diff --git a/clickup_runner/__main__.py b/clickup_runner/__main__.py
index 0e2780b..30495a5 100644
--- a/clickup_runner/__main__.py
+++ b/clickup_runner/__main__.py
@@ -9,8 +9,10 @@ from __future__ import annotations
 import logging
 import signal
 import sys
+import tempfile
 import time
 from datetime import datetime, timezone
+from pathlib import Path
 
 from .autocora import archive_result, scan_results, submit_job
 from .claude_runner import (
@@ -436,6 +438,37 @@ def _dispatch_autocora(
     )
 
 
+def _download_attachments(
+    client: ClickUpClient,
+    task: ClickUpTask,
+    work_dir: Path,
+) -> list[str]:
+    """Download all task attachments to the work directory.
+
+    Returns list of local filenames (not full paths) that were downloaded.
+    """
+    downloaded = []
+    for att in task.attachments:
+        title = att.get("title", "")
+        url = att.get("url", "")
+        if not title or not url:
+            continue
+
+        dest = work_dir / title
+        if client.download_attachment(url, dest):
+            downloaded.append(title)
+        else:
+            log.warning("Skipping attachment %s -- download failed", title)
+
+    if downloaded:
+        log.info(
+            "Downloaded %d attachment(s) to %s: %s",
+            len(downloaded), work_dir, ", ".join(downloaded),
+        )
+
+    return downloaded
+
+
 def _dispatch_claude(
     client: ClickUpClient,
     cfg: Config,
@@ -460,20 +493,20 @@ def _dispatch_claude(
         )
         return
 
-    # 3. Gather .xlsx attachment URLs for the prompt
-    xlsx_urls = [
-        a.get("url", "")
-        for a in task.attachments
-        if a.get("title", "").lower().endswith(".xlsx")
-        or a.get("url", "").lower().endswith(".xlsx")
-    ]
+    # 3. Create work dir and download all attachments into it
+    work_dir = Path(tempfile.mkdtemp(prefix="clickup_runner_"))
+    downloaded_files = _download_attachments(client, task, work_dir)
 
-    # 4. Build prompt
-    prompt = build_prompt(task, route, skill_content, xlsx_urls or None)
+    # 4. Build prompt (reference local filenames, not URLs)
+    prompt = build_prompt(task, route, skill_content, downloaded_files)
 
     # 5. Run Claude
     log.info("Starting Claude for task %s (%s)", task.id, task.name)
-    result = run_claude(prompt, route, cfg)
+    result = run_claude(
+        prompt, route, cfg,
+        work_dir=work_dir,
+        exclude_files=set(downloaded_files),
+    )
 
     if not result.success:
         _handle_dispatch_error(
diff --git a/clickup_runner/claude_runner.py b/clickup_runner/claude_runner.py
index 8aa5bf3..8fef05c 100644
--- a/clickup_runner/claude_runner.py
+++ b/clickup_runner/claude_runner.py
@@ -36,13 +36,17 @@ def build_prompt(
     task: ClickUpTask,
     route: SkillRoute,
     skill_content: str,
-    xlsx_urls: list[str] | None = None,
+    attachment_filenames: list[str] | None = None,
 ) -> str:
     """Assemble the prompt sent to `claude -p`.
 
     Structure:
         1. Skill file content (system-level instructions)
         2. Task context block (name, description, customer, URL, attachments)
+
+    Args:
+        attachment_filenames: Local filenames (not paths) of attachments
+            already downloaded to the working directory.
     """
     parts: list[str] = []
 
@@ -74,11 +78,11 @@ def build_prompt(
         ctx_lines.append("### Description")
         ctx_lines.append(task.description.strip())
 
-    if xlsx_urls:
+    if attachment_filenames:
         ctx_lines.append("")
-        ctx_lines.append("### Attached Cora Reports (.xlsx)")
-        for url in xlsx_urls:
-            ctx_lines.append("- %s" % url)
+        ctx_lines.append("### Attached Files (in working directory)")
+        for fname in attachment_filenames:
+            ctx_lines.append("- %s" % fname)
 
     # Tell Claude where to write output
     ctx_lines.append("")
@@ -94,11 +98,20 @@ def build_prompt(
     return "\n\n".join(parts)
 
 
-def _collect_output_files(work_dir: Path) -> list[Path]:
-    """Return all files Claude created in the working directory."""
+def _collect_output_files(
+    work_dir: Path,
+    exclude: set[str] | None = None,
+) -> list[Path]:
+    """Return all files Claude created in the working directory.
+
+    Args:
+        exclude: Set of filenames to skip (e.g. downloaded attachments
+                 that were in the dir before Claude ran).
+    """
     if not work_dir.exists():
         return []
-    files = [f for f in work_dir.iterdir() if f.is_file()]
+    exclude = exclude or set()
+    files = [f for f in work_dir.iterdir() if f.is_file() and f.name not in exclude]
     # Sort for deterministic ordering
     files.sort(key=lambda p: p.name)
     return files
@@ -109,6 +122,7 @@ def run_claude(
     route: SkillRoute,
     cfg: Config,
     work_dir: Path | None = None,
+    exclude_files: set[str] | None = None,
 ) -> RunResult:
     """Run `claude -p` as a subprocess and return the result.
 
