diff --git a/apps/llm-mcp-reference/README.md b/apps/llm-mcp-reference/README.md
index 83b7ef0..5ca979e 100644
--- a/apps/llm-mcp-reference/README.md
+++ b/apps/llm-mcp-reference/README.md
@@ -2,6 +2,8 @@
 
 Secondary App-References tier: **LLM product façades** that call Traverse **workflows/capabilities** through MCP instead of encoding business logic in prompt “skills”.
 
+Shared prompts/workflows follow Specs 001/002 presentation vocabulary (`idle|loading|loaded|blocked|ended|error`) and still forbid inventing business fields — see [`shared/prompts/system-boundary.md`](shared/prompts/system-boundary.md).
+
 Plan: [`docs/llm-reference-apps-plan.md`](../../docs/llm-reference-apps-plan.md)
 
 ## Layout
diff --git a/apps/llm-mcp-reference/shared/prompts/system-boundary.md b/apps/llm-mcp-reference/shared/prompts/system-boundary.md
index 6c70641..a9c647a 100644
--- a/apps/llm-mcp-reference/shared/prompts/system-boundary.md
+++ b/apps/llm-mcp-reference/shared/prompts/system-boundary.md
@@ -6,9 +6,11 @@ You are a **façade** over the Traverse runtime. You do **not** invent business
 
 1. Use MCP / Traverse tools to **discover, validate, and execute** capabilities or workflows.
 2. When presenting results, only report fields returned by the runtime (for example title, tags, noteType, suggestedNextAction, status, action items). If a field is missing, say it is missing — **do not invent it**.
-3. Do not re-implement starter / doc-approval / meeting-notes pipelines in natural language “skills”.
-4. If the tool fails or the runtime is offline, report the tool error. Do not fabricate a successful structured result.
-5. Prefer deterministic tool execution over long chain-of-thought process control.
+3. When describing session progress, use Spec 001 presentation states from runtime evidence only: `idle | loading | loaded | blocked | ended | error`. Do not invent a success/`loaded` state without a completed tool/runtime result.
+4. For multi-capability runs (Spec 002), report invoke/result order from tool/runtime evidence — never from local step timers or guessed stages.
+5. Do not re-implement starter / doc-approval / meeting-notes pipelines in natural language “skills”.
+6. If the tool fails or the runtime is offline, report the tool error. Do not fabricate a successful structured result.
+7. Prefer deterministic tool execution over long chain-of-thought process control.
 
 ## Why
 
diff --git a/apps/llm-mcp-reference/shared/workflows/meeting-notes.md b/apps/llm-mcp-reference/shared/workflows/meeting-notes.md
index f8d9f67..7ff13b1 100644
--- a/apps/llm-mcp-reference/shared/workflows/meeting-notes.md
+++ b/apps/llm-mcp-reference/shared/workflows/meeting-notes.md
@@ -15,6 +15,8 @@ OS twins: [`apps/meeting-notes/`](../../../meeting-notes/) (Web / iOS / macOS /
 
 Discover this entrypoint through `traverse-mcp` — do **not** hard-code private crate paths. Prefer the **governed workflow/entrypoint** the server lists for `meeting-notes.process`.
 
+Presentation chrome (Specs 001/002): keep the session at `loading` while tools run; move to `loaded` only when runtime output fields are present; use `error` on tool/runtime failure. If the workflow exposes multiple capabilities, narrate invoke/result order from evidence only.
+
 ## Prerequisites (Mode A)
 
 1. Traverse checkout with working `traverse-mcp` (see [MCP stdio server](https://github.com/traverse-framework/Traverse/blob/main/docs/mcp-stdio-server.md)).
diff --git a/apps/llm-mcp-reference/shared/workflows/traverse-starter.md b/apps/llm-mcp-reference/shared/workflows/traverse-starter.md
index 62a4a4d..4151136 100644
--- a/apps/llm-mcp-reference/shared/workflows/traverse-starter.md
+++ b/apps/llm-mcp-reference/shared/workflows/traverse-starter.md
@@ -8,14 +8,14 @@ Use the Traverse MCP server’s discovery/description tools to locate the starte
 
 ## Happy path
 
-1. Confirm MCP server is reachable (`describe_server` / discovery).
-2. Execute with a note string input (example: `"Ship the MCP façade docs tomorrow"`).
-3. Render **only** runtime fields from the execution report (e.g. title, tags, noteType, suggestedNextAction, status / validation issues / summary when present).
+1. Confirm MCP server is reachable (`describe_server` / discovery) — presentation `idle` until submit.
+2. Execute with a note string input (example: `"Ship the MCP façade docs tomorrow"`) — while tools run, presentation is `loading` (or `blocked` only if the runtime reports a waiting-for-human condition).
+3. Render **only** runtime fields from the execution report (e.g. title, tags, noteType, suggestedNextAction, status / validation issues / summary when present) — presentation `loaded` when renderable output is present, else `ended` if the runtime completed without product fields.
 
 ## Success looks like
 
-Same as Web starter: structured fields filled by the runtime, not by the model.
+Same as Web starter: structured fields filled by the runtime, not by the model. Session chrome matches Spec 001 (`idle|loading|loaded|blocked|ended|error`).
 
 ## Failure
 
-If execution fails, show the tool error. Do not invent a title/tags payload.
+If execution fails, show the tool error and presentation `error`. Do not invent a title/tags payload.