Copydesk (AI Editorial QA)
Copydesk is the AI copy desk for WordPress. It reads each post when it enters the statuses you choose and flags objective typos only — misspelings, duplicated words, broken capitalization and punctuation — right in the editor. Your own OpenAI or Anthropic key, your rules, your content untouched.
Install
This is the repository archive. It installs, but the folder name will carry the branch suffix and updates will not flow.
Install straight from GitHub with WP-CLI:
wp plugin install https://github.com/tedslaterdev/copydesk/archive/refs/heads/main.zipCopydesk (formerly Editorial QA)
Your AI copy desk for WordPress. One click in the post editor — Check this post — checks the story for typos and clear grammar errors and, if you like, adds Editor's Notes: two paragraphs (or one, or three) written out in plain words, ready to paste into the story where each one goes — plus smaller ideas such as missing context, open questions, other perspectives, and sources. Works with OpenAI or Anthropic (Claude), on your own API key.
Description
Copydesk puts a small Copydesk box in the post editor's sidebar. When a writer or editor clicks Check this post, Copydesk sends the last saved version of the post to your chosen AI provider — OpenAI or Anthropic (Claude) — and shows the result right there, while the person who can fix it is still in the post:
- Typos & grammar — Clean or N to fix beside the section's title, and a short list of fixes that read wrong → right (
teh budget→ the budget), plus places where the story contradicts itself, such as "Doe, 52 or Doe, 54 — the story says both; check which is right". - Editor's Notes (optional) — paragraphs to paste (two by default), each placed after the paragraph it belongs with and written out in plain words, ready to copy and paste, then up to five smaller suggestions folded under More suggestions.
The two run side by side, and each shows as soon as it's ready, so the typo list usually appears first.
The typo check is intentionally conservative. It flags only:
- Misspellings — including a name spelled differently from the same name elsewhere in the post
- Typos that spell another real word — "form the start" for "from the start", or "the sate budget" when the post says "the state budget" elsewhere. A word that makes sense as written is left alone.
- Duplicated words (e.g. "the the")
- Broken capitalization (e.g. "thE", random mid-word caps) — it does not enforce title case or headline style
- Broken punctuation (e.g. "??", "!!", "..")
- Clear grammar errors — the wrong word form (its/it's, their/there/they're, then/than, your/you're), a missing or extra small word, subject–verb agreement, or a wrong verb form
- The story contradicting itself — the same fact given two ways, such as two ages for one person or two dates for the same event. Copydesk shows both versions and leaves it to you to say which is right.
It deliberately ignores:
- Quotation marks and apostrophes (curly or straight)
- Style and preference — comma choices, passive voice, sentence fragments in news style, tone, and political/editorial judgment
- How brands are styled ("Youtube" vs "YouTube") and compound-word choices
- Names it doesn't recognize. It never changes a name to a different name — people, companies, and products get renamed, and an AI's knowledge can be out of date.
- Anything inside
<blockquote>(quoted source material is removed before the typo check, so you don't get "corrections" to other people's words)
Each request also carries today's date, so a recent date or a just-announced event isn't "corrected".
The model is constrained with a structured-output (JSON schema) contract — OpenAI's strict Responses-API format or Anthropic's output_config.format — and the list editors see is rebuilt server-side from the validated results only: a fix whose text isn't actually in the post is dropped, and the model can't inject free-form commentary.
Editor's Notes (optional)
Tick Include Editor's Notes in Copydesk → Settings (on by default for new installs) and every check also helps the writer grow the story — it suggests what to add, never what to correct.
- Paragraphs to paste. Two by default — one, two or three under Paragraphs to add in Settings — listed in story order. Each says where it goes ("Add after ¶3" plus that paragraph's opening words, or "Add at the end") and what it does — sums up where things stand, says what happens next or why it matters, explains what something in the story means, connects one part to the next, sets up a quote, or gives background a reader needs. None goes between someone's words and their quote.
- Written out for you, in plain words. Each paragraph comes in full — usually 3 to 6 short sentences at about a 5th-grade reading level, in the story's tense and tone — with a Copy button. It reads fine as is: paste it, or edit it first. It tells the reader something the paragraphs around it don't — never a repeat of the one above it — and adds a new fact only when the AI is sure of it. In the block editor, a two-paragraph draft pastes as two paragraphs.
