1. Home
  2. Blog
  3. Artificial Intelligence (AI)

Let a WordPress AI Copilot Make Changes Only After You Approve Them

There are two kinds of AI assistant you can put inside WordPress. The first can only talk. It reads your question, writes an answer, and hands you a snippet to paste somewhere yourself. The second can act. It creates the draft, sets the categories, fixes the function it just found in your child theme.

The second kind is far more useful, and far more worrying. An assistant that can write to your site can also write the wrong thing to your site. The question is not whether an AI copilot should be able to make changes. It is who decides when a change actually happens.

Generatify, the DaftPlug AI plugin, answers that in a specific way. Its copilot can read your site freely, but every change it makes stops and waits for your confirmation, and that rule is enforced in the plugin rather than in the prompt. This guide covers what the WordPress AI copilot can read, what it can change, how the approval step works, and the limits that apply when you let it touch files.

What the copilot is

The copilot lives at Generatify > Chatbot. The documentation describes it as an assistant that knows your site, not a general chatbot that happens to live in WordPress. It can query your content, and it can make changes, with your approval, through a defined set of abilities.

It is a Pro feature and needs a connected provider. Under Generatify > Settings > Connectors you add your own key for ChatGPT, Claude, Gemini, or Kimi. You pay the provider directly, and the model selector in the prompt box lets you pick one or leave it on Auto. If the selector reads “No AI Model Available”, no connector is set up.

The prompt box has three controls worth knowing:

  • Add Context attaches images, PDFs, or CSVs for the assistant to work from.
  • Mention WordPress content points at a specific post or page, so the assistant works on exactly what you mean rather than searching for it.
  • Select AI Model chooses which connected provider answers.

What it can read without asking

Reading changes nothing, so these abilities run without a confirmation step:

AbilityReturns
generatify/get-site-infoWordPress and PHP versions, active theme, active plugins
generatify/search-postsPosts matching a query
generatify/list-recent-postsRecent posts, filterable by type and status
generatify/get-postOne post’s content, excerpt, author, terms, featured image, and meta
generatify/list-usersUsers, optionally with email addresses

This is what makes questions about your own site work. The docs give examples such as “which posts mention returns”, “what did I publish in March”, and “who wrote the most posts last year”. We covered that side of the assistant in our guide to asking your WordPress admin AI chatbot about your own site. This article is about the other side: what happens when it wants to write.

What it can change, and why every change stops

These abilities modify your site, and every one of them stops and asks first:

AbilityEffect
generatify/create-postCreates a post
generatify/update-postEdits an existing post
generatify/trash-postMoves a post to trash
generatify/set-post-termsSets categories and tags
generatify/set-featured-imageSets a featured image
generatify/generate-imageGenerates an image
generatify/moderate-commentApproves, unapproves, spams, or trashes a comment
generatify/update-plugin-settingChanges a Generatify setting

How the approval step works

According to the copilot guide, nothing is written when the assistant decides to write it. Instead, three things happen:

  1. It builds a preview. The change is prepared and held, not applied.
  2. It describes the change to you, in plain language, in the conversation.
  3. You confirm or you do not. Only on your confirmation does the change actually happen.

The important detail is where this is enforced. The docs are explicit: it is enforced in the plugin, not in the prompt. The write path returns a preview and a token instead of performing the action, and the action only runs when the pending change is applied. An assistant that decides to skip asking still cannot write anything.

That distinction matters more than it sounds. Many AI tools rely on a system prompt that says “always ask before acting”. Models do not always follow instructions. A rule in the prompt is a request. A rule in the write path is a guarantee.

You can also ask to see what is waiting at any point, and pending changes are listed back to you. If the copilot ever says it made a change and nothing happened, the troubleshooting guide has the answer: changes are held until confirmed. If you did not confirm, nothing was written.

Read the preview, not the summary

One line from the docs is worth putting on a sticky note: the description is generated, the preview is what will actually be written, and when they disagree, the preview is right. The plain-language summary is helpful for a quick read. The preview is the source of truth. Check it before you confirm, especially for edits to existing posts.

Developer abilities: working on the site itself

The copilot can also work on the site’s code. The docs describe this as the difference between an assistant that hands you a snippet and one that creates the child theme and fixes the broken function it just found.

AbilityEffect
generatify/list-filesLists files under wp-content
generatify/read-fileReads one file
generatify/write-fileCreates or overwrites a file
generatify/delete-fileDeletes a file
generatify/manage-pluginActivates or deactivates a plugin
generatify/get-optionReads an option
generatify/update-optionWrites an option

The same approval rule applies to the ones that write. On top of it, the documentation lists a set of hard limits.

The limits

Capability checks. These abilities require manage_options at minimum, plus WordPress’s own file-editing capabilities. They are checked against the logged-in user both when the preview is built and again when it is applied.

