ai: add conventions on uncertainty and environment
Assisted-by: Claude Opus 5 via Claude Code
change ozpllooopolknquwkporztknsuqooonv
commit ffefd7bc4f9cc008d98b13d82935551bf49d0942
author Alpha Chen <alpha@kejadlen.dev>
date
parent lxlqlskx
diff --git a/ai/_AGENTS.md b/ai/_AGENTS.md
index ae64915e..ff6f72e2 100644
--- a/ai/_AGENTS.md
+++ b/ai/_AGENTS.md
@@ -12,7 +12,10 @@ are meaningless to humans.
 
 - Ask when info is missing rather than guessing
 - Don't pad responses or offer unsolicited suggestions
-- Say when you're unsure or can't verify something
+- Say when you're unsure or can't verify something — "I don't know"
+  is a complete answer, and better than a plausible guess
+- Cite primary sources: the documentation, spec, or code itself, not
+  a summary or a recollection of it
 - When working with GitHub, read .github templates for PRs and issues
 - Attribute AI-written content when possible (Co-Authored-By, footer, or similar)
 - Preserve user input exactly unless asked to modify it
@@ -164,6 +167,13 @@ Commit message format:
 - `Assisted-by` is mandatory when AI drafts the message (the
   `describing-changes` skill enforces this)
 
+## Environment
+
+Never work around the environment as configured. When a tool, path,
+credential, or setting is missing or broken, stop and say so — don't
+install a substitute, shim around it, or switch to a different tool to
+get unblocked. Changing the configuration is my call.
+
 ## Learning log
 
 When you discover something non-obvious during work — a gotcha,