- Guest posts stay in the author's words. In a first-person piece — a guest post, column, or personal essay — nothing is drafted in the author's voice; the suggestions say what to ask the author for.
- Facts flagged, not invented. A draft uses the story's facts and settled background (what a law is, who a long-known group is, what happened years ago). Anything recent or uncertain that the story doesn't say — a current title, a latest figure, someone's response — the AI writes around instead of guessing. Numbers a draft uses that the story doesn't have are highlighted so the writer checks them. Every draft is checked before it's shown: no quotes, nothing like "declined to comment" or "didn't respond", no blanks to fill in, nothing cut off.
- Up to five more suggestions, folded under More suggestions so the paragraphs come first. Each is tagged Context, Open question, Perspective, or Source and pinned to the paragraph it's about.
- Turn drafts off with Write the paragraphs out, ready to paste in Settings; each paragraph to add then says what it should cover, for the writer to write.
- A warning for blanks left behind. If you mark a missing fact with a
[TK: …]blank yourself, the Copydesk box says so while the saved post still contains one — no request is sent for that. - Suggestions, never rewrites. Copydesk never edits the post; the writer decides what to paste. The prompt forbids inventing quotes or sources and guessing at anything recent.
- Stays honest about freshness. If the post changes after a check, the box says so.
- Knows what's already linked. Linked words are marked as links in what the notes read (the addresses themselves are never sent), so notes don't ask for a source the story already links.
- Editable prompts. Both prompts are shown in Settings with a Reset to default button.
Keeping costs predictable
Copydesk uses your OpenAI or Anthropic API key and bills to your account with that provider. A check is at most two requests — one for typos, one for notes — and often fewer:
- An unchanged post reuses its typo result for free. Notes always run fresh, since a second look can differ. With notes off, checking an unchanged post again sends nothing at all.
- A per-post limit — by default each post can be checked 3 times in any 20 minutes, shared by everyone working on it. When it's reached, the box says when the post can be checked again and the button turns itself back on then. Only checks that send a request count. Set the limit (or turn it off) in Settings.
- An optional monthly budget stops checks once estimated spend for the month reaches it.
- The Costs page shows estimated spend by day, by model, and by part (typos vs. notes), and lists the last 50 OpenAI requests: who started each, how long it waited and took, and how it ended.
Installation
- Upload the
copydeskfolder to/wp-content/plugins/, or install the ZIP via Plugins → Add New → Upload. - Activate Copydesk (AI Editorial QA).
- Go to Copydesk → Settings, choose your AI provider (OpenAI or Anthropic), provide that provider's API key (see below), choose a model, and pick which post types show the Copydesk box.
Providing the API key
Each provider has its own key field and its own optional wp-config constant. You only need to configure the provider you select.
-
Settings field — paste the key into the matching field on Copydesk → Settings. It is stored in
wp_options. -
wp-config.php constant (recommended) — keep the secret out of the database:
define('OGM_EQA_OPENAI_API_KEY', 'sk-...'); // OpenAI define('OGM_EQA_ANTHROPIC_API_KEY', 'sk-ant-...'); // AnthropicWhen a constant is defined, it takes precedence and the matching settings field is ignored at runtime.
Frequently Asked Questions
Who can use the Check this post button?
Anyone who can edit the post — writers on their own drafts, editors on any post they can edit. Settings and the Costs page are for administrators.
Does it check posts automatically?
No — not since 2.0. Results from an automatic scan arrived after the writer had moved on, so flagged typos often went live anyway. Now nothing is sent until someone clicks Check this post, and the result appears where they can fix it.
Does it check my latest edits?
It checks the last saved version of the post. If the editor has unsaved changes when you click, the box says edits since your last save may not be included — save first to be sure. (An autosave of your own draft may already include them.)
How does the check limit work?
