TL;DR
Comment replies are the cheapest, highest-frequency place your brand voice shows up — and the easiest place for it to drift once more than one person is replying. A one-page style guide fixes that. Feed an AI model 30-50 of your best past replies plus 5-10 you regret, and it will draft a usable guide in minutes: voice rules, length norms, per-comment-type examples, a retire list, and an escalation ladder. Then paste that guide into a reusable AI workspace (a Claude Project or a custom GPT) so every drafted reply already knows your voice. The human’s job is the part AI can’t do: reading community culture and signing off on the examples.
Why a reply style guide is worth one afternoon
A single warm, specific reply to a stranger does more for retention than a polished post, and a single tone-deaf reply can undo a month of careful work. Yet most accounts reply ad-hoc, with whoever is on shift, and the voice slides. Sprout Social’s own guidance is blunt about the fix: keep “a shared document or spreadsheet where you log approved microcopy for common questions” — a system, not vibes.
A style guide is that system in one page. It records exactly how this account talks: opening words, sentence length, when emoji is allowed, and what you never say. AI is well suited to drafting it because the raw material — your past replies — already exists. You are not asking the model to invent a voice. You are asking it to reverse-engineer the one you already have.
When AI is the right tool here
- You or a team reply to comments daily and the voice is visibly inconsistent.
- You have 30-50 past replies you consider on-brand and 5-10 that went wrong.
- You’re growing, so soon more than one person will be hitting “reply.”
Where a human still has to own it
AI cannot read your community’s actual culture. The community manager’s gut about what flies on your particular subreddit, Discord server, or Instagram comment section is irreplaceable, and it is also where the legal and reputational risk lives. Use AI to draft the framework; have a human sign off on every example.
AI also over-warms. Left unguided, GPT-5.5 and Claude Sonnet 4.6 both default to a peppy, exclamation-heavy register (“Love this!!”) that reads as inauthentic on most accounts. Bake an explicit “no more than one exclamation point per reply” rule into the guide and the prompt, or you will spend your edits stripping sugar.
One more caution worth writing down: do not wire a model to auto-post. Even the platforms that ship AI replies keep a human in the loop — Hootsuite’s Smart Replies suggests an on-brand response and a person still approves each one before it publishes. Treat any tool that promises fully autonomous public replies as a brand-safety incident waiting to happen.
What to feed the model
| Input | Why it matters |
|---|---|
| Account purpose, one sentence | Anchors tone to audience (a fintech reply ≠ a meme page reply) |
| Voice attributes, 3-5 words | e.g. “direct, dry, never condescending” — the model imitates these |
| 30-50 on-brand replies, verbatim | The single most important input; this is your real voice |
| 5-10 regretted replies + one line each | Negative examples teach faster than positive ones |
| Platforms you reply on | Norms differ: LinkedIn is “friendly but formal,” TikTok stays light |
Paste replies verbatim. Paraphrasing them launders out the exact word choices and rhythm that make your voice yours.
The prompt
Draft a comment-reply style guide for my account.
Account purpose: [one sentence]
Voice attributes: [3-5 words]
Platforms: [list]
Sample on-brand replies (verbatim): [paste 30-50]
Sample off-brand replies and why they were wrong: [paste 5-10 with reasons]
Output the guide, total length under one page:
1. Voice rules: 5-7 bullets, each one rule + one short example.
Include an emoji rule and a max-one-exclamation-point rule.
2. Length norm: short (1 sentence) vs medium (2-3) vs long (4+) — when each.
3. By comment type — Genuine question / Praise or fan /
Constructive criticism / Hostile or troll / Spam or off-topic.
For each: 2 on-brand example replies, then 2 do-not-do examples
with one line on why.
4. Five phrases to retire (overused or off-brand).
5. Five phrases or moves that are uniquely "us."
6. Escalation ladder: when to stop replying and pass to a human / moderator.
Do not invent praise the samples don't support. Match the register of the
verbatim samples, not a generic friendly-brand voice.Turn the guide into a reusable AI workspace
A document in a shared drive is step one. The leverage comes when the guide lives inside the tool that drafts replies, so every suggestion is already on-voice. As of June 2026 you have three good options:
- Claude Project — create a Project, paste the full guide into Project Instructions, and upload your 30-50 sample replies as a knowledge file. Project Instructions have no practical character limit, so the entire guide fits. Best when your guide is long or you want the samples retrievable.
- Custom GPT — in ChatGPT, build a GPT and put condensed rules in the instructions field. Note the cap: custom-instruction text is limited to roughly 1,500 characters, so you’ll trim the guide to its core rules and attach the examples as a knowledge file instead.
- A saved prompt — lowest effort: keep the guide as a text snippet and paste it above each batch of comments. Fine for a solo account; brittle once a team is involved.
Whichever you pick, the guide is the asset. The workspace just keeps it always-on.
Don’t skip platform-native moderation
A style guide governs what you say. Platform tools govern what you see. On Instagram, turn on the Hidden Words filter (Settings → Privacy → Comments) and add a Manual Filter keyword list so spam and slurs are auto-hidden before a human reads them. Two limits to know, both as of June 2026: the filter has no context awareness (“this product is fire” and “a dumpster fire” both trip on “fire”), and Hidden Words does not support languages without spaces between words, including Chinese, Japanese, and Thai. So filters reduce the volume your team replies to; they do not replace the guide.
How to pressure-test the draft
- Send it to one person who has been replying for months. Their pushback is the highest-value input you’ll get.
- Run the guide against last week’s actual comments. If a real comment doesn’t fit any bucket, the guide is incomplete — add the bucket.
- Read the retire list aloud. If it names phrases everyone on the team uses daily, you have an internal argument to win before the guide ships.
Common mistakes
- “Be kind and authentic” guidance. Every brand says this; none of it is operational.
- No anti-examples. The do-not-do replies are what stop the 8pm-Wednesday mistakes.
- Skipping the hostile/troll section. Voice slips hardest exactly where the guide goes quiet.
- Writing it once and freezing it. Revisit quarterly, or it describes an account you no longer are.
Keep the guide alive
Every month, pull the 10 most-liked replies from your account and the 3 worst. Add the winners as new examples; add the losers, with a one-line lesson, to the do-not list. Then re-paste the updated guide into your Claude Project or custom GPT so the drafting workspace learns the same lesson the team did. The guide should grow with the account, not date with it.
FAQ
- How long should the style guide be? One page, two at the absolute most. A team member won’t read more at 8pm on a Wednesday, which is exactly when they need it.
- Should I share the guide with my community? No. It’s internal. Published reply playbooks get gamed — trolls learn precisely which words trigger your escalation ladder.
- Do I need a separate guide per platform? Usually no. Keep one core guide plus platform-specific footnotes (LinkedIn formal, TikTok loose) rather than maintaining five drifting documents.
- Can I let the AI post replies automatically? Not for public comments. Keep a human approving each one, the way Hootsuite’s Smart Replies and Sprout’s AI Assist both do. Auto-posting is how a single hallucinated reply becomes a screenshot.
- Which model drafts the best guide? Either flagship works. Claude Opus 4.7 or Sonnet 4.6 tends to imitate a verbatim voice closely; GPT-5.5 is fine too. The quality of your 30-50 sample replies matters far more than the model choice.