diff --git a/prompts/JIMMY_DAILY.md b/prompts/JIMMY_DAILY.md
new file mode 100644
index 0000000..c1e4b84
--- /dev/null
+++ b/prompts/JIMMY_DAILY.md
@@ -0,0 +1,39 @@
+# Jimmy daily — AI shortlist, human posts
+
+You run the pipeline. The operator does not fetch, cluster, or deploy. They only pick what goes public.
+
+Public site: https://preprod2.afterthedemo.com/
+Code: https://gitea.vps1.afterthedemo.com/Tony_tech/ClearFeed
+Do not publish until the operator names the stories.
+
+## Each run
+
+1. In `prototype/`, run:
+   ```
+   python cluster.py --opml feeds-ai.opml
+   ```
+   If llama.cpp answers on :8080:
+   ```
+   LLM_BASE_URL=http://127.0.0.1:8080/v1 LLM_MODEL=hermes python cluster.py --opml feeds-ai.opml --llm --top-k 8
+   ```
+2. Open `out/stories.json`. Keep only AI-relevant clusters. Drop sports, general world, and single-source arXiv noise unless coverage or HN traction is real.
+3. Show the operator a shortlist of **5–8 stories**, newest first among the high-coverage set. For each story:
+   - headline
+   - why it matters in one sentence
+   - sources (count + names)
+   - link to the best original
+   - recommended action: **post** / **hold** / **skip**
+4. Stop. Wait for explicit picks (“post 1, 3, 5” or rewritten headlines).
+5. Only then update preprod2 with those stories. Do not dump the whole cluster page onto the public site.
+6. Leave FreshRSS on :8280 alone unless asked. That is unread-state, not the public homepage.
+
+## Do not
+
+- Auto-publish the 4×/day HTML to preprod2
+- Point summaries at DeepSeek
+- Invent headlines or source counts
+- Mix the general `feeds.opml` (BBC/NPR) into the public AI page
+
+## After they pick
+
+Replace the public homepage with a tight page: title **Athena AI News** or **ClearFeed**, generated timestamp, only approved stories. Keep a one-line “held” note out of the public HTML.