Each post can be checked a set number of times in a rolling window — 3 checks in any 20 minutes by default. The limit is per post and shared by everyone who works on it, so an editor reviewing ten posts isn't held up. Only checks that send a request to the AI count; checking an unchanged post with notes off is free and never counts. Change both numbers in Settings → Copydesk box, or set the limit to 0 to turn it off. (Developers can adjust it per user with the ogm_eqa_check_limit filter.)
What is "Max output tokens"?
It caps how long the AI's typo list can be for a single check — i.e. how many fixes it can list for one post. It is not a daily or monthly budget or a cost cap over time. The default (1000) comfortably holds dozens of findings; raise it only if a very long post has its list cut off. For models that "think" before answering, Copydesk adds separate room for the thinking on top of this cap — it's a ceiling, and you're billed only for tokens actually used. Editor's Notes has its own fixed cap and isn't affected by this setting.
Does it change my content?
No. Copydesk never edits your posts. It only reports suspected typos and grammar errors — and, with Editor's Notes on, suggestions — for a human to act on.
Can I paste the suggestions?
Yes — that's the idea. With Write the paragraphs out, ready to paste on (the default), each paragraph to add comes written out in full, in plain words, with a Copy button. Paste it where the label points ("Add after ¶3"); it reads fine as is, and you can edit it like anything else you write — there are no blanks to fill in. The smaller suggestions under More suggestions are pointers, not text to paste. In a first-person piece such as a guest post there's nothing to paste: Copydesk doesn't write in the author's voice, so the suggestions say what to ask the author for.
Can I get more (or fewer) paragraphs?
Yes. Paragraphs to add in Settings sets how many each check writes: one, two (the default), or three. More paragraphs means a slightly longer reply, so a check costs a little more.
A check keeps saying "the server may have cut the request off"
Before 2.5.0, that happened on every check of a story with an emoji (often in an embedded tweet or post) when the site's database tables can't store emoji (older utf8mb3 tables). From 2.5.0 Copydesk stores such characters safely. If Copydesk ever can't save a check, it now says so ("Copydesk couldn’t save this check in the database…") and stops the request instead.
Can I trust the facts in a draft?
Check them. A draft mostly explains your story's own facts and what they mean, and can include background the AI is confident of — what a law or program is, who a long-known organization is, what happened years ago — so it reads like a finished paragraph. Copydesk highlights every number in a draft that isn't in your story, and each paragraph to add says where to confirm its facts. Names and other details aren't checked automatically, so treat a draft like copy from a new reporter: read it against your reporting before you publish.
What's [TK: …]?
"TK" is the newsroom mark for "to come": a placeholder for a fact that isn't in hand yet. From 2.4.0 Copydesk's drafts never contain one (versions 2.1–2.3 left blanks such as [TK: vote count] in drafts). If you type your own, the Copydesk box warns while any [TK: …] blank is left in the saved post, and the typo check never flags one.
Why did it flag something that isn't a typo?
AI is not perfect. The box notes this explicitly. Treat the list as suggestions, not corrections. If your site has spellings it keeps flagging, add them to the typo & grammar instructions in Settings → Advanced.
Does it cost money?
Yes — it uses your OpenAI or Anthropic API key and bills to your account with that provider. With a recommended low-cost model, a typo check costs about a tenth of a cent and a notes request a little more, depending on the model and post length. Drafted paragraphs add about $0.0002 to a notes request on the recommended models. See Keeping costs predictable above; the Costs page lists typo checks and notes separately, and both count toward the monthly budget cap.
Which provider/model should I use?
Pick whichever provider you already have a key for — typo detection is an easy task and any inexpensive current model handles it well. The model list marks one model per provider as recommended:
- OpenAI:
gpt-6-luna(Responses API + strict structured outputs) — OpenAI's newest low-cost model. - Anthropic:
claude-haiku-5-5(Messages API + structured outputs) — Claude's newest low-cost model.
Flagship models cost 20–100× more per check without finding more typos. New installs start on the recommended model; an existing site keeps whatever model it already uses until you change it (a site that never saved Settings stays on the earlier built-in default, gpt-4o-mini or claude-haiku-4-5).
Each provider's Model setting is a dropdown with its own Refresh models button: click it to populate the list from the models available on your account for that provider (save the matching API key first). The list keeps the newest models plus the low-cost tier, and leaves out models that don't suit this job (coding, research, "-pro", image, and audio models). Your previously-selected model always stays selectable.
