Teams still argue about "AI-assisted" versus "vibe coded" without agreeing what changed in the workflow. The 0rwx framework offers a compact vocabulary: four permission-style levels, from zero AI access through full agentic generation, named with the Unix letters 0, r, w, and x. The site stresses that the levels track how much you can scrutinize the result, not which product logo appears in your dock.
This explainer walks through each level, ties them to examples the author lists on 0rwx.com, and connects the idea to disclosure fights on game stores, coding-agent guardrails, and review habits we have covered elsewhere on explainx.ai.
TL;DR: the four 0rwx levels
| Level | Unix letter | Tagline on 0rwx.com | Human role | Example from the site |
|---|---|---|---|---|
| ZERO | 0 | "I think" | Hand-made work, no model reads or writes your artifact | You read the docs and write the code yourself |
| READ | r | "I make" | AI may read your work to advise, not change it | Spell-check style suggestions, chord ideas you reject or accept consciously |
| WRITE | w | "I fix" | AI may edit, but you can still read and correct the output | Autocomplete loops, refactors, collages you inspect before shipping |
| EXECUTE | x | "I curate" | Output is opaque; you choose among results and test behavior | Full agent runs where you cannot realistically audit every line |
The Dijkstra quote on the landing page is decorative philosophy, not part of the taxonomy. The operational definition is scrutiny.
Why Unix permissions are a useful metaphor
Unix file modes answer a narrow question: who may read, write, or execute a file. They do not name the editor. The 0rwx author applies the same shape to LLM workflows because tool churn is fast—today's autocomplete plugin becomes tomorrow's autonomous agent—while the review obligation changes more slowly.
The WRITE band on 0rwx.com lists examples such as reading generated code, building collages, and loop generation—work you can still open and edit. That matches how many senior engineers describe comfortable agent use: digital pair programming where you retain diff review, not blind delegation.
If you manage agents in production, the metaphor also lines up with how security folks talk about least privilege. Our posts on Octomind's TOML guardrails for coding agents and loop engineering for unattended Claude Code runs push the same idea in a different vocabulary: constrain what automated tools may touch, and keep a human verification step where stakes are high.
Level ZERO: hand-made work
At ZERO, the model has no read or write access to the artifact you are shipping. You might still use web search or offline references, but the creative or technical object—source file, score, painting—never enters an LLM context window for editing.
0rwx.com lists classic hand work: reading official documentation, painting with physical paint, playing an instrument without generative backing tracks. In software terms, this is the "I wrote it in vim and no assistant saw the buffer" story.
ZERO is rare in day-to-day product engineering, but it still matters as a baseline when teams negotiate credit, liability or academic authorship. Kevin Boone's essay on AI-generated web content and trust argues readers deserve to know when prose never passed through a model; 0rwx gives a process label for the opposite case.
Use ZERO when:
- You need a hard line for compliance or licensing (some clients still forbid model training on their specs).
- You are benchmarking models and must keep the task set pristine.
- You are teaching fundamentals and want learners to struggle without autocomplete.
Level READ: AI as a research assistant
READ grants the model read-only access. It may comment on your draft, suggest chords, or answer questions about an API, but it does not commit changes into your artifact. The tagline is "I make"—you remain the author of record.
Image: 0rwx
Think of IDE hover help, a grammar checker that underlines mistakes without rewriting paragraphs, or a chat where you paste a function signature and ask for pitfalls. You accept or reject advice; the file on disk only changes when you type.
On 0rwx.com, READ includes "I write the code" style workflows (the page links the Brainmade mark for this)—reference images and chord inspiration where you remain the author. It maps cleanly to early Copilot modes that would not apply a patch until you tab-complete, and to the lighter end of Microsoft Copilot as a research assistant rather than an autonomous coworker.
READ is underrated in hype cycles because it sounds boring. For many senior engineers, it is the highest band they will stay in for production code: models accelerate lookup and rubber-ducking without touching git history.
