notion-anchored-comments

model	sonnet
name	notion-anchored-comments
description	Post block-anchored review comments on Notion pages via the Notion MCP. Use when posting RFC/doc review feedback to Notion, anchoring on code blocks or prose, replying to existing discussions, or hitting selection_with_ellipsis validation errors.

Notion Anchored Comments

Post review/feedback comments on Notion pages anchored to specific blocks, not the page as a whole. Use this for RFC reviews, design-doc feedback, or any structured commentary that needs to land next to the relevant content.

Hard Tool Constraints

The Notion MCP exposes:

• mcp__claude_ai_Notion__notion-create-comment: create
• mcp__claude_ai_Notion__notion-get-comments: read
• mcp__claude_ai_Notion__notion-fetch: fetch page (use include_discussions: true to see anchor positions)

There is no notion-update-comment and no notion-delete-comment. Comments are append-only via MCP. To fix a posted comment you must ask the user to delete it in the Notion UI, then repost. Never assume you can edit.

Notion comments do not render fenced code blocks (triple-backtick) when posted via rich_text. The only available code styling is the code: true annotation on individual rich_text segments. Plan body content accordingly (see Wrapper Format).

Targeting Modes

notion-create-comment has three targeting modes. Pick exactly one:

1. Page-level: page_id only. Avoid for review feedback. Page-level comments lose context and are usually a mistake; if anchoring fails, stop and report rather than degrade to page-level.
2. Block-anchored: page_id + selection_with_ellipsis. Anchors to a specific block. Whether the UI also captures a sub-selection within that block depends on the block type, not on anchor tightness (see Block-Type Sub-Selection Behaviour).
3. Discussion reply: page_id + discussion_id. Replies into an existing thread. Use this when explicitly continuing a conversation rather than starting a new finding.

selection_with_ellipsis Rules

Format: <start>...<end>. Notion finds the first occurrence of <start> and the nearest <end> after it. Both halves combined must match exactly once across the whole page.

Common failure modes (all return validation_error: Multiple occurrences found: N occurrences):

• Generic ends like ?? null, : null, }), }: match many times in code samples.
• Identifiers that appear in multiple sections (e.g. a variable named in prose, in code, and in a checklist).
• End anchor that occurs both inside the intended span and again later (e.g. cioanalytics.page() inside an if block and again inside a callback).

Recovery recipe:

1. Read the page once with notion-fetch (include_discussions: true) so you have the full content in context.
2. Pick a <start> that names a specific declaration or unique phrase (const currentPosIntegrationType, Blocking prerequisite, Flip ENABLE_CUSTOMER_IO_ANALYTICS).
3. Pick an <end> that occurs exactly once between the intended start and the next ambiguous boundary. Closing punctuation ():, )., :) and uniquely-spelled identifiers help.
4. The selection must fit within a single block. You cannot anchor across blocks; pick the most representative one.
5. Snippet length only affects uniqueness (the start+end combo must match the page exactly once). It does not drive sub-selection on paragraph or bullet blocks (see Block-Type Sub-Selection Behaviour).
6. Uniqueness matching is strict and operates on substring level, not word boundaries. A start anchor like initi will match every occurrence of that substring on the page (initiative, initialise, etc.), so even a single-word visual target often needs surrounding context to combine into a unique start+end pair. Verified 2026-05-06: initi...ative failed with Multiple occurrences found: 3 occurrences because initi appeared in initiative twice plus a third matching token elsewhere.

If after two attempts the selection still has multiple matches, switch the anchor to a different unique line in the same logical area rather than fighting the matcher.

Block-Type Sub-Selection Behaviour

selection_with_ellipsis always anchors at block level. Whether the UI also captures a sub-selection inside the block (visible in the API as a text-context attribute on the discussion, and in the UI as an in-block underline of the targeted phrase) depends on the block type, not on how tight the anchor is.

Sub-selection captured (in-block underline):

• Table cells.
• Code blocks.
• Reference data point: the Fox PaaS BullMQ thread on 3563c6719946813e9fbfedf7efe77685 produced text-context="BullMQ" because the anchor was a single word inside a table cell.

Sub-selection NOT captured (whole block highlights):

• Paragraph blocks.
• Bulleted list items.
• Heading blocks (treat as no-sub-selection until proven otherwise; not yet directly tested).
• Test data, 2026-05-06 against Notion page 3503c6719946806d9bfcffaafead7219: three variants on the same paragraph block (long anchor with ~50 char start and ~22 char end, tight ~10-char-per-side anchor, and a tightest-feasible word-scoped anchor) all returned context="block" with no text-context. The Notion UI highlighted the whole paragraph in every case.

Implication: tightening the anchor on a paragraph or bullet block will not produce a tighter UI highlight; anchor tightness there only matters for uniqueness validation. Set the user's expectations up front: on prose blocks, a comment underlines the whole block in the UI regardless of how surgical the anchor looks. The comment still lands correctly; readers just see a wider visual scope.

