Humans.md: tell your AI agents how to work with you A new HUMANS.md specification lets users declare interaction preferences that AI coding agents load at session start, with plugins demonstrated across Claude Code 2.1.302 (plugin 0.2.1), Codex CLI 0.163.0 (plugin 0.3.0), Devin CLI 3000.11.92 (plugin 0.3.0) and OpenCode 2.0.32 (plugin 0.3.0) on 17 Oct 2026. The example profile sets interrupting.interrupt_for to blocking_only, deciding.batch_questions to end_of_task, deciding.reversible to decide_and_report and deciding.question_style to yes_confirms_recommendation, and was last updated on 2026-10-13. The spec states preferences shape how agents write to users but never grant a permission, and that how closely a model follows the file is not measured. $ claude -p "What does my HUMANS.md say about interruptions?""WhatdoesmyHUMANS.mdsayaboutinterruptions?" TL;DR: Your HUMANS.md says to interrupt you only when blocked. Everything else waits until the end of the task. Setting Value Meaning interrupting.interrupt for blocking only Interrupt only when work can’t continue deciding.batch questions end of task Non-blocking questions are grouped at the end deciding.reversible decide and report Reversible choices are made, then reported deciding.question style yes confirms recommendation A plain “yes” accepts my recommendation Example: -Missing credentials for a required service → I interrupt you. -Picking a variable name or a table layout → I decide and mention it at the end. The profile was last updated on 2026-10-13 and isn’t flagged as stale. Real session, Claude Code 2.1.302 with plugin 0.2.1, 17 Oct 2026, with the example profile from the spec. The plugin loads your file at session start; how closely a model follows it is not measured here. Preferences shape how agents write to you. They never grant a permission. Another agent? Cursor, GitHub Copilot CLI, Google Antigravity, JetBrains Junie, Cline and Warp: install commands for each. codex $ codex ⠋ Loading your HUMANS.md ›What does my HUMANS.md say about interruptions?WhatdoesmyHUMANS.mdsayaboutinterruptions? •Only interrupt you when work is blocked. Save non-blocking questions for the end of the task, with a recommended answer for each. Worked for 8s Real session, Codex CLI 0.163.0 with plugin 0.3.0, 17 Oct 2026. The plugin loads your file at session start; how closely a model follows it is not measured here. Preferences shape how agents write to you. They never grant a permission. Another agent? Cursor, GitHub Copilot CLI, Google Antigravity, JetBrains Junie, Cline and Warp: install commands for each. devin $ devin -p "What does my HUMANS.md say about interruptions?""WhatdoesmyHUMANS.mdsayaboutinterruptions?" --respect-workspace-trust false Your HUMANS.md specifies that I should interrupt only for blocking issues. TL;DR : You only want interruptions when I'm blocked. | Setting | Value | |---------|-------| | Interrupt for | Blocking only | This means I should continue working autonomously unless I encounter a blocker I cannot resolve myself. Real session, Devin CLI 3000.11.92 with plugin 0.3.0, 17 Oct 2026, with the example profile from the spec, printed as is. The plugin loads your file at session start; how closely a model follows it is not measured here. Preferences shape how agents write to you. They never grant a permission. Another agent? Cursor, GitHub Copilot CLI, Google Antigravity, JetBrains Junie, Cline and Warp: install commands for each. opencode $ opencode run "What does my HUMANS.md say about interruptions?""WhatdoesmyHUMANS.mdsayaboutinterruptions?" TLDR: Your HUMANS.md says interrupt only for blocking issues. Since nothing is blocking me right now, I won’t interrupt you. No HUMANS.md file exists in your workspace it’s empty , so this comes from the profile data in your environment context: Setting Value interrupting.interrupt for blocking only Per your reading preferences, I’ve led with the TLDR and stayed under your 149-word limit. I’ll proceed silently unless I hit something that blocks progress. One non-blocking question answer at end of task : Your reading settings cap responses at 149 words and prefer tables/examples — shall I keep replies that compact going forward? Real session, OpenCode 2.0.32 with plugin 0.3.0, 17 Oct 2026, with the example profile from the spec; the reply text, from its JSON output. It first searched the folder for a HUMANS.md and found none: the values came from the plugin. How closely a model follows them is not measured here. Preferences shape how agents write to you. They never grant a permission. Another agent? Cursor, GitHub Copilot CLI, Google Antigravity, JetBrains Junie, Cline and Warp: install commands for each. Your agents run in parallel. Your attention doesn’t. La nouvelle salle de téléphones de Gutenberg : demoiselles téléphonistes. Agence Meurisse, 1912. Bibliothèque nationale de France, Gallica btv1b9021257n, public domain, via Wikimedia Commons. Every line went through an operator. She decided which call got through, and when. That was the Gutenberg exchange, Paris, 1912. Your agents ring one desk too: yours. HUMANS.md tells them which calls may ring now, and which wait for the end of the task. Tableau standard pour Science et Vie : standard téléphonique. Agence Meurisse, 1913. BnF, Gallica btv1b9021818m, public domain. One file. Every session. One human. The noiseThe labelThe calm Ten sessions, one