What does the Effort setting do?
It sets how hard the model thinks before it writes Editor's Notes. Typo checks always use Low (or the model's nearest level): thinking harder doesn't find more typos, it just makes writers wait. The dropdown lists only the levels the selected model accepts; models without a thinking step, such as gpt-4o-mini or claude-haiku-4-5, have no Effort setting.
- Auto (the default) uses Medium for notes. That's the right choice for most sites.
- Higher levels can give more considered notes, but are slower and cost more, because thinking is billed as output.
- Extra high and Max can take minutes per request. How that plays out depends on the provider:
- OpenAI: each request runs as a background job. Copydesk waits up to 10 seconds, then the page checks on the job every few seconds (checks are free) and shows the result when it lands — after a reload, too. If the writer closes the page, Copydesk still collects the result in the background, so a request you've paid for is never lost. Notes still running after 3 minutes are stopped (5 at High, 10 at Extra high, 15 at Max), with a message to try again; a typo check, also after 3 minutes. On a busy day OpenAI can leave a background request waiting in its queue without starting it; one it hasn't started within 30 seconds is cancelled (it isn't billed) and sent again the ordinary way.
- Anthropic: a request has to finish inside your server's time limit. Many hosts stop requests after 30–60 seconds, and Cloudflare after about 100. If the server stops one, the box says so and you can try again at a lower Effort.
If you switch to a model that doesn't offer your chosen level, Copydesk uses the nearest level that model does offer.
Do I need Action Scheduler or WP-Cron?
No. Checks run when someone clicks. Copydesk uses Action Scheduler (if another plugin provides it) or WP-Cron only as a backstop: to collect an OpenAI result the page didn't wait for, so it's recorded and billed in your Costs page even if nobody opens the post again.
Where is my data sent?
For the typo check, the post title and body (with shortcodes, blockquotes, and HTML removed) and today's date are sent to the API of the provider you selected — OpenAI or Anthropic. For Editor's Notes, the title and body are sent the same way except that quoted passages (blockquotes and embedded posts) are included and linked words are marked as links, because whether a story already quotes a response or links a source is exactly what the notes need to judge. Link addresses (URLs) are never sent, and a line that is only a web address — an embedded video or post — is left out of the typo check and reaches Editor's Notes only as the name of the site it's from. An address typed into a sentence is sent as part of the text. Password-protected posts are never sent. Nothing else leaves your site.
With OpenAI, requests run in OpenAI's background mode, which means OpenAI keeps each request and its reply until Copydesk collects the result. Copydesk has always used the Responses API's default storage setting, so this isn't new retention. If your account or model doesn't accept background requests, Copydesk sends ordinary ones instead.
Screenshots
- The Copydesk box in the post editor: typo and grammar fixes, then Editor's Notes — paragraphs to add, drafted to copy — and Additional notes.
- The check limit, with the time the post can be checked again.
- Settings: provider, model, Effort, the check limit, and Editor's Notes.
- The Costs page, with typo checks and Editor's Notes listed separately.
Changelog
2.6.3
A request OpenAI never starts is sent again.
- No more waiting in a stuck queue. OpenAI sometimes accepts a background request and then leaves it waiting in its queue for minutes without starting it. Now, if OpenAI hasn't started a request within 30 seconds, Copydesk cancels it (a request that never started isn't billed) and sends it again the ordinary way, while the page keeps waiting. An ordinary request has to finish within your host's time limit; if the host stops it, the box says so and the writer can check again. Developers can change the wait, or turn this off with 0, using the
ogm_eqa_queue_max_waitfilter. - Typo checks stop after 3 minutes, like Editor's Notes at Auto, instead of 15. The message suggests a faster model rather than Effort, which applies to notes only.
- A clearer message. A request OpenAI still hasn't started when it's stopped now says so, and to try again in a minute, instead of suggesting a lower Effort.
- Costs → Recent AI requests lists a request sent again twice: "Sent again: never started", then the ordinary request ("Done (sent directly)" or "Failed (sent directly)").
2.6.2
Slow Editor's Notes are stopped sooner, and every request is on record.