Location. Every path is resolved and must land inside wp-content. Symlinks cannot escape it, and neither can ... That keeps its file abilities away from wp-config.php and WordPress core, which live outside it.

File types. An allowlist of source and text extensions: php, js, mjs, jsx, ts, tsx, css, scss, json, txt, md, html, svg, xml, po, pot, yml, and yaml. Anything else is refused rather than guessed at. .htaccess is deliberately excluded, because one bad directive there takes down the admin you would need to undo it.

Site policy. DISALLOW_FILE_MODS and DISALLOW_FILE_EDIT are honored. A site that has locked file editing stays locked. If the copilot cannot edit files, these constants in wp-config.php are the first thing the troubleshooting guide tells you to check, followed by the allowlist and the path.

Reversibility. Every write and delete copies the previous file into uploads/generatify-backups first and returns the backup path. Nothing is removed outright.

Protected options. A fixed list stays out of reach: siteurl, home, active_plugins, template, stylesheet, admin_email, new_admin_email, users_can_register, default_role, cron, db_version, initial_db_version, and recently_activated. Those are the options that can lock you out, change who can register as what, or break the site’s address.

The limits are not a backup

The documentation is honest about this: these protections are real, and they are not a substitute for backups. An approved change to a theme file is still a change to a theme file. The advice is to work on staging when asking the assistant to modify code, and to confirm you have a working backup before approving anything structural.

The per-file copy in uploads/generatify-backups lets you undo a single edit. A full backup, such as one from Snapshotify, lets you undo an afternoon.

Where your conversations live

Copilot conversations are kept in your browser’s local storage, not in your database. The FAQ spells out what that means:

  • They do not appear in your database and do not follow you to another machine.
  • They are cleared when you clear site data.
  • Other administrators cannot see your conversations. Each admin has their own history in their own browser.

New Chat discards the current conversation. If the copilot “forgot” an earlier discussion, this is why.

As for what reaches the AI provider, the FAQ says only what the feature needs: for the copilot, the content it reads to answer you.

Getting good results

The copilot guide closes with four habits, and each one saves time or money.

Mention the content instead of describing it

Use the mention control to point at the exact post. Searching for “that post about shipping” wastes a round trip and sometimes finds the wrong one.

Ask for one change at a time

A request bundling five edits produces one confirmation you have to accept or reject as a whole. Separate requests are easier to review and easier to undo.

Read the preview, not the summary

Covered above, and worth repeating. The preview is what will be written.

Start a new chat when you change topic

Long conversations carry everything forward. That costs more per request and gives the assistant more chances to conflate two tasks. The troubleshooting guide names long copilot conversations as one of the two usual causes of surprise provider bills, because every message resends the whole conversation. DaftPlug cannot see or cap your provider usage, so set a spending limit in your provider’s dashboard. It is the only hard ceiling that exists.

A note on images

The copilot can generate an image and set it as a featured image, both with approval. Image generation needs a provider that offers an image model. The connectors docs note that ChatGPT is also the provider used for image generation, and the editor checks image capability first, so you see a specific message rather than a failure when only Claude or Kimi is connected. Generated images are, by the docs’ own description, the most expensive item by a wide margin.

A safe first session

If you have never let an assistant write to your site, start with low stakes:

  1. Connect a provider and set a spending limit in its dashboard.
  2. Ask a read-only question about your own content to confirm it sees your site.
  3. Ask for one draft post on a topic you know. Read the preview, confirm, then open the draft in the editor.
  4. Ask it to set categories and tags on that draft. Check the preview.
  5. Ask to see pending changes before confirming anything, to get used to the flow.
  6. For code work, move to staging. Take a backup, ask for one file change, confirm, and note the backup path it returns.

Free and Pro

The copilot is a Pro feature, as are content generation, AI meta descriptions, the connectors, and helpdesk conversation mode, because they call an AI provider with your key. Free Generatify features such as AI related posts and the toxicity checker run without one. Pro comes with the DaftPlug membership: one license unlocks Generatify along with Progressify, Lightify, Rankify, and Snapshotify, and plans differ only by the number of sites. See the pricing page.

Quick checklist

  1. Connect your own provider key and set a spending limit.
  2. Use mentions to point at exact posts.
  3. Ask for one change per request.
  4. Read the preview before confirming, not just the summary.
  5. Ask for pending changes if something seems not to have happened.
  6. Keep DISALLOW_FILE_EDIT on production if you never want file edits there.
  7. Do code changes on staging, with a full backup first.
  8. Start a new chat when the topic changes.

An assistant that can act is worth having only if you stay the one who decides. Generatify’s copilot is built so that the decision cannot be skipped, whatever the model says.