of you Two tools, ten sessions running at once. Half of what they send needs you, a decision or a check; half is reports that do not. Both arrive the same way, in the middle of your work. By the evening: AI fatigue, brain fog. The label When a session starts, it reads ~/.humans/HUMANS.md once: how you read, how you decide, when to interrupt you. Every session gets the same copy. The calm Sessions that read it can write the way you asked: a TL;DR first, questions saved for the end of the task, and only a blocked session reaching you now. An illustration of how the file is used, not a measurement Preferences shape how agents write to you. They never grant a permission. The same work, reported with and without HUMANS.md. Switch it off to compare. Preferences shape how agents write to you. They never grant a permission. An illustration written for this page auth-tests · agent session · 14:40 HUMANS.mdOffOn Migrate the auth tests to the new session API. Migrated 38 test files · 212 tests pass 3 files need the staging identity provider Stopped at 14:12 and waited 18 min: “Is it okay to rename authHelper.ts?” Renamed authHelper.ts without stopping: reversible, reported below Without HUMANS.md Great news I’ve been working through the auth test migration and wanted to give you a full update on where things stand, since there were quite a few moving parts along the way. First, I looked at the existing setup. The old tests used the legacy SessionStore mock, which was tightly coupled to the cookie middleware. Before I go further: do you want me to keep the legacy mock around for the admin tests, or remove it entirely? Anyway, I went ahead and migrated 38 of the 41 test files to the new session API. Along the way I noticed that the token refresh tests were flaky, which seems to be caused by a hard-coded clock. Should I switch them to fake timers, or would you rather I leave them as they are for now? I also renamed authHelper.ts to session-helpers.ts to match the rest of the folder. Let me know if that’s okay or if you’d prefer the old name. Three files still fail: oauth-callback.test.ts, sso.test.ts and logout-all.test.ts. They depend on the staging identity provider, which I can’t reach from here. Could you check whether the staging credentials have changed? Overall I think the migration is in a good place, although there is definitely more we could do, such as consolidating the fixtures and maybe adding contract tests for the provider. Let me know how you’d like to proceed, and whether you want me to open a PR, merge it and deploy to staging With HUMANS.md TL;DR: 38 of 41 auth test files run on the new session API. 3 are blocked on the staging identity provider. Files Result Migrated 38 212 tests pass Blocked 3 staging IdP unreachable Decided, easy to undo: Token refresh tests now use fake timers; the hard-coded clock made them flaky. Renamed authHelper.ts to session-helpers.ts to match the folder. 2 questions. “yes” approves my pick: Delete the legacy SessionStore mock? My pick: yes, nothing uses it now. Rotate the staging IdP credentials so I can finish the 3 files? My pick: yes. Reply to the agent↑ 245 words · 4 questions spread through the text · 1 interruption · merge and deploy offered91 words, cap 149 · 2 questions, at the end, each with a pick · 0 interruptions Illustration of a scripted session, not a measurement. The plugin loads the profile; how closely a model follows it is not established. Lit: the fields this reply follows.Switch it on to see which fields change the reply. The session and both replies are written for this page, not recorded. The profile is the real example from the spec, byte for byte. What the profile changed in this replyWithout the profile, the same work reads like this TL;DR first 149 words at most A table, not paragraphs Reversible choices made, then reported Questions at the end, each with a pick No stop mid-task Pleasantries first, result later Long paragraphs, no table Four questions scattered in the text One stop mid-task, 18 minutes waiting Merge and deploy offered, unasked How it works ~/.humans/source.mdoptional TL;DR first, under 150 words. Show me a table or an example. If it can be undone, decide and tell me. Batch your questions at the end of the task, always with your pick, so a yes approves it. Only interrupt me when you are blocked. I often read on my phone. First you say it, in plain words. Any language. Include only the personal information you want kept. Compilingyour assistant “under 150 words”max words: 149 “decide and tell me”reversible: decide and report “only when you are blocked”interrupt for: blocking only Today the compiler is the assistant you already use. In Claude Code, the plugin’s /humans-md:compile does it; in other agents, the humans-md-compile skill. Review before writingyou lead with: tldrok max words: 149ok reversible: decide and reportok interrupt for: blocking onlyok visibility: privateok Left out or guessed: shown to you first. You approve every field. And who reads it. Write visibility: private explicitly: a profile without it is read as org. Session startintegration loaded~/.humans/HUMANS.md Reader checked. Recognized fields passed to the agent as data. not passed source.md, comments In Claude Code, Codex or Cursor, the humans-md plugin loads it. Agents read it at session start. To share it, a team repository holds a copy as humans/