your work
The active task
Inside each project
Coordination & bookkeeping
Brief, instructions, sources, data.
Tasks, dependencies, gates, deadlines.
Progress, decisions, outputs.
Example: update a literature table.
Hook setup and delivery checks ↗
Technical installation and recovery guide ↗
Literature-table exampleFind the existing table, inclusion criteria, source folder, and notes from the previous session.
Literature-table exampleAgree which sources qualify, which columns to update, and what counts as a verified entry.
Literature-table examplePlan the search and update; bring changes to inclusion criteria back to the researcher.
Literature-table exampleUpdate the claimed table, retain source links, and mark unresolved metadata for checking.
Literature-table exampleCheck source details, inclusion decisions, and missing fields; flag entries needing researcher review.
Literature-table exampleLeave the updated table, its verification status, and a short list of unresolved sources for the next task.
Merged release and validation ↗
The user supplies preferences; the agent explores measures and sources; the user selects the operationalization.
Collect each source and record its provenance. Preserve the original data.
Clean each source separately. Save the transformation scripts and checks.
Summarize each cleaned source: coverage, distributions, and missingness. The user resolves merge rules before the sources are combined.
Join the cleaned sources using the agreed rules. Check unmatched records and produce one analysis-ready dataset.
Package the sourcing, cleaning, descriptive analysis, and merging code with instructions. Verify that it reproduces the dataset; the user accepts the package as ready for analysis.