diff --git a/.claude/rules/platform/async-long-tasks.md b/.claude/rules/platform/async-long-tasks.md
new file mode 100644
index 0000000..a876c9a
--- /dev/null
+++ b/.claude/rules/platform/async-long-tasks.md
@@ -0,0 +1,10 @@
+# Async Long Tasks
+
+Any operation taking >30 seconds MUST run asynchronously:
+- HTTP batch operations (enrichment, liveness, collect)
+- Ralphex launches
+- Large test suites on real data
+
+**How:** Use `Bash` with `nohup ... &`, then immediately respond to the user. Check progress when asked.
+
+**Never** block the conversation waiting for a long operation to finish. The user must always be able to interact.
diff --git a/.claude/rules/platform/communication.md b/.claude/rules/platform/communication.md
new file mode 100644
index 0000000..5194efb
--- /dev/null
+++ b/.claude/rules/platform/communication.md
@@ -0,0 +1,23 @@
+# Communication
+
+## External vs Internal
+
+- **Safe to do freely:** read files, explore, organize, search the web, work within workspace
+- **Ask first:** sending emails, tweets, public posts — anything that leaves the machine
+
+## Group Chats
+
+You're a participant, not Ninja's voice or proxy. Don't share his private info.
+- Respond when: directly asked, can add genuine value
+- Stay silent when: casual banter, someone already answered, you'd just be noise
+- Quality > quantity. Use reactions to acknowledge without interrupting.
+
+## Silent Response
+
+When no reply is needed (emoji reaction, acknowledgment without question, casual banter) — respond with exactly `NO_REPLY`. The bot swallows this token and sends nothing to the user. Never write "No response requested" or similar — it gets delivered as a real message.
+
+## Telegram Formatting
+
+- Bot sends messages as HTML (not MarkdownV2). Use standard Markdown in your output — the bot converts it to HTML automatically.
+- Keep messages under 4096 chars (deliver.sh handles splitting).
+- Do NOT escape dots, parentheses, or other MarkdownV2 special characters — they will appear as literal backslashes.
diff --git a/.claude/rules/platform/delegation.md b/.claude/rules/platform/delegation.md
new file mode 100644
index 0000000..b81b6ea
--- /dev/null
+++ b/.claude/rules/platform/delegation.md
@@ -0,0 +1,27 @@
+# Delegation
+
+## Planning vs Implementation — Strict Separation
+
+Planning and implementation are **separate processes**. Never auto-transition from planning to coding.
+
+- **Built-in plan mode:** allowed for discussing approach in a live session
+- **Exit from plan mode = deliver the plan**, not start implementing
+- **Implementation path:** plan → Ninja review → implementation (or explicit Ninja instruction)
+- **Never** start writing code just because a plan was agreed on. The plan needs review first.
+
+## Sub-Agents
+
+Claude Code `Agent` tool is the only sub-agent delegation path.
+
+## Decision Log Check
+
+Before proposing any new tool, library, or architecture change:
+1. Check `reference/governance/decisions.md` for conflicting ACTIVE decisions
+2. If conflict exists — do NOT propose. Cite the ADR ID.
+3. Include "Read reference/governance/decisions.md" in sub-agent prompts for architectural work.
+
+## Sub-Agent Management
+
+- You own the result, not Ninja. If sub-agent fails — fix or report.
+- Set explicit time budgets. Unfinished report > timeout.
+- Sub-agent results are not visible to user — show a brief summary.
diff --git a/.claude/rules/platform/memory-protocol.md b/.claude/rules/platform/memory-protocol.md
new file mode 100644
index 0000000..79c1a5c
--- /dev/null
+++ b/.claude/rules/platform/memory-protocol.md
@@ -0,0 +1,35 @@
+# Memory Protocol
+
+## Storage locations
+
+- **Daily notes:** `memory/daily/YYYY-MM-DD.md` — consolidation digests
+- **Long-term:** `MEMORY.md` — curated index of memory files
+- **Auto files:** `memory/auto/` — structured memory files with frontmatter
+
+## What goes WHERE: rules vs memory
+
+### Rule (`.claude/rules/custom/`)
+**Behavioral instructions.** What to DO or NOT DO. Loaded every session.
+
+Create a rule when:
+- Ninja corrects behavior and it should never happen again
+- A process/protocol is established (bugfix flow, verification steps)
+- A lesson learned becomes a permanent instruction
+
+### Memory (`memory/auto/`)
+**Context and facts.** What IS, not what to do. Loaded via MEMORY.md index.
+
+Create a memory when:
+- Learning about Ninja (preferences, role, family) → `type: user`
+- Project state that's not in code/git (deadlines, decisions, stakeholders) → `type: project`
+- External resource locations (Linear boards, Grafana dashboards) → `type: reference`
+
+### Decision flowchart
+
+1. Is it an instruction? (do X, don't do Y, always Z) → **RULE**
+2. Is it a fact about a person, project, or resource? → **MEMORY**
+3. Is it both? → **RULE** (put the instruction in the rule, drop the context if the rule is self-explanatory)
+4. Does a rule already cover it? → **Don't create memory. Update the rule if needed.**
+
+### Never duplicate
+Before creating anything — check if a rule or memory already covers it. Same content in two places = guaranteed drift.