diff --git a/AGENTS.md b/AGENTS.md
index 96c5e4b..b463fd5 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -124,3 +124,29 @@ This repository uses one trust gate. Every session must use it and must not bypa
 - Keep generated trust configuration and this managed block in place.
 
 <!-- CorvidLabs trust toolchain: END -->
+
+
+## Human intent
+
+This repository should use [hi (Human Intent)](https://corvidlabs.xyz/hi): plain
+sentences saying what people want, each with a permanent id, kept in `hi/`.
+Tickets and specs are generated from them.
+
+If there is no `hi/` here yet, start one from real work rather than from the code:
+
+1. When I ask for a feature, draft its criteria first — one plain sentence each,
+   about what somebody **wants**, not what the code does. Private test: you should
+   be able to put *As a ___,* in front of it. Leave those words out of the file.
+2. Show them to me and stop. Capture nothing I have not agreed to.
+3. Capture what I confirm, one per command: `hi SEND-1 "the sentence, in my words"`.
+   You pick the id. Letters are cases and numbers are steps (`SEND-1.a.1`), and an
+   id is permanent and never reused, so choose like you will say it out loud.
+4. Then build. `hi check` fails only on a structurally broken file, never on
+   unfinished intent.
+
+Do this before every feature, not only the first one. Do **not** bulk-generate
+criteria from the code or from existing tickets: a hundred plausible sentences
+nobody said is worse than five real ones, and an id spent on a wrong sentence is
+spent forever.
+
+Install: `brew install corvidlabs/tap/hi`, or `cargo install human-intent`.