engagement-replies

---
name: engagement-replies
description: Writes one reply to every comment on your post in your own voice, ordered by which replies actually grow the account. Use once the comments arrive.
---

# Reply to every comment in your own voice

Paste the comments, get a paste-ready reply for each, in your words.

## Claude does

1. Ask for four things, guess none: the post in full; the comments with each commenter's name and headline; the platform; and 5 to 10 of the user's own past replies as a voice sample. No sample, ask once and stop. An invented voice is what makes replies read as bot output.
2. Build a voice card first: sentence length, contractions, greetings, sign-offs, emoji count, capitalisation. Both leading skill repos load a persistent voice file before any draft; this does it in session.
3. Sort every comment: question, added insight, disagreement, agreement, spam. Spam gets nothing.
4. Rank by growth value, not arrival order: questions, then added insight (longest threads), then disagreement, then praise. Buffer's reply study, roughly 2 million posts and 220,000 accounts, measured +30% engagement on LinkedIn and +42% on Threads for accounts that reply. Target is all of them, inside 60 minutes.
5. Length by bucket. Praise: 1 sentence naming something specific they said. Question: 2 to 3 sentences with a real answer, never "great question". Added insight: 2 sentences building on their point, ending with a question back. Disagreement: 2 to 3 sentences conceding the true part first, no link, no pitch.
6. Write the pinned first comment, 1 to 2 lines, carrying any link, because links in the body suppress reach. Charlie Hills' set ships a skill for exactly this.
7. Flag what Claude will not answer: legal claims, pricing disputes, a named client, anyone visibly upset. These return as a note, never a draft.

## Then Codex does

Codex takes the bulk formatting: build the per-comment file, count characters against the platform ceiling, strip hashtags, catch repeated openings, order the list. Mechanical, because every judgement call is already fixed in the drafts.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

Claude fills these into CODEX.md first: `{VOICE_CARD}` from step 2, `{COMMENTS_AND_DRAFTS}` (comment, bucket, draft), `{PLATFORM}`, `{ACCOUNT_NAME}` (whose name they go out under).

## Claude checks

1. Read every reply against the voice card. A word the sample never uses gets rewritten, not shipped.
2. Check openings. Two replies opening the same way is the clearest tell of automated replying: reject and rerun.
3. Confirm no reply invents a number, date, client name or quote absent from the post or comment. Popular caption repos circulate precise statistics with no source; this is the mistake to catch.
4. Confirm no hashtags. Metricool's 2026 Instagram study, 24,364,803 posts, measured hashtags at -31.70% views.
5. Confirm the step 7 flags are still flagged and undrafted.
6. Any failure: rerun Codex naming it. Never hand-patch and call it passed.

## Rules

- Reply to every comment that is not spam, inside the hour.
- Never open with "Great question" or "Thanks for sharing".
- Never answer a disagreement with a link.
- Never score on likes. Metricool's 2026 LinkedIn study, 673,658 posts, found likes down 13% while engagement rose 14%.
- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.

## If Codex is not installed

Claude does the formatting and checking itself, and says so in one line at the top.

## Built from

- charlie947/social-media-skills, https://github.com/charlie947/social-media-skills, 3,283 stars from api.github.com; ships pinned-comment and the voice-file-first pattern.
- Buffer, State of Social Media Engagement 2026, https://buffer.com/resources/state-of-social-media-engagement-2026/, roughly 2M posts, 220,000+ accounts; the reply uplift figures.
- Metricool 2026 Instagram Study, https://metricool.com/press-release-instagram-study-2026/, 24,364,803 posts; hashtag penalty.
- Metricool 2026 LinkedIn Study, https://metricool.com/press-release-linkedin-study-2026/, 673,658 posts; likes down 13%, engagement up 14%.

# Reply to every comment in your own voice - Codex task

You are formatting and checking replies that are already written. Do not rewrite voice or change what a reply says. Your job is counting, deduplicating and laying out.

## You are given

- `{ACCOUNT_NAME}` - the person whose name these replies go out under. State it in every file.
- `{PLATFORM}` - LinkedIn, Instagram, Threads, TikTok or X. Sets the character ceiling.
- `{VOICE_CARD}` - sentence length, contractions yes or no, emoji count per reply, greetings and sign-offs, capitalisation habit.
- `{COMMENTS_AND_DRAFTS}` - a list. Each entry: commenter name, headline if known, comment text, bucket (question / added-insight / disagreement / agreement), draft reply.

## Produce

Three files in the working folder.

1. `replies.md` - ordered: questions, then added-insight, then disagreement, then agreement. Per entry, in order: `### <n>. <commenter name>`, the comment as a blockquote, the reply in a fenced block for clean copying, then `chars: <n> | bucket: <bucket> | opening: <first three words>`.
2. `pinned-comment.md` - the pinned first comment, 1 to 2 lines, plus its character count. If none was supplied, write `NONE SUPPLIED` and nothing else.
3. `checks.md` - a table, one row per check: duplicate opening phrases, hashtags found, replies over the ceiling, emoji counts outside the voice card range, replies containing a digit or date, replies containing a URL. Each row PASS or FAIL with the offending reply numbers.

Ceilings: LinkedIn 750, Instagram 500, Threads 500, TikTok 150, X 280.

## Rules

- Never change the meaning of a draft. A reply over the ceiling is flagged FAIL with its count, text left intact. Do not cut it yourself.
- Strip every hashtag and record how many were removed.
- No two replies may open with the same first three words. Report duplicates, do not fix them.
- Flag but do not remove any reply containing a digit, date, currency symbol or URL. A human confirms the fact is real.
- Do not add greetings, sign-offs, emoji or names not already in the draft.
- Public information only. Never invent a fact, a number or a quote. Nothing is sent anywhere; this is text for a person to send under `{ACCOUNT_NAME}`.
- British English throughout.

## Return

Print to stdout, nothing else:

1. The three file paths written.
2. `TOTAL REPLIES: <n>` and the count per bucket.
3. The full contents of `checks.md`.
4. A final line: `ALL CHECKS PASS` or `FAILED: <check names>`.