An approve-or-reject checklist for AI-drafted replies asks six questions before a reply goes out: does it answer what the post actually says, did it pick one angle, does it add something or only agree, is every fact in it also in the post, is it one to three short lines, and would you delete it as generic praise on your own post.

Playbook · November 2026

Approve or Reject: A Checklist for AI-Drafted Replies

A reply is read under someone else's post, by people who did not come for you, and the author decides in a second whether it was worth their thread. An AI draft is fluent enough to pass a skim, which is how the weak ones get sent. These six checks are the ones we run on every reply a model drafts before a person sends it, and they are the same six whatever tool wrote the draft.

Read the reply against the post

Keep the post you are answering open beside the draft. Each question below is answered by comparing the two, never by reading the reply alone.

  1. 1

    Does it answer what the post actually says?

    Catches: A reply to the category, not the post

    A post about hiring a first salesperson gets a reply about hiring in general, because the model saw the word and answered the genre. The author knows the difference at once, and so does everyone reading the thread.

    Fix: Use a term the post itself uses. If the reply would fit under ten other posts, it answers none of them.

    Weak
    “Hiring is always the hardest part of scaling. Great reminder.”
    Better
    “Did the first salesperson get a quota from day one, or a ramp?”
  2. 2

    Did it pick one angle?

    Catches: Three half-points in three lines

    A long post offers five threads. A reply that touches all of them adds a sentence to each and a thought to none. The one that goes a layer deeper on a single thread is the one that gets answered.

    Fix: Keep the line that goes deepest. Cut the rest.

    Weak
    “Pricing matters, onboarding matters, and the churn point is real too.”
    Better
    “On the churn point: was it the month-two drop, or people who never came back after the first week?”
  3. 3

    Does it add something, or only agree?

    Catches: Agreement dressed as a reply

    The author has seen a hundred replies that agree. A specific point from your own work, or a question about how they did it, is what makes someone open your profile.

    Fix: Add one point or one question. If you have neither, do not reply.

    Weak
    “Could not agree more, this is exactly right.”
    Better
    “We made the same cut and kept the weekly email. Did you drop yours or move it?”
  4. 4

    Is every fact in the reply also in the post?

    Catches: Invented numbers, dates, timelines and "we did this"

    Ask a model to reply as a builder and it invents the builder's history: a day count, a launch last quarter, a percentage. None of it is in the post it answers, and none of it happened to you. A number you cannot back up is the fastest way to lose the author.

    Fix: Rewrite the claim as a question, or as a general observation that needs no number.

    Weak
    “We cut our onboarding time 40% doing this last month.”
    Better
    “Did the shorter onboarding change who stayed, or only how fast they started?”
  5. 5

    Is it one to three short lines?

    Catches: The essay under someone else's post

    A reply is read in a feed, under a post that is not yours. Past three lines it stops reading as a reply and starts reading as a takeover, and your point is now the last thing anyone sees.

    Fix: Cut to the line that carries the point. Say the rest in a post of your own.

  6. 6

    Would you delete it as generic praise if it were left on your post?

    Catches: "Great insight!" in all its spellings

    You already know this one from the other side. The replies you skim past on your own posts are the ones a model writes most fluently.

    Fix: Delete it. No reply is better than this one.

Edit or delete

If cutting one line leaves a reply worth sending, edit it. If the line is the reply, delete it and move on to the next post.

Edit when the problem is

  • One invented detail you can turn into a question
  • A second angle you can cut
  • A fourth line

Delete when

  • It agrees and adds nothing
  • It answers the category and not the post
  • You would not say it to the author across a table

Reply, then be worth following

A good reply earns a profile visit, and that is all it earns. The visit turns into a follow only if the profile has something to offer: a bio that says what the company does, and a last post from this week rather than this spring. Replies without posts of your own send people to an empty room.

So pair ten to fifteen replies a day with a steady week of your own posts, planned from the two-lane idea system and each read against the post checklist before it goes out.

The checklist

  • It uses a term from the post itself.
  • It holds one angle.
  • It adds a point or asks a question.
  • Every number, date and "we did" is in the post it answers.
  • It is one to three short lines.
  • You would not skim past it on your own post.

Get the next playbook in your inbox

Practical rules for running social media at a small company, from what we learn building XreplyAI.

No spam. Unsubscribe anytime.

Frequently asked questions

How do I review an AI-drafted reply before sending it?
Read it against the post it answers, not on its own. Check that it uses a term from the post, holds one angle, adds a point or a question, puts no number or date in your mouth that the post did not supply, and stays under three lines. Then ask whether you would skim past it on your own post.
Why do AI-written replies invent numbers and timelines?
A model asked to reply as a builder writes the builder's history: a day count, a result, a launch. The post it answers gives it none of that, so it fills the gap. Rewrite any such claim as a question or a general observation.
Should I edit a weak reply or delete it?
Edit when the problem is one line: an invented detail, a second angle, a fourth line. Delete when the reply is the problem: it only agrees, it answers the category instead of the post, or you would not say it to the author in person.

Keep building without disappearing

XreplyAI is the social media person you haven't hired yet: it learns your company, finds what's worth saying, and writes it. Drafts wait for your approval, so this checklist stays yours to run.

Start your free trial →