- Notes stop after 3 minutes. An Editor's Notes request that is still running after 3 minutes is stopped, and the box says to try again, instead of keeping the writer waiting for up to 15 minutes. A higher Effort gets longer: 5 minutes at High, 10 at Extra high, 15 at Max. Typo checks are unchanged. Developers can change the limits with the
ogm_eqa_job_max_agefilter. - Recent AI requests on the Costs page. Administrators see the last 50 OpenAI requests: the post, who started it, how long it waited in OpenAI's queue and how long it took, how many tokens it wrote and how many of those were thinking, the processing tier OpenAI used, and how it ended. A fingerprint shows when two requests sent exactly the same thing. Only these numbers are kept, never story text or the AI's reply.
2.6.1
Fixes from a review. A review of everything since 2.1.1 found three small problems in 2.6.0's box:
- A fix that reads two ways is shown as written. When the words in a typo fix themselves hold quoted phrases joined by "to" (say, 'yes' to 'no'), the box could strike out the wrong words. Such a line is now shown as written.
- Saving keeps More suggestions open. In the block editor, saving the post closed an open More suggestions; now it stays open, and keyboard focus stays where it was.
- Long web addresses wrap. A web address in a note or a Check line no longer pushes the block editor's sidebar sideways.
2.6.0
A quieter Copydesk box. Nothing it checks or finds has changed; it's easier on the eye:
- Status in the title. Clean, "3 to fix" or Error sits beside each section's title; a clean check no longer says so twice.
- Fixes read wrong → right. The wrong words struck through, an arrow, the right words in bold — no numbers and no BODY/TITLE labels. A fix in the headline is tagged Headline. A contradiction shows both versions and says to check which is right.
- Flatter paragraph cards. Copy sits in each card's top line, what the paragraph does is a small caption, and the draft reads like story text, without a box around it. A paragraph that opens with a quote no longer shows two quotation marks.
- Plainer reminders. "AI can be wrong" is one grey line, without the box and icon. The "Checks the last saved version" hint sits beside the button.
- Costs for admins only. The estimated cost of each check shows only to administrators; writers and editors see when the post was checked. The Costs page is unchanged.
2.5.2
Better placement. From the staging test of 2.5.1:
- Never splits a speaker. A paragraph to add no longer goes between a paragraph about what someone said and their quote in the next one; it goes after the quote. (A quoted back-and-forth was already kept together.)
- If you saved Settings with 2.5.1's notes prompt unchanged, you get the new one automatically. An edited prompt keeps its words.
2.5.1
Paragraphs that move the story along. From the first staging test of 2.5.0:
- No repeats. A drafted paragraph no longer restates the paragraph it follows, and doesn't end on a sentence that says nothing. With two or more, they're different kinds — not all explanations of a term — and "what happens next" and "why it matters" are now among what a paragraph can do.
- In story order. The paragraphs to add are listed by where they go ("at the end" last), so you can paste them top to bottom. The AI still picks which ones to suggest.
- If you saved Settings with 2.5.0's notes prompt unchanged, you get the new one automatically. An edited prompt keeps its words; the new drafting rules apply to it too.
2.5.0
Paragraphs to paste. From the newsroom's staging tests:
- Paragraphs written to paste in. Each check writes two paragraphs (one, two or three — new Paragraphs to add setting) in plain words at about a 5th-grade reading level, each placed after a named paragraph. They sum up, explain, connect, or set up what comes next; a new fact only when the AI is sure of it. Every paragraph gets written — around a missing fact, not with it.
- More suggestions, folded. The smaller notes (context, open questions, perspectives, sources) now sit under More suggestions (n), closed by default, and are no longer drafted.
- Fixed: "the server may have cut the request off" on stories with an emoji. On sites whose database tables can't hold emoji (utf8mb3), the notes for any story with an emoji — usually in an embedded tweet — were never saved, and every check of that story ended with that error. Copydesk now stores such characters safely and shows them as they are; if a check still can't be saved, it says so and stops the request.
- Settings: the drafts checkbox is now Write the paragraphs out, ready to paste.