Verifying which behaviour you got: in notion-get-comments output, look at the <discussion> element. A text-context="<phrase>" attribute means Notion captured a sub-selection. Its absence means block-level highlight only. For paragraph and bullet blocks that is normal, not a failure; do not retry to force it. The create call returns success either way; the difference only shows up here.

Block-Discussion Merging Behavior

If a block already has any discussion attached, sub-selections within that block merge into the existing discussion rather than spawning a new sub-anchored thread. Multiple findings on the same code block will end up as sequential comments in one thread.

Implications:

• Tell the user up-front when several findings will share a thread (e.g. multiple comments on a single code block).
• Use discussion_id only when intentionally replying to a specific existing thread.
• The first post against a fresh block creates a new discussion; later posts on the same block append.

Wrapper Format

Use this body shape for review findings (matches the reviewing-branch-changes skill's LLM note: wrapper, adapted for Notion):

LLM note: <one-line summary, lowercase verb after the colon>

<paragraph: clear issue statement and concrete impact, with inline-code for files/identifiers>

Recommended change: <prose for the fix>

<optional multi-line code snippet, single rich_text segment with code: true>

(ref: <FINDING-ID>)

Rules:

• LLM note: prefix on the first line. Always.
• (ref: <ID>) suffix on the last line for findings with stable IDs from a checklist (e.g. CIO-PAGE-1). Skip the suffix for replies that are not findings.
• File paths, identifiers, function names, and code fragments: separate rich_text segments with annotations: { code: true }.
• Multi-line snippets: one rich_text segment with embedded \n and code: true. This renders as inline code with line breaks, the closest Notion comments get to a fenced block.
• Plain prose between code segments lives in unannotated rich_text segments.

Build the body as an array of rich_text objects, alternating plain and code: true segments. Example skeleton:

[
  {"text": {"content": "LLM note: <summary>\n\nThe code sets "}},
  {"text": {"content": "everIdentified.current = true"}, "annotations": {"code": true}},
  {"text": {"content": " in "}},
  {"text": {"content": "components/Tracking/CustomerIoTracking.tsx"}, "annotations": {"code": true}},
  {"text": {"content": ".\n\nRecommended change: ...\n\n"}},
  {"text": {"content": "const wasIdentified = ...\nif (...) {\n  ...\n}"}, "annotations": {"code": true}},
  {"text": {"content": "\n\n(ref: CIO-PAGE-1)"}}
]

Posting Workflow

1. Fetch and orient. Call notion-fetch with include_discussions: true. Note existing discussions, their text-context, and the IDs of any threads you might reply to.
2. Set expectations. Before posting, tell the user which findings will land on prose blocks (whole-block highlighting) and which will share a thread through block-discussion merging.
3. Test one anchor first. Post the most-likely-tricky finding as a single test, then notion-get-comments (include_all_blocks: true) and confirm context="block" (not page-level) and that the discussion landed on the intended block. Presence or absence of text-context follows block type (see Block-Type Sub-Selection Behaviour); either is normal.
4. Post the rest. Parallel calls are fine. Each anchored comment that lands on a fresh block creates its own discussion; ones that share a block get bundled.
5. Verify. Final notion-get-comments (include_all_blocks: true). Tally discussions and per-thread comment counts. Report the breakdown to the user.
6. Stop on failure. If anchoring repeatedly fails (multiple-occurrence errors with no good unique anchor), stop and report the limitation. Do not fall back to page-level.

Verification Output

Default report shape after a batch post:

Final state: N discussions, M comments

| Block / Discussion | Comments |
|---|---|
| <text-context excerpt> | <count>: <ID-1>, <ID-2>, ... |
...

Confirm context="block" (never page-level) and that each discussion landed on the intended block. Treat text-context as informational, not pass/fail. Call out any thread-merging in the report.

Quick Reference: Common Anchor Picks

• Code declarations: const <varName> is usually unique within a single code block.
• Section headings: full heading text often unique (This PR fixes...identity model: works, Production rollout alone does not).
• Bullets: lead with a distinctive opening phrase (Blocking prerequisite..., Flip ENABLE_<FLAG>...).
• Quoted strings: API names, env vars, file paths in backticks make excellent anchors when present.

When to Use This Skill

• User asks to "post review comments to a Notion page / RFC / design doc".
• User shares a Notion URL plus a draft of review findings.
• Recovering from a previous session where comments landed page-level instead of anchored.
• After running reviewing-branch-changes against a doc-style target and the user wants the findings posted to Notion rather than GitHub.

When Not to Use

• One-off page-level comment that genuinely targets the whole page (rare for reviews).
• Editing or deleting an existing comment: not possible via MCP. Tell the user to delete in the UI; repost from a corrected draft.