diff --git a/skills/distill-knowledge/SKILL.md b/skills/distill-knowledge/SKILL.md
index e86c592..a120767 100755
--- a/skills/distill-knowledge/SKILL.md
+++ b/skills/distill-knowledge/SKILL.md
@@ -11,14 +11,14 @@ compatibility: Requires uv, ffmpeg, ffprobe, Python ≥3.10, and OPENAI_API_KEY
 license: MIT
 metadata:
   author: Dim Kharitonov <dimds@fastmail.com> (https://github.com/dimdasci)
-  version: "1.1.0"
+  version: "1.2.0"
 ---
 
 # Convert Recording → Knowledge Markdown
 
 Emit `outbox/{meeting-slug}/transcript.md`; screenshots inline when useful; structured docs only on request. `{meeting-slug}` = `kebab-case-topic-YYYYMMDD`. Never touch `inbox/` or `knowledge/`.
 
-References (load on demand): [setup](references/setup.md) · [output templates](references/output-templates.md) · [ffmpeg](references/ffmpeg.md) · [transcribe CLI](references/transcribe-cli.md) · [prep audio CLI](references/prep-audio-cli.md) · [chunked transcription](references/chunked-transcription.md) · [structured docs](references/structured-docs.md) · spot-check: [`scripts/extract_clip.py`](scripts/extract_clip.py).
+References (load on demand): [setup](references/setup.md) · [output templates](references/output-templates.md) · [ffmpeg](references/ffmpeg.md) · [transcribe CLI](references/transcribe-cli.md) · [prep audio CLI](references/prep-audio-cli.md) · [chunked transcription](references/chunked-transcription.md) · [structured docs](references/structured-docs.md) · [cleanup CLI](scripts/cleanup.py) · spot-check: [`scripts/extract_clip.py`](scripts/extract_clip.py).
 
 Scripts run via `uv run --script` (PEP 723). All support `--help`. On first run verify [setup prerequisites](references/setup.md).
 
@@ -85,9 +85,19 @@ Technical / multilingual / mumbled audio → VTT-aligned path strongly preferred
 
 ### Steps 7–10 — Structured docs (conditional)
 
-**Gate 2** — ask: structured docs or transcript only? If transcript only → report + stop.
+**Gate 2** — ask: structured docs or transcript only? If transcript only → cleanup + report + stop.
 
-Otherwise → [structured docs reference](references/structured-docs.md): plan topics, emit `summary.md` + `topics/{slug}.md`, report, stop.
+Otherwise → [structured docs reference](references/structured-docs.md): plan topics, emit `summary.md` + `topics/{slug}.md`, cleanup, report, stop.
+
+### Step 11 — Cleanup (mandatory, after user accepts)
+
+Once the user confirms the results are acceptable, remove temporary prep artifacts via [`cleanup.py`](scripts/cleanup.py):
+
+```bash
+uv run --script scripts/cleanup.py --slug {meeting-slug}
+```
+
+The script verifies `outbox/{meeting-slug}/transcript.md` exists before deleting `tmp/prep/{meeting-slug}/`. Use `--dry-run` to preview. **Never use `rm` directly — all temp removal goes through this script.**
 
