I want images, voice, or live answers

The meetup flyer came back from Imagine with “Analtyics Nite” in 72-point type, a photoreal stranger labeled as the VP of Sales, and a live-flavored caption that cited a blog post from 2019 as “this morning.” Eleven regenerations. The room still needed a readable title, a real headshot with permission, and a date someone had opened. Pixels, voice, and live answers are not a writing chat with extra sparkles. They are different surfaces with different failure modes.

This is Part 5 of Which AI product should I use? You want images, voice, or “live” answers. Stay out of the coding agent. Stay out of the inbox bot unless the asset is the job. Re-check product names (ChatGPT images, Gemini image tools, Grok Imagine, voice modes) the week you ship a flyer. A writing chat can describe a picture. That description is not the picture. Open the studio when the artifact is pixels, and inspect those pixels like you inspect a chart axis. If you would not ship a truncated axis, do not ship a misspelled title.

What you’ll learn

Why a still, a microphone, and a search-ish reply are three jobs
Which big-four doors to open for each, as of writing
A verify-before-you-post checklist (type, links, likeness, quota)
When a clean still beats a wobbly six-second clip
The Analtyics Nite failure, in a table you can steal

Three outputs, three doors

Pixels, voice, live answers, and short video as four different surfaces

Output	Try first (as of writing)	Verify
Still image	ChatGPT image tools, Gemini image tools, or Grok Imagine	Every letter, every face, no invented logos
Voice	The voice mode in the chat app you already use	You are not dictating secrets in a cafeteria (see Part 4)
Live / search-ish answer	Grok chat, or the browsing/live mode in ChatGPT / Gemini / Claude if offered	Open the link. Dates and names first
Short video	Imagine-style video or the vendor’s motion tool	Quota, likeness, and whether a still would have done

Claude, as of writing, is less of an image studio and more of a language product. Do not punish it for a job it does not sell. Gemini and ChatGPT both make stills; which one looks better depends on the brief and the week. Grok Imagine is a whole surface, not a chat afterthought. Details live in the Grok Imagine parts and the other brand series. This page is the hallway.

Verify before you post

Verify before you post: read every letter, open live citations, check likeness and brands, keep the still if video is garnish

Read every letter. Overlay real type if the model cannot spell Analytics. Eleven regenerations will not fix a bad brief.
Open every live citation. If you cannot click it, you cannot ship the claim.
Likeness and brands: no fake VP. No invented official logos. Use a real photo with permission or an illustration that is obviously not a person you employ.
Quota: a still that prints beats a six-second clip of the same misspelling.

Rule of thumb: If the artifact is pixels, inspect pixels. A confident caption does not spell the title.

A brief that survives a flyer

Still image, 16:9, meetup flyer background only.
Do not put any words on the image. We will overlay type ourselves.
No people faces. No logos. No watermarks.
Metaphor: a clean classroom table with printed charts, evening window.
If you cannot avoid letters, return a blank table and stop.

What that brief does: it stops Analtyics Nite at the source by refusing on-image type. You overlay “Analytics night” in real software. Voice is simpler: talk, then read the transcript before you paste it into Slack. Live answers need a second tab. Grok’s live flavor is a feature. It is not a footnote.

Worked example: Analtyics Nite

Check	Regeneration 11	Shipable
On-image title	Analtyics Nite	Type overlaid: Analytics night
Face	Invented VP	No person, or Maya’s real photo with OK
Caption source	“this morning” + 2019 URL	Link opened, date in the sentence
Motion	6s clip of the misspelling	Still only

The 11th regeneration was a budget problem disguised as taste. Change the brief. Overlay type. Rights still apply even when the picture is pretty. School and work honesty is a later safety series; the short version is: do not submit generated art as photography, and do not deepfake a colleague.

Voice and live, without mythology

Voice is hands-busy. It is not more truthful. A cafeteria is still a Part 4 problem. Live or browsing modes are a retrieval claim: the model looked somewhere. Your job is the same as a junior analyst with a search bar. Open it. If Claude, ChatGPT, or Gemini offers a live/web mode the week you read this, use it the same way. Do not memorize which brand “has search.” Verify the link.

Quota, voice transcripts, and the still that prints

Image and video meters are separate from “I already pay for chat” more often than people think. Grok’s Imagine surface is the cleanest example on this site: a studio with its own habits, not a second chat window. ChatGPT and Gemini also meter generations. Eleven regenerations of a misspelled title is how a team burns the week’s budget on a flyer nobody can hang. Write the brief. Overlay type. Stop.

Voice modes are a gift when your hands are full and a leak when the room is public. The transcript is still data. If you dictate a customer list while waiting for coffee, you have done Leah’s cafeteria move with a microphone. Save voice for chores that are already allowed in that account: outline a public blog post, rehearse a talk track, capture a grocery list. