Nothing to paste? — a resolved rate-limit ticket with exact steps, one hedge and one internal-only note. Or , rendered from a saved result with no run and no charge.
Paste the thread, not a summary
The raw material carries the exact error text, the exact steps and the exact paths a searchable article needs. The source scan runs in your browser as you type — free, no sign-in, and nothing you paste is sent anywhere until you press Forge the article — and flags what the article can be built from.
Facts from the source — gaps named, not filled
Every step, path, limit and version in the article comes from what you pasted. A hedge like “I believe this also applies to…” lands under Gaps to fill, and anything marked internal-only stays out entirely.
Read the verdict, then publish
Every draft carries its publish-readiness: an unfilled gap rules out Publish as is. The title linter and article lint re-check the draft in your browser, against the same searchability rules the writer follows.
| Type | Purpose | When the source calls for it |
|---|
| Good title | Bad title | Why |
|---|
From the source skill's searchability table — customers search the symptom, the tool name and the exact error text.
What file do I get, and will my knowledge base accept it?
Two, and they are the two that publish. The .md download opens with a real
--- YAML front matter block carrying the title, category, tags, audience and article type,
so a docs repo (Docusaurus, MkDocs, Hugo, Jekyll), Notion or Obsidian reads them as metadata rather
than as body text. The .html download is a complete HTML document, which is what the
article body field of Zendesk Guide, Salesforce Knowledge, Intercom Articles, HubSpot KB and
Confluence expects; pasting raw markdown into those rich-text editors publishes literal
### and literal backticks to customers. The file is named after the article title, not
a fixed filename.
Do the internal notes end up in the customer-facing file?
No. Publishing notes, Gaps to fill and Related articles name ticket numbers, customer names and facts nobody has confirmed yet. Both article exports carry the article body and its metadata only. The full brief is a separate, separately labelled download for the editor.
Does what I paste leave my browser before I run it?
No. The source scan, the title linter and the article lint all run inside this page. The running cost estimate is priced from filler text of the same length as what you pasted, not from the text itself, so no keystroke sends your ticket anywhere. Your material is sent to the model once, when you press Forge the article.
What happens if my ticket thread is longer than the limit?
The middle is dropped and both ends are kept, on whole line boundaries, and the cut is announced inside the text that is sent. A ticket thread carries the symptom at the top and the resolution — the numbered steps, the exact path, the root cause — at the bottom, so cutting the tail would delete the article. The model is told how much was removed and reports it in the summary.
What happens if the response is cut short?
Whatever fully arrived is recovered and shown, and every field that did not arrive is labelled "not received" — never "None", which in this contract means the model looked and found nothing. The last part-written line is discarded rather than displayed as a finished sentence, and a verdict that never arrived is never guessed at.
Can I run the same source again without being charged twice?
Each press of Forge the article is one billable run, and a re-run of unchanged source really does run again rather than replaying the old answer. What is protected is the accidental duplicate: every run carries an idempotency key derived from your input and that press, so a dropped connection or an automatic reformat of a malformed reply cannot bill the same press twice.