 ## Fidelity rule
 
@@ -107,3 +117,4 @@ When two transcripts agree on meaning → faithful. When they disagree and neith
 - **Don't load rich `--prompt`** on non-diarize model. Prompts leak as fabrications. Vocab list + 1-line topic max. See [prompt-hallucination warning](references/transcribe-cli.md#prompt-hallucination-warning).
 - **Don't auto-pick VTT decisions.** Always surface assessment; user confirms at Gate 1.
 - **Don't write outside `outbox/{meeting-slug}/`.** Temp files → `tmp/`; finals → `outbox/`.
+- **Don't run `rm` on temp files.** Always use [`cleanup.py`](scripts/cleanup.py) — it validates output exists and only operates under `tmp/`.
diff --git a/skills/distill-knowledge/scripts/cleanup.py b/skills/distill-knowledge/scripts/cleanup.py
new file mode 100755
index 0000000..a1a647a
--- /dev/null
+++ b/skills/distill-knowledge/scripts/cleanup.py
@@ -0,0 +1,138 @@
+#!/usr/bin/env -S uv run --script
+# /// script
+# requires-python = ">=3.10"
+# dependencies = []
+# ///
+"""Remove temporary prep artifacts after the user accepts the transcript.
+
+Safety constraints:
+- Only deletes directories under `<project-root>/tmp/`.
+- Refuses to act if the final transcript is missing in outbox.
+- Never shells out to `rm`; uses Python's shutil.rmtree.
+
+Run with:
+    uv run --script skills/distill-knowledge/scripts/cleanup.py \
+        --slug <meeting-slug> [--project-root .]
+"""
+
+from __future__ import annotations
+
+import argparse
+import shutil
+import sys
+from pathlib import Path
+from typing import NoReturn
+
+
+def _die(message: str, code: int = 1) -> NoReturn:
+    print(f"Error: {message}", file=sys.stderr)
+    raise SystemExit(code)
+
+
+def _resolve_project_root(raw: str) -> Path:
+    """Resolve and validate the project root directory."""
+    root = Path(raw).resolve()
+    if not root.is_dir():
+        _die(f"Project root is not a directory: {root}")
+    return root
+
+
+def _assert_under_tmp(path: Path, tmp_root: Path) -> None:
+    """Hard guard: path must be strictly inside tmp_root."""
+    try:
+        path.resolve().relative_to(tmp_root.resolve())
+    except ValueError:
+        _die(
+            f"Refusing to delete {path}: it is not under {tmp_root}. "
+            "This script only removes directories inside the tmp/ folder."
+        )
+
+
+def _dir_size(path: Path) -> int:
+    """Total bytes of all files in a directory tree."""
+    return sum(f.stat().st_size for f in path.rglob("*") if f.is_file())
+
+
+def _human_size(nbytes: int) -> str:
+    """Format byte count as a human-readable string."""
+    for unit in ("B", "KB", "MB", "GB"):
+        if nbytes < 1024:
+            return f"{nbytes:.1f} {unit}"
+        nbytes /= 1024  # type: ignore[assignment]
+    return f"{nbytes:.1f} TB"
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        description="Remove temporary prep artifacts for a completed meeting."
+    )
+    parser.add_argument(
+        "--slug",
+        required=True,
+        help="Meeting slug (e.g. architecture-dima-mark-20260211)",
+    )
+    parser.add_argument(
+        "--project-root",
+        default=".",
+        help="Project root directory (default: current directory)",
+    )
+    parser.add_argument(
+        "--dry-run",
+        action="store_true",
+        help="Show what would be deleted without actually deleting",
+    )
+
+    args = parser.parse_args()
+    slug: str = args.slug
+    root = _resolve_project_root(args.project_root)
+    tmp_root = root / "tmp"
+    prep_dir = tmp_root / "prep" / slug
+    outbox_dir = root / "outbox" / slug
+    transcript = outbox_dir / "transcript.md"
+
+    # --- Pre-flight checks ---
+
+    if not tmp_root.is_dir():
+        _die(f"tmp/ directory does not exist: {tmp_root}")
+
+    if not prep_dir.is_dir():
+        _die(f"Nothing to clean: {prep_dir} does not exist")
+
+    _assert_under_tmp(prep_dir, tmp_root)
+
+    if not transcript.is_file():
+        _die(
+            f"Final transcript not found at {transcript}. "
+            "Refusing to delete temp artifacts before the output is confirmed. "
+            "Finish the transcript first, then re-run cleanup."
+        )
+
+    # --- Inventory ---
+
+    files = list(prep_dir.rglob("*"))
+    file_count = sum(1 for f in files if f.is_file())
+    dir_count = sum(1 for f in files if f.is_dir())
+    total_bytes = _dir_size(prep_dir)
+
+    print(f"Slug:        {slug}", file=sys.stderr)
+    print(f"Prep dir:    {prep_dir}", file=sys.stderr)
+    print(f"Transcript:  {transcript} ✓", file=sys.stderr)
+    print(f"Files:       {file_count}", file=sys.stderr)
+    print(f"Directories: {dir_count}", file=sys.stderr)
+    print(f"Size:        {_human_size(total_bytes)}", file=sys.stderr)
+
+    if args.dry_run:
+        print(
+            "\n[dry-run] Would delete the above. Pass without --dry-run to execute.",
+            file=sys.stderr,
+        )
+        return
+
+    # --- Delete ---
+
+    shutil.rmtree(prep_dir)
+    print(f"\nDeleted {prep_dir} — freed {_human_size(total_bytes)}", file=sys.stderr)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/skills/distill-knowledge/scripts/tests/test_cleanup.py b/skills/distill-knowledge/scripts/tests/test_cleanup.py
new file mode 100644
index 0000000..c754dd4
--- /dev/null
+++ b/skills/distill-knowledge/scripts/tests/test_cleanup.py
@@ -0,0 +1,106 @@
+"""Tests for cleanup.py safety logic."""
+
+from __future__ import annotations
+
+import json
+import subprocess
+import sys
+from pathlib import Path
+
+SCRIPT = str(Path(__file__).resolve().parent.parent / "cleanup.py")
+
+
+def _run(
+    tmp_path: Path, slug: str, *, dry_run: bool = False, extra_args: list[str] | None = None
+) -> subprocess.CompletedProcess[str]:
+    cmd = [
+        sys.executable,
+        SCRIPT,
+        "--slug",
+        slug,
+        "--project-root",
+        str(tmp_path),
+    ]
+    if dry_run:
+        cmd.append("--dry-run")
+    if extra_args:
+        cmd.extend(extra_args)
+    return subprocess.run(cmd, capture_output=True, text=True)
+
+
+def _setup_project(tmp_path: Path, slug: str, *, with_transcript: bool = True) -> None:
+    """Create minimal project structure with tmp/prep and optionally outbox."""
+    prep = tmp_path / "tmp" / "prep" / slug
+    prep.mkdir(parents=True)
+    # Create some temp files
+    (prep / "stripped.ogg").write_bytes(b"\x00" * 1024)
+    (prep / "manifest.json").write_text(json.dumps({"version": 1}))
+    chunks = prep / "chunks"
+    chunks.mkdir()
+    (chunks / "chunk_00.ogg").write_bytes(b"\x00" * 512)
+
+    if with_transcript:
+        outbox = tmp_path / "outbox" / slug
+        outbox.mkdir(parents=True)
+        (outbox / "transcript.md").write_text("# Test transcript\n")
+
+
+def test_refuses_without_transcript(tmp_path: Path) -> None:
+    """Cleanup must refuse if outbox transcript is missing."""
+    slug = "no-transcript-slug"
+    _setup_project(tmp_path, slug, with_transcript=False)
+
+    result = _run(tmp_path, slug)
+    assert result.returncode != 0
+    assert "Final transcript not found" in result.stderr
+    # Prep dir must still exist
+    assert (tmp_path / "tmp" / "prep" / slug).is_dir()
+
+
+def test_refuses_nonexistent_slug(tmp_path: Path) -> None:
+    """Cleanup must fail if prep dir doesn't exist."""
+    (tmp_path / "tmp").mkdir(parents=True)
+    result = _run(tmp_path, "ghost-slug")
+    assert result.returncode != 0
+    assert "does not exist" in result.stderr
+
+
+def test_dry_run_preserves_files(tmp_path: Path) -> None:
+    """Dry run should report but not delete."""
+    slug = "dry-run-test"
+    _setup_project(tmp_path, slug)
+
+    result = _run(tmp_path, slug, dry_run=True)
+    assert result.returncode == 0
+    assert "dry-run" in result.stderr
+    # Nothing deleted
+    assert (tmp_path / "tmp" / "prep" / slug / "stripped.ogg").exists()
+
+
+def test_actual_cleanup(tmp_path: Path) -> None:
+    """Actual cleanup should remove prep dir when transcript exists."""
+    slug = "real-cleanup"
+    _setup_project(tmp_path, slug)
+
+    prep = tmp_path / "tmp" / "prep" / slug
+    assert prep.is_dir()
+
+    result = _run(tmp_path, slug)
+    assert result.returncode == 0
+    assert "Deleted" in result.stderr
+    assert "freed" in result.stderr
+    # Prep dir gone
+    assert not prep.exists()
+    # Transcript untouched
+    assert (tmp_path / "outbox" / slug / "transcript.md").is_file()
+
+
+def test_reports_file_count_and_size(tmp_path: Path) -> None:
+    """Dry run output should include file count and size."""
+    slug = "inventory-check"
+    _setup_project(tmp_path, slug)
+
+    result = _run(tmp_path, slug, dry_run=True)
+    assert "Files:" in result.stderr
+    assert "Size:" in result.stderr
+    assert "Directories:" in result.stderr