2.4.0
Drafts you can paste as is. From a review of the staging tests:
- No more [TK: …] blanks. Drafted paragraphs and suggested wording are now complete sentences, ready to copy, adjust to your voice, and paste. Nothing to fill in.
- No guessing instead. A draft is written only when every fact in it is in the story or settled background. If an addition needs a fact the story doesn't have — who is running a review, what a defense said, a latest figure — the suggestion comes without a draft and its "Check:" line says where to find it. Expect fewer drafts, and no invented facts in their place.
- Any draft that still arrives with a blank or placeholder is dropped before it's shown. Notes saved before the update keep their blanks until the post is checked again.
- The box still warns about
[TK: …]blanks you type into a post yourself.
2.3.0
Typos that spell real words. From a test where one run caught "a long holdout" (meant: "a lone holdout") and later runs didn't:
- Real-word slips. The typo check now flags a word that is one or two letters off from the word the sentence clearly needs — "form the start", "an untied front" — especially when the post gets the phrase right elsewhere. A word that makes sense as written is never flagged. In a planted test story, three models each caught all six such slips with no false alarms.
- Why a draft is missing. When a paragraph to add or a note comes back without its draft, the box now records which check dropped it (no draft sent, too long, too many blanks, a quotation mark, and so on) in a
data-no-draftattribute for whoever is tuning the prompts. Writers don't see it. - A site still holding a saved copy of 2.2's built-in typo prompt picks up the new one automatically. Because the prompt changed, each post's next check runs the typo check once instead of reusing the last result. If you edited the typo prompt, click Reset to default in Settings → Advanced to get the new rule.
2.2.3
Guest posts and shorter drafts. From a test on a first-person guest post:
- Nothing in the author's voice. In a guest post, column, or personal essay, drafts no longer put sentences in the author's mouth ("I released the song on…"). Paragraphs to add and notes there say what to ask the author for, and skip background the author's readers already know.
- No padding. When most of a draft's facts would be
[TK: …]blanks, it's now a sentence or two instead of a paragraph filled out with sentences that say nothing ("putting a specific measure behind the response"). A draft needs 10 words outside its blanks, down from 20. - A site still holding a saved copy of 2.2.2's built-in Editor's Notes prompt picks up the new one automatically. Typo results are reused as before.
2.2.2
Tighter suggestions, and embeds stay private. From the next test on a real story:
- Tied to the news. Paragraphs to add now connect to what the story reports — what led to it, what it changes, what happens next, who is affected — instead of a textbook account of a well-known job or institution.
- No prying. Notes no longer ask the writer to find out private medical, family or personal details that the people in the story haven't shared.
- Embeds aren't sent. A line that is only a web address (an embedded video, post or tweet) is left out of the typo check, and Editor's Notes see only which site it's from. Posts with such a line show "changed since this check" once after upgrading, and their next typo check runs fresh.
- A site still holding a saved copy of 2.2.1's built-in Editor's Notes prompt picks up the new one automatically.
2.2.1
Sharper drafts. From a test on a real story:
- Specific, not vague. Drafted paragraphs now name the year, the place, the office, the law and the people — or leave a
[TK: …]blank — instead of phrases like "earlier reporting has examined". - Facts, not caveats. Drafts no longer add sentences about what the story's sources haven't shown, what something doesn't prove, or that two things are "separate questions". Every sentence should tell the reader something new; weighing the story's claims stays with the writer.
- Better placement. A paragraph to add is never placed in the middle of a quoted back-and-forth; it goes after the exchange ends.
- A site still holding a saved copy of 2.2.0's built-in Editor's Notes prompt picks up the new one automatically. Typo results are reused as before.
2.2.0
Editor's Notes now help you grow the story — with paragraphs drafted for you.
- Paragraphs to add. Editor's Notes now lead with up to two paragraphs the story could easily add — background a reader needs, what came before, what happens next, a reaction the story leaves out — each placed after the paragraph it belongs with ("Add after ¶3") or at the end, with a note on where to check its facts.
- Drafted in full. With Draft the paragraphs for writers to paste on (the default), each comes written out — 60 to 150 words in the story's style, sometimes two short paragraphs — with a Copy button. In the block editor, a two-paragraph draft pastes as two paragraphs.
