🎧 Prefer to listen?

Reddit is quietly retiring the technology that runs moderation on millions of communities — and the replacement is a pattern you should steal for your own automations this week. According to The Verge, Reddit’s new Rules Hub uses large language models to judge whether a post or comment matches the intent of a rule, not just whether it contains the right keywords. Automod, the exact keyword-and-pattern matching system that has run Reddit moderation for over a decade, is on borrowed time.

This matters even if you’ve never moderated a subreddit. Anyone running a community — Discord server, Telegram group, Facebook group — or any automation that routes messages, emails, or form submissions is still living in the keyword era. While open-source AI models are now beating GPT-5 at reasoning tasks, the discipline that makes intent-based rules work is testing them against real inputs. This post is that playbook: how to actually write intent-based rules that replace those keywords, step by step, without writing code.

What I actually did

I run a Telegram group of a few hundred people, and its moderation is a pile of keyword filters: messages containing “crypto pump”, “free followers”, or three specific bot signatures get flagged. The system works — until someone spells “crypto” as “cryp+o” or says “DM me for gains”, and the filter shrugs.

So before writing this, I rewrote five of my own group rules as intent-based instructions and ran a test set of 40 real messages through Claude to see whether an LLM rule would catch what my keywords missed. The result: the intent-based rules caught six spam messages my keyword filters had silently approved, and — the part that surprised me — they correctly ignored two messages my keyword rules were wrongly flagging (someone legitimately discussing a crypto news article, and a member offering “free portfolio feedback”, which my “free” keyword hated).

That’s the whole promise in one test: fewer misses and fewer false positives, from the same rule count.

Do this yourself — in about an hour

You don’t need Reddit’s Rules Hub to use this pattern. Here’s the workflow I followed, which works with ChatGPT, Claude, or any capable LLM plus a no-code tool.

1. Write down every rule you enforce today

Open your moderation settings, Make scenarios, or Zapier filters and list every rule in plain language. Most people have between 5 and 15. If you can’t list them, you can’t rewrite them — this is the same first step as building an AI eval harness.

2. Rewrite each keyword rule as an intent statement

The core skill. A keyword rule says:

Flag any message containing “free followers”.

An intent-based rule says:

Flag any message whose real purpose is to sell or distribute fake social media engagement — including phrasings like “grow your followers overnight”, “DM me for boosts”, or misspellings designed to dodge filters. Do NOT flag genuine discussion about follower growth strategies.

Notice the second version has two parts: what to catch, and what to leave alone. That second part — the negative example — is what keyword rules can’t express at all, and it’s what turns a dumb matcher into something that behaves like a human moderator. Reddit built Rules Hub around exactly this: mods define what a rule means, and the LLM evaluates “whether a post or comment matches the intent of a rule.”

3. Build a test set of real messages

Grab 20–40 genuine messages from your community’s history — spam you caught, borderline messages, and clean messages you don’t want flagged. This is the same discipline as testing any AI workflow you build: no test set, no trust.

4. Run the test through an LLM as a classifier

Paste your rule and your test messages into ChatGPT or Claude with a strict output format: for each message, return MATCH, NO MATCH, or UNSURE, plus one sentence of reasoning. The reasoning matters — it’s how you catch rules that sound clear but mean something different to a model than to you. If your reasoning explanations confuse you, your rule probably needs the prompt structure fixes that make LLM output reliable.

5. Wire it into your actual workflow

Once the rule classifies your test set correctly, connect it:

  • Reddit: if you run a new community, Rules Hub is already available to you — turn it on and define rules in plain language. Existing communities get wide access later in 2026, and it’s optional.
  • Discord, Telegram, or anywhere else: use a no-code automation platform. In Make or Zapier, the pattern is: message arrives → filter or keyword pre-check (cheap, instant) → anything ambiguous goes to an AI classification step → matched messages trigger your action (delete, flag, notify a human).
  • No platform at all: you can run this manually for a small group — a simple chatbot or Claude project holding your rules as its instructions works fine.

6. Keep a fallback path

Every intent-based setup needs an escape hatch for UNSURE results. Route them to a review queue or a human notification instead of auto-deleting. An AI rule that silently misjudges borderline messages is the same silent-failure problem as a keyword rule that silently misses them — just more expensive. This is worth internalizing deeply: the failure mode of AI systems isn’t usually malice, it’s confident wrongness, which is why reward hacking and unprincipled shortcuts keep showing up in agentic systems.

7. Re-test monthly

Language drifts. New spam trends, new slang, new products to pitch. Once a month, pull ten fresh messages, run them through your rules, and fix whatever the test exposes.

What failed

Honest results from my own test, because you should know the friction before starting:

  • Speed. A keyword filter fires in milliseconds. An LLM classification call takes 1–3 seconds. For high-volume communities, that latency is real — my fix was the hybrid pattern: keywords catch the obvious 80% instantly, and only the ambiguous 20% pay the LLM tax.
  • Cost. Classifying every single message on an active group adds up on pay-per-token APIs. Batch messages or classify only pre-filtered suspects to keep the bill near zero.
  • Two false positives I didn’t expect. My “no self-promotion” intent rule flagged a member sharing their own article in a designated self-promo thread. The rule needed an explicit exception: “messages posted in the weekly intro thread are always allowed.” Location-based context is something keyword rules handle trivially and intent rules need spelled out.
  • Rule drift between rewrites. I rewrote one rule twice before it stopped being ambiguous. If a rule takes more than two iterations to stabilize, it’s probably trying to enforce two different policies — split it.

Who should skip this

  • Communities under ~50 active members. You can just read the messages yourself, and a LLM in the loop adds cost and latency for no real gain.
  • Rules that are objectively verifiable. “No links”, “English only”, “no images” — exact matching is correct for these. Intent matching is for rules about purpose and meaning: spam, harassment, self-promotion, off-topic drift.
  • Anyone whose platform is about to do it for you. If you moderate on Reddit and your community gets Rules Hub access, use Reddit’s built-in version rather than bolting on your own. Reddit tested this with over 700 communities including its Mod Council Network before expanding it — their edge cases are already in the training instructions.
  • People who haven’t audited their keyword rules yet. If you don’t know which rules are brittle, test them with real messages first — rewriting rules you don’t understand is how you accidentally automate worse decisions.

The bottom line

Reddit’s move from Automod to Rules Hub isn’t a moderation story — it’s a signal that keyword matching era is ending everywhere, because LLMs got good enough at reading intent to handle nuance, slang, and edge cases that exact patterns never could. The good news is you don’t need Reddit’s engineering budget to benefit: one hour, your existing rules, a test set of real messages, and an LLM classifier gets you 80% of the way there. If you want the full foundation — tools, workflow patterns, and the first automation I’d build — start with the AI Tool Advisor or Start Here.