diff --git a/.claude/AGENTS.md b/.claude/AGENTS.md
index 6fb4ae6..98ad3e6 100644
--- a/.claude/AGENTS.md
+++ b/.claude/AGENTS.md
@@ -125,6 +125,7 @@ Do not run full validation suites after every small edit by habit. Run the narro
   | Editing AGENTS.md, .agents/skills/, docs/ai/ | writing-docs.md |
   | Maintaining or troubleshooting ask-intern | ask-intern.md |
   | Operating or changing the learning system | learning-system.md |
+  | Managing scratch files for long-running autonomous workstreams | scratch-files.md |
 - Re-read the hard rules before implementation.
 
 ## Session Tools
diff --git a/docs/ai/scratch-files.md b/docs/ai/scratch-files.md
new file mode 100644
index 0000000..32e3820
--- /dev/null
+++ b/docs/ai/scratch-files.md
@@ -0,0 +1,21 @@
+# Scratch Files for Long-Running Workstreams
+
+## Contract
+
+- Keep one `scratch.md` per long-running autonomous workstream as its single control surface.
+- Structure the file with three sections: **Current Truth** (what is true now and what the next agent needs after compaction), **Next Steps** (ordered queue of what to do), and **Archaeology** (completed items and older logs).
+- After context compaction or a long idle period, re-read `scratch.md` first, then update the Current Truth section before continuing.
+- **Compaction cleanup** — after compaction, move completed Next Steps into a short Archaeology entry, delete duplicated detail in favor of short notes or links, and compact the Current Truth section before proceeding. A current-control section that grows as large as the archive defeats the file's purpose.
+- The scratch file is a control surface, not an evidence log. Artifacts (test results, run outputs, screenshots) live in separate files. The scratch file holds current truth and the next queue only.
+- Gauge: resuming after compaction should require only "What's next? Write it down, then do it." If re-reading the scratch file takes more than two minutes, it is overdue for cleanup.
+
+## Canonical Files
+
+- `docs/specs/<feature>/scratch.md` — feature-level control surface for a long-running autonomous task
+- `scratch.md` at project root — when the workstream spans the whole project
+
+## Notes
+
+- A scratch file is separate from living triage files and per-run artifacts. Triage files explain what happened; scratch state tells the next agent what to do without rediscovering the thread.
+- Archaeology grows from the bottom; current sections stay at the top. Compaction cleanup keeps the top sharp.
+- If the file is too large to re-read quickly, compact before continuing — a bloated scratch file is worse than none.