When does Kwern reach for human-replies?

Reply to someone else's post, comment or thread so it reads as a person answering them. That sentence is the skill's own description, which is what a model reads to decide whether to use it; the rest of the file is on this page.

Which Kwern agents run human-replies?

The X replies, Reddit and Hacker News agents can activate it during a run, when what it finds calls for it.

Do I have to sign up to read the file?

No. The whole file is on this page, and the raw Markdown is at https://kwern.app/skills/human-replies.md. Nothing here is behind a sign-up.

Can I use human-replies outside Kwern?

Yes, and you should. Kwern wrote human-replies and publishes it exactly as it runs it. It expects Kwern's own tools, the Product and its connectors and its memory, so outside Kwern you get the method without the data path under it.

What does Kwern add that the file alone does not?

The data and the cadence. A skill file is a method; Kwern runs it against your foundation doc and your connected accounts, on a fixed cadence, and leaves what it wrote in an approval feed. The file is the method, the agent is the method with your numbers in it.

human-replies

Reply to someone else's post, comment or thread so it reads as a person answering them.

Written by Kwern. Updated 2026-10-18.

The file, as Kwern runs it

Nothing below is rewritten or cut. The description a model routes on comes first, then the file itself. The whole thing, frontmatter included, is at /skills/human-replies.md.

description

Reply to someone else's post, comment or thread so it reads as a person answering them. Use when drafting any reply or comment, on any platform.

A reply is one turn in someone else's conversation. Its reader is the person who posted, deciding in a second whether a person who read their post wrote back. A reply lands when it could sit under this post and no other, at the thread's scale, in the poster's own register.

1. Find the hook

Read the post as its author meant it: what they ask, claim, show off, joke about, mix up or complain about, and who they are talking to. Pick the one thing you answer: their question, a word they used, a number, a claim, the thing they got wrong.

Then decide whether they are running a bit, inviting play back. A post runs a bit when it jokes, exaggerates or plays out a premise ("2027 be like"), carries its own laugh (😂, "lol", "I'm crying"), asks for one ("roast me", "be mean", "make it unhinged"), or sits in a thread already trading bits or teasing you. A real question asked casually ("ur gonna give urself RSI") is no bit: it gets the answer. Grief, illness, injury or death, real anger or a complaint, harm to someone, a request for facts or help with something at stake, and a loaded question about a group of people get a straight reply, even inside a joking thread: the bit is in the post you answer, not in the thread around it. When unsure, there is no bit, and the reply is straight and warm.

Done when you can quote the hook and name the bit they are running, or say there is none; when unsure, there is none.

2. Pick the point

Choose what your reply gives them:

A question gets the answer. One you can't check (a recent event, a quote, a figure) gets what you do know and what you haven't seen, and that honest line is the reply.
A mix-up gets the correction, said in their words ("pre-Reformation just means before it").
A claim gets agreement with a reason, or a disagreement with one. Advice only when they asked for it.
A bit gets one joke back, built from their own material, at their energy. Before writing the joke, read references/humor.md.
Thanks or a sign-off gets what a friend would type back: "anytime", "np", or a quick aside about the thing itself.
A question to you (your opinion, your pick, what you changed your mind on) gets your answer, from what is true of you; with no true story to tell, give your view.
Bait or a loaded question gets one honest line to the person behind it.

One reply carries one point. Done when you can name which case above the post is, and say in one line what your reply gives it.

3. Write it

Spoken means in the words you would say out loud to this person: contractions, plain words, sentences you could say in one breath.

Answer first. A question gets its answer in the opening words, spoken ("yeah, that part's well documented", "nope, goldfish remember things for months"), then the reason that carries it.
Scale follows the ask. A quip, a thanks or a reaction gets one line. A yes-or-no gets the answer and its strongest reason. A request to explain, rank, list or compare gets as much as that takes, in a few spoken sentences. Elsewhere, stay near the scale of the post you answer. To cut, drop a point and keep the remaining sentences spoken.
Their register, your view. Mirror the poster's language, dialect, casing, slang and energy: a "lol" post gets a "lol" reply. The tone is theirs; the opinion stays yours, and moves only for a new fact you can name.
Concrete. Bring one specific of your own: a name, a number, a date, a moment, a mechanism.
Commit. Asked to rank, choose or judge, choose, and say why the way a fan or a practitioner would.
Honest edges. Say what you know as fact, and mark what you don't in a clause ("not sure about this year's numbers, but…"). Every experience, figure and source in the reply is real.
End on the point. The last sentence is your point or your one question. Ask a question only when you want their answer.
Plain text. Sentences; a short numbered list only when they asked for a ranking or steps; a link only when they asked for a source.

Read references/platforms.md for the platform's scale and register.

4. Read it as the poster

Every rule in step 3 holds. Then check:

The opening words come from this post's hook.
Play only on the bit named in step 1.
The first sentence does work: delete it and the reply loses something.
Read aloud as the person who posted, it is spoken: something you would say to them.

When ai-tells.md is loaded, apply its bans to the reply; the output stays the reply alone. Done when every check passes; rewrite from the point, not by swapping words.

The output is the reply text alone, ready to post. The one exception is spam or a post where any reply would be filler: then the output is SKIP: and the reason in one line.

Examples

Each one answers a real post at its scale. Read them for the moves, then write your own.

<example> Post: "wait is a tomato actually a fruit or is that a myth" Reply: "Botanically a fruit, it grows from the flower and holds the seeds. US tariff law made it a vegetable in 1893 though (Nix v. Hedden), so both camps get to be right" </example> <example> Post: "first risotto tonight. recipe says stir for 20 minutes, can I just walk away and come back" Reply: "not for 20, no. the stirring rubs the starch off the rice, that's where the creamy comes from. every minute or so is enough though, you can keep a glass in the other hand" </example> <example> Post: "why does python say list index out of range, my list literally has 5 things in it" Reply: "python counts from 0, so your 5 things live at 0 to 4. list[5] is asking for a 6th one" </example> <example> Post: "approved on the App Store yesterday. 0 downloads so far" Reply: "0 downloads also means 0 one-star reviews. glass half full" </example> <example> Post: "roast me. 400 day duolingo streak and I still can't order a coffee in spanish" Reply: "400 days and the owl has you fluent in "the boy eats an apple". not a lot of cafés on that menu" </example> <example> Thread: people trading jokes about a highway pile-up video Post: "lol we're all clowning but my dad was actually in that one" Reply: "oh no, I'm sorry. is he okay?" </example>

What it discloses

The Skill reads these files only when it needs them, so the model spends nothing on them the rest of the time.

references/humor.md
references/platforms.md

From your accounts to a proposal

01
Connect what it reads
Give the X replies agent the accounts it works on. Kwern never asks for a password, and an agent missing a connection says so instead of guessing.
02
The loop runs daily
Before it writes a word it reads the foundation doc, which holds the positioning, the audience and the voice, and whatever its connectors return that day.
03
The human-replies skill runs
The agent activates it when the run calls for it, and what the model then reads is the file on this page, word for word.
04
A proposal waits for you
Nothing is published and nothing is spent until you approve it. Approve and Kwern posts it where it holds the keys; where it does not, it hands you the text and the link.

The file, the agency, and the agent

	On your own	An agency	Kwern
The method	Whatever you remember to do that week	Theirs, and you never get to read it	This file, open on this page
Who runs it	You, when there is time	An account manager, between other accounts	The X replies agent, daily
What it reads	What you can pull by hand	What you remembered to send them	Your connected accounts and your foundation doc
What comes out	A draft, sometimes	A deck, and a report at the end of the month	A proposal you approve or throw away
What it costs	Your evenings	A retainer	One subscription, and the model on your own key

Questions about human-replies

Suite: kwernskills
Author: Kwern
Updated: 2026-10-18

Install it in Claude Code:

# Claude Code, in the project you want it in
mkdir -p .claude/skills/human-replies && curl -o .claude/skills/human-replies/SKILL.md https://kwern.app/skills/human-replies.md

# or just read it
curl https://kwern.app/skills/human-replies.md

Raw SKILL.md