- Numbers to check are highlighted. A draft may use background the AI is sure of; any number it uses that isn't in the story is highlighted, with a line saying to check it. Anything recent or uncertain is still a
[TK: …]blank. - Up to five smaller notes follow under Additional notes: Context, Open question, Perspective and Source. Verify notes are gone — Editor's Notes no longer point out errors.
- The typo check now finds a story that contradicts itself — the same fact given two ways, such as two ages for one person: "BODY: 'Doe, 52' doesn't match 'Doe, 54' — check which is right". Because its built-in prompt changed, an unchanged post's first check after updating runs the typo check once instead of reusing the stored result.
- The At a glance sidebar shows Editor's Notes as "On, with drafts", "On" or "Off".
- A notes request that was still running when you updated finishes the 2.1 way. Notes written before 2.2 show as they did.
- If you edited the Editor's Notes prompt, reset it (Settings → Reset to default) to get paragraph suggestions; an edited prompt is kept as written. Sites still on a saved copy of an earlier built-in prompt pick up the new one automatically.
2.1.1
Notes in two groups, plus fixes from a full review of 2.0 and 2.1.
- Top Editor's Notes and Additional notes. The three most important notes appear under Top Editor's Notes; any others (up to five) under Additional notes. Still eight at most. Notes written before 2.1 stay in one list, marked "In story order."
- The check limit can't be sidestepped. Only a post's newest check can run each part, so replaying an older check's request no longer sends another AI request.
- Suggested wording is checked more strictly. A decimal must match the story exactly ("$5.4 million" no longer passes because the story has a 4 and a 5 somewhere). HTML entities, non-breaking spaces and zero-width characters are decoded or removed before the checks, so they can't hide a quotation mark or "didn't respond". More no-comment phrasings are caught ("has yet to respond", "was not immediately available for comment", "declined to discuss", "calls went unanswered"), as are more double-quote characters. What the box shows and copies is exactly the text that was checked.
- A request the server cut off shows as an error. If the host stops a check before it finishes, the box now says so instead of showing the previous result, and the next click runs it again. A result that arrives after all replaces the error.
- A click while another check is still running now waits on it instead of using up one of the post's checks for nothing.
- Billing: replacing a background job that had already finished still records its cost.
- Screen readers now hear each part's result when it finishes ("Typos & grammar: 2 to fix.", "Editor's Notes: 5 notes."), the limit message, and the blanks warning. After a check, keyboard focus returns to Check this post (or to the limit message while the button is off).
- Block editor: after a save, the box now shows "The post has changed since this check." without a reload.
- The unsaved-changes note now says edits since your last save may not be included — an autosave of your own draft may already have them.
- The limit's "check it again at" time is rounded up to the minute, so it's never early.
- Clearer focus rings on the Settings and Costs tabs and the Save button, and darker small grey text (WCAG AA contrast).
- Upgrades: a site that last saved Settings before 1.16 and still held a copy of the old built-in typo prompt now follows the current one. A fresh install on a multisite subsite now gets the new-site defaults (Editor's Notes on).
- Uninstall on a multisite network now cleans up every site, API keys included.
2.1.0
Editor's Notes, most important first — with wording you can paste.
- Most important first. Notes now come in order of importance instead of story order: anything that would be an error if the story published as is (the story contradicting itself, a claim its own text doesn't support), then the questions most readers will ask, then a missing response or affected party, then sources and background. Each note still says which paragraph it's about, and the eight that make the cut are the eight most important.
- Suggested wording. Where a sentence fits, a note includes one or two in the story's style — its tense, how it names people, how it attributes — with a Copy button. Facts the story doesn't have are left as
[TK: …]blanks for the writer to fill in. Every suggested sentence is checked before it's shown: no quotation marks, no numbers the story doesn't contain, nothing like "declined to comment", nothing cut off; one that fails is dropped and its note kept. Verify and Source notes never carry wording. On by default; turn it off with Suggest wording writers can paste in Settings → Copydesk box. Wording adds about $0.0002 to a notes request. - A warning for blanks left in the post. While the saved post still has a
[TK: …]blank, the box says "This post still has a blank to fill in: [TK: vote count]." No request is sent for it. In the block editor it updates each time you save. - The typo check never flags a
[TK: …]blank. Because its built-in prompt changed, an unchanged post's first check after updating runs the typo check once instead of reusing the stored result. - The At a glance sidebar shows Editor's Notes as "On, with wording", "On" or "Off".
