What the steps mean

While a turn runs, rows stack up in the chat. Each one is Pretty narrating a piece of its own work: “Running Snapshot…” while it happens, then a plain-language label when it finishes. Click any row to expand it and see the detail behind it: what it was given, what came back. Collapsed, the row is a progress marker; you do not have to read any of it.

Labels are written for you, not taken from the internal tool names. The two tables below cover every label you can see.

Steps that read or research

Nothing in your deck changes while these run.

The read and research steps, as they appear in a turn.

Step label	What it does	What to expect
Reading skill	Loads one of your skills so the agent can follow its instructions.	Fast. Usually the first row after you run a `/` command.
Searching docs	Looks up how to build something in Pretty’s own slide API.	Fast, and frequent before a PowerPoint step.
Find icon	Searches the icon library by meaning, not filename.	A handful of candidates; the agent picks and recolours to your master.
Find logo	Finds official company logos by image search, then a vision model picks the best.	Takes a few seconds. Accepts a variant such as symbol only or wordmark. See Logos.
Find map	Fetches a country, region, bloc or administrative-subdivision map.	Comes back as native editable freeform shapes, one per region, never a flat picture.
Researching	Live web research on your topic.	The slowest read step. Returns prose plus sources, and the agent adds a “Sources:” footer to slides built from it.
Reading page	Reads one named URL and pulls out the parts relevant to your prompt.	Use it by pasting a link in your message.
Searching layouts	Ranks your slide pool by structure and content, then re-ranks by looking at thumbnails.	Only appears if your workspace has a pool uploaded.
Reading file	Opens an attachment or a working file: text, DOCX, images, PDFs.	Images and PDFs come back as something the model can look at.
Extracting deck text	Dumps your deck’s text into a temporary working file so it can be searched.	Runs once, then later steps search that file instead of re-reading the deck.
Searching your deck (this may take a while)	A separate agent answers a question about the slides already in your deck.	Slow by design, and the label says so. It does not read speaker notes. See Editing existing slides.

Steps that touch your deck

These run inside the add-in. Only PowerPoint writes to your slides.

Step rows for PowerPoint, Snapshot, Reviewing design, Suggesting layouts, Asking you, Updating file, Running command, Planning steps and Summarizing — The steps that run in the add-in. Only PowerPoint carries a slide line.

Step label	What it does	What to expect
PowerPoint	Runs the agent’s code against the presentation you have open.	The one that changes your deck. Expands into a line per slide (Inserted slide, Edited slide, Moved slide), and each line jumps you to that slide.
Snapshot	Renders one to four slides to a picture so the agent can see its own work.	Two to four slides come back as a single 2×2 grid. Nothing is changed.
Reviewing design	A vision reviewer critiques the slide that was just changed.	Often followed by a second PowerPoint step fixing what it found.
Suggesting layouts	Shows you candidate layouts from your slide pool to choose from.	Titled Choose a slide layout, with a note field (“Optional: note for the AI…”). The turn waits for you.
Asking you	A clarifying question, two to four choices plus an Other box.	Nothing moves until you answer.
Updating file	Replaces exact text in a working file.	Fails rather than guess when the text it is looking for appears more than once.
Running command	Works over the chat’s working files: listing, searching, converting, zipping.	Shows a plain-language summary such as “Listing your files” instead of the raw command.
Planning steps	Writes a visible task plan, usually one task per slide.	Tasks flip from pending to in progress to completed as it goes. Good early signal of scope.
Summarizing	Compresses the conversation to reclaim context.	Fires on long chats. See Context and compaction.
Stopping agent	Stops a subagent that is still running.
Checking agents	Checks on subagents running in the background.

PowerPoint is the only step that writes to your deck, and it rolls back on failure. A step that errors leaves the slide as it was, not half-changed. If a change you expected never appeared, look for a failed PowerPoint row rather than assuming the agent ignored you.

Controls on a step row

Expanded failed PowerPoint step showing a tool issue, the error text and the code that ran — A failed PowerPoint step, expanded: the error it hit, the code it ran, and a copy button on each.

Expand a row and you get the buttons that are useful for that step.

Control	What it is for
Hints	Not an error. A geometry or overflow notice Pretty picked up and worked around.
Copy error	Copies the failure text. Paste it into a support email.
Copy hints	Copies the hints block.
Copy output	Copies what the step returned.
”Step 3/7”	This step is part of a longer script the agent is working through.
”Output truncated — Full output saved to …/tool-output.txt”	The result was too long to show. The full version is in the chat’s working files, and the agent can still read it from there.

Chatting with Pretty

Where the step rows sit, and what a whole turn looks like.

Troubleshooting

What to do when a step fails or the result is wrong.

Get started

Customize

Chat

Building slides

Assets

Reference

Support

What the steps mean

Steps that read or research

Steps that touch your deck

Controls on a step row

Chatting with Pretty

Troubleshooting

​Steps that read or research

​Steps that touch your deck

​Controls on a step row

​Related

Chatting with Pretty

Troubleshooting

Steps that read or research

Steps that touch your deck

Controls on a step row

Related