/review-comments

name: webflow-mcp:review-comments
description: Review open comment threads on a Webflow site and triage each one.

Review open comment threads on a Webflow site and triage each one.

**Input:** `$ARGUMENTS` — a site name (e.g. "Workhaus"), site ID (e.g. `6808fd4eff835ee3af009d6f`), or either with the `-reply` flag (e.g. `Workhaus -reply`).

**Flags:**

• `-reply` — in addition to writing the report, post a bot reply to each non-open thread. Without this flag, the skill runs in report-only mode (read-only).

Parse `$ARGUMENTS` at the start: strip `-reply` from the input to get the site identifier, and set `replyMode = true` if `-reply` was present, `false` otherwise.

---

Step 1 — Resolve the site

If the site identifier is empty or blank — cross-site comment survey

1. Call `data_sites_tool > list_sites` to get all sites. Page through until all are collected. 2. For each site, call `data_comments_tool > list_comment_threads` with `isResolved: false` and `limit: 100`. Page through until all unresolved threads are collected. If the API does not support `isResolved` filtering, fetch all threads and filter client-side to `isResolved === false`. Process sites in batches of 20 (batch multiple actions in a single tool call). After each batch completes, log a progress line: `Batch {N}/{total} done (sites {start}–{end}): {summary of findings, e.g. "all 0 threads" or "SiteName has X threads, rest 0"}.` 3. For each site, compute:

• `unresolvedCount` = total unresolved threads
• `newestDate` = the maximum `lastUpdated` value across all unresolved threads for that site (ISO → human-readable date, e.g. "Apr 3, 2026"). If no threads, show `—`.
• `oldestDate` = the minimum `lastUpdated` value across all unresolved threads for that site (ISO → human-readable date). If no threads, show `—`.

4. Sort sites by `unresolvedCount` descending. Take the top 10. 5. Above the table, show a heading line: `### Checked {totalSiteCount} sites — {sitesWithUnresolved} have unresolved comments.` 6. Display a table in this format (link just the site name to `https://webflow.com/design/{siteId}`):

| Site | Unresolved Comments | Newest / Oldest |
|------|---------------------|-----------------|
| [Site Name](https://webflow.com/design/{siteId}) ({siteId}) | {N} | {newestDate} / {oldestDate} |

7. After the table, tell the user: "Run `/webflow-mcp:review-comments <site name or ID>` to review a specific site." 8. Write the survey output to a file:

• Ensure `comment-reviews/` exists (create with `mkdir comment-reviews` if not).
• Filename: `comment-reviews/triage-report-{YYYY-MM-DD-HH-MM}.md` using the current local time (zero-padded).
• File contents: a bold H1 title `# **Webflow Comment Review**`, then a blank line, then an H3 line `### Report created on: {human-readable date, time, and timezone, e.g. "April 16, 2026 at 10:39 AM PDT"}`, then a blank line, then the H3 heading line (`### Checked …`) and the full table from steps 5–6 above, in markdown.
• Log the path after writing, e.g. `Report written to comment-reviews/triage-report-2026-04-16-14-30.md`.

9. **Stop** — do not proceed to Step 2.

If the site identifier looks like a Webflow site ID (24-char hex), use it directly.

Otherwise call `data_sites_tool > list_sites` and find the site whose `displayName` matches the identifier (case-insensitive). If no match, tell the user and stop.

---

Step 2 — Fetch all open threads

**Always make a fresh API call here — never reuse thread data from earlier in the conversation. The user may have added or resolved comments since the last run.**

Call `data_comments_tool > list_comment_threads` with `isResolved: false` and `limit: 100`. Page through results until all threads are collected.

Log: `Site: {displayName}` and `Found {N} open thread(s)`.

---

Step 3 — Build page-level element map

From the already-fetched thread list, build a frequency map of `elementId.element → Set<pageId>` across all threads. Any `elementId.element` that appears on **2 or more distinct pages** is almost certainly the page root/body element (a real element ID would be page-scoped; only shared structural roots repeat across pages).

No API calls needed — this is a local computation on the thread data.

---

Step 4 — Triage each thread

For each thread:

4a — Fetch replies

Call `data_comments_tool > list_comment_replies` for this thread.

4b — Dedup check

Look for replies whose `content` includes the string `— 🤖 Comment Review Agent`.

If found, note the most recent one (`lastAgentReply`). If no human reply exists with a `createdOn` after `lastAgentReply.createdOn`, **skip this thread** (increment skipped count, continue to next thread).

4c — Compute element context

• `elementId` = `thread.elementId?.element`
• `isPageLevel` = `elementId` appears on 2 or more distinct `pageId`s in the frequency map from Step 3

4d — Classify the thread

Use the following criteria:

**noise** — No real design or engineering value:

• Test/placeholder text ("hello world", "testing", "asdf", random characters)
• Casual reactions with no ask ("looks nice", "nice!", "hey hey")
• Duplicate sentiments that add nothing

**stale** — Real concern, but old and likely handled:

• Substantive comments older than 14 days with no replies and no follow-up
• Questions that are probably resolved ("beta for how long?", "is this good contrast?")
• Action items that normal review would have caught

**open** — Real, actionable concern needing attention:

• Explicit tasks ("Should be sentence case", "fix image", "Look at name wrt L10N")
• Design decisions still required
• Specific and concrete concerns

**page-level** — Comment is on the page root, not a specific element:

• Use this when `isPageLevel` is true
• The comment is not anchored to any specific element — it may be intentional or may be an orphan from a deleted element

4e — Compose reply

Always compose the reply text (it appears in the report regardless