Level WRITE: collaboration you can still audit
WRITE allows the model to modify your work inside a loop you can inspect. The motto is "I fix"—you still read diffs, run tests, and reject bad patches.
Image: 0rwx
Examples on 0rwx.com include reading generated code, autocomplete you watch line by line, chord loop generation you tweak, and image collages built from pieces you can still open in an editor.
This is where most "AI pair programming" should live. Addy Osmani's widely shared essay on cognitive surrender when using Claude Code warns that speed without comprehension drifts toward EXECUTE even if the tool UI still shows a diff view. The 0rwx label asks an honest question: could you explain the patch in a code review? If yes, you are in WRITE.
Practical habits that keep WRITE honest:
- Keep changesets small enough for human review, the same discipline we recommend in Worktrunk's parallel agent workflow.
- Run tests before merge; agents do not replace CI.
- Log which prompts produced which hunks when working on regulated code.
Platform disclosure rules rarely distinguish READ from WRITE, but your internal retros should.
Level EXECUTE: curation without line-by-line audit
EXECUTE is agentic generation where you cannot realistically read everything the model produced. The human job is to curate outputs, run behavioral tests, and decide what ships. 0rwx uses "I curate" and warns that you may no longer read your own artifact—think generated albums, whole mini-apps from a single prompt, or long agent trajectories.
Image: 0rwx
This is the territory marketplace moderators worry about. Our coverage of Steam Next Fest filling with AI slop and Nexus Mods' mandatory AI disclosure shows platforms struggling to label content after the fact. 0rwx instead labels process: did the author have scrutiny bandwidth?
EXECUTE is not automatically bad. Prototyping, throwaway marketing assets, or exploratory game jam builds can live here if stakeholders know the review model is "test the binary, not the source." Risk rises when EXECUTE output ships to paying users without automated checks. Epic's vibe-coded launcher preview is an example of a major vendor openly embracing the band while claiming performance wins—buyers still need independent verification.
If you run agents that act without per-step human approval, pair EXECUTE-level honesty with hard guardrails. Tools such as AgentBeam, the agent security platform from the explainx.ai team, exist to stop agents before they take dangerous actions when scrutiny breaks down.
How 0rwx compares to store disclosure checkboxes
Valve, Nexus, Etsy and others ask whether AI generated visible assets. Those policies matter for consumers browsing screenshots. They do not tell a maintainer whether you understood the networking layer your agent added.
| Question | Platform AI checkbox | 0rwx level |
|---|---|---|
| Can a reviewer audit the diff? | Not usually asked | Central |
| Which vendor you used | Sometimes | Irrelevant |
| Legal/branding for store listing | Yes | Complementary |
| Team retros on process | Rare | Designed for this |
Use both: check the store box and agree on a 0rwx band in your README or pull request template.
Applying 0rwx on a real team
- Pick a default band per repo. Libraries might cap at WRITE; hackathons might allow EXECUTE with a warning label.
- Put the label in the PR description. "0rwx: w" beats a vague "AI helped."
- Escalate scrutiny when stakes rise. Security modules, billing code, or medical content should drop from WRITE to READ unless a named reviewer signs off.
- Revisit after tool upgrades. A plugin that used to suggest single lines may start multi-file refactors; your label should move with it.
Hacker News carried the 0rwx launch on November 7, 2026 (story id 50405734, six points, zero comments as of this review). Treat the framework as early shared vocabulary, not an industry standard. The site's source lives in a public GitHub repository, where issues and PRs are welcome.
Credit and licensing for the logos
0rwx.com states the level logos are released into the public domain and may be reused. That makes the labels easy to drop into slide decks or CONTRIBUTING.md files without a separate trademark fight—still cite 0rwx.com so readers know where the definitions live.
Bottom line
0rwx does not solve disclosure law or agent security. It gives builders a permission metaphor that survives tool rebrands: ZERO, READ, WRITE, EXECUTE track how much you could scrutinize before shipping. Start labeling internal work there; pair it with platform rules and automated tests when EXECUTE output faces the public.