2.0.0
Copydesk is now a button. The automatic scan is gone: writers and editors click Check this post in the editor, and the result appears while they can still fix it. (On a busy news site, about one in eight published posts went out with an automatic scan's typo flags still standing.)
- One Copydesk box replaces the typo box and the Editor's Notes box. Check this post runs the typo check and Editor's Notes side by side, and each section fills in as soon as its part is ready — typos usually first. If one part fails, the other still shows.
- Grammar, clear-cut only. The typo check now also flags the wrong word form (its/it's, then/than, their/there), a missing or extra small word, subject–verb agreement, and wrong verb forms. Style is still out of scope, and so are how brands are styled and compound-word choices.
- Names are left alone. The check never changes a name to a different name — names change, and an AI's knowledge can be out of date. Only clear misspellings are flagged. Each request also carries today's date.
- A per-post check limit: 3 checks in any 20 minutes by default, shared by everyone on the post, adjustable in Settings (0 turns it off). Only checks that send a request count; the box says when the post can be checked again, and the button turns itself back on.
- Unchanged posts are free. Checking an unchanged post reuses its typo result; Editor's Notes always runs fresh.
- Anyone who can edit the post can check it — writers on their own drafts included (Rescan used to need editor rights).
- Effort applies to Editor's Notes only. Typo checks always use Low: thinking harder doesn't find more typos.
- Editor's Notes is on for new installs. An existing site keeps its choice; a site that never chose stays off.
- Removed: automatic scans and the Scan triggers settings, the Reports page, the Posts-list Issues column, the separate Rescan and Get Editor's Notes buttons, and Dismiss / Clear all error alerts (an error now clears with the next check). The old Reports address goes to Settings.
- Nothing is lost on upgrade. Stored typo results and notes show in the new box as the last check. Scheduled scans are cancelled; a background job that was still running is finished and billed as usual.
- Behind the scenes: each part is its own request, so a direct (Claude) request gets your server's full time limit; a backstop collects OpenAI results the page didn't wait for, for notes too; a part your server stops frees its lock at once. "No obvious typos detected." now reads "No typos or grammar errors found."
- The Costs page calls them typo checks: "Avg / typo check", "Typo checks (30d)", and "Typos & grammar" in the "Used for" column. Earlier automatic scans stay in the totals.
1.16.0
- New: Editor's Notes (opt-in, off by default). A separate post-editor box returns up to eight on-demand suggestions — missing context, open questions, other perspectives, facts to verify, and sources — each pinned to a numbered paragraph with its opening words. Notes never rewrite the post, are validated server-side (type, paragraph, length, duplicates) before anyone sees them, and warn when the post has changed since they were written. Settings gains an Editor's Notes panel (on/off + an editable prompt with Reset) and a sidebar status line.
- Costs page: typo scans and Editor's Notes are tracked separately — a "Used for" column in the by-model table, an Editor's Notes stat card, and Avg / scan now counts scans only. The cost ledger gains a
kindcolumn; the upgrade runs automatically, labels existing rows as scans, and is verified before it's recorded. - Fix: saving Settings without touching a prompt no longer turns it into a "customized" copy. Browsers send textarea line breaks as CRLF, so the saved text never matched the built-in prompt and the site stayed frozen on that copy; prompts are now normalized, and copies saved by earlier versions are recognized as the default again, so prompt improvements reach your site.
- New: Effort setting next to each provider's model — Auto, or any level the selected model accepts (None through Max), applied to both typo scans and Editor's Notes. Auto uses Low for scans and Medium for notes. Copydesk adds thinking headroom on top of Max output tokens and longer timeouts at higher levels, and turns on thinking for Claude Opus 4.6–4.8 and Sonnet 4.6 when you pick a level explicitly (those models don't think unless asked).
- **Model
This README is longer than the copy stored here. Read the rest on GitHub →