@@ -118,6 +132,8 @@ def run_claude(
         cfg: Runner config (timeout, etc.).
         work_dir: Directory for Claude to write files into.
                   If None, a temp directory is created.
+        exclude_files: Filenames to exclude from output_files
+                       (e.g. pre-existing attachments).
     """
     created_tmp = False
     if work_dir is None:
@@ -169,11 +185,11 @@ def run_claude(
                 output=stdout,
                 error="Claude exited with code %d: %s"
                 % (result.returncode, stderr[:2000]),
-                output_files=_collect_output_files(work_dir),
+                output_files=_collect_output_files(work_dir, exclude_files),
                 work_dir=work_dir,
             )
 
-        output_files = _collect_output_files(work_dir)
+        output_files = _collect_output_files(work_dir, exclude_files)
         log.info(
             "Claude completed successfully. %d output file(s).", len(output_files)
         )
@@ -195,7 +211,7 @@ def run_claude(
             output="",
             error="Claude timed out after %d seconds"
             % cfg.runner.claude_timeout_seconds,
-            output_files=_collect_output_files(work_dir),
+            output_files=_collect_output_files(work_dir, exclude_files),
             work_dir=work_dir,
         )
     except FileNotFoundError:
diff --git a/clickup_runner/clickup_client.py b/clickup_runner/clickup_client.py
index 246de4a..15381f0 100644
--- a/clickup_runner/clickup_client.py
+++ b/clickup_runner/clickup_client.py
@@ -272,6 +272,26 @@ class ClickUpClient:
                 )
             return []
 
+    def download_attachment(self, url: str, dest: Path) -> bool:
+        """Download a ClickUp attachment to a local file.
+
+        ClickUp attachment URLs are pre-signed S3 URLs that don't need
+        auth headers, so we use a plain httpx request (not the API client).
+
+        Returns True on success, False on failure.
+        """
+        try:
+            with httpx.stream("GET", url, follow_redirects=True, timeout=60.0) as resp:
+                resp.raise_for_status()
+                with open(dest, "wb") as f:
+                    for chunk in resp.iter_bytes(chunk_size=8192):
+                        f.write(chunk)
+            log.info("Downloaded attachment to %s", dest)
+            return True
+        except Exception as e:
+            log.warning("Failed to download attachment from %s: %s", url, e)
+            return False
+
     def get_custom_fields(self, list_id: str) -> list[dict]:
         """Get custom field definitions for a list."""
         try:
diff --git a/skills/content_draft.md b/skills/content_draft.md
index 1b0884e..e4a3830 100644
--- a/skills/content_draft.md
+++ b/skills/content_draft.md
@@ -82,34 +82,14 @@ After the main content, write the FOQ section from the outline:
 - Mark the section: `<!-- FOQ SECTION START -->` and `<!-- FOQ SECTION END -->`
 - FOQs are excluded from the main word count
 
-## Step 4: Verify Against Cora
+## Step 4: Self-Check
 
-Run the optimizer scripts against your draft to check coverage.
+Before writing the output file, review your own draft:
 
-### 4a. Entity Check
-
-```bash
-cd .claude/skills/content-researcher/scripts && uv run --with openpyxl python entity_optimizer.py "{draft_path}" "{cora_xlsx_path}" --format json --top-n 30
-```
-
-Review the output:
-- Are all "Must mention" entities present? If any are missing, add them.
-- Are any entities at 0 that should have at least 1 mention? Fix.
-- Do NOT chase high deficit numbers for outlier entities.
-
-### 4b. LSI Check
-
-```bash
-cd .claude/skills/content-researcher/scripts && uv run --with openpyxl python lsi_optimizer.py "{draft_path}" "{cora_xlsx_path}" --format json --top-n 30
-```
-
-Review the output:
-- Are the top 20 LSI terms from the outline present? If any key ones are missing, weave them in.
-- Do NOT force LSI terms where they don't fit.
-
-### 4c. Word Count Check
-
-Count the words in your draft (excluding FOQ section). Compare against the Cora target. If you're over by more than 10%, trim. If under by more than 10%, expand the thinnest sections.
+- Count words per section (excluding FOQ). Compare against the outline's per-section targets. If any section is over by more than 10%, trim. If under by more than 10%, expand.
+- Scan for each entity in the "Must mention" list from the Writer's Reference. If any are missing, add them.
+- Scan for the top LSI terms. If key ones are absent, weave them in where natural.
+- Re-read for keyword stuffing -- if any paragraph sounds forced, rewrite it.
 
 ## Step 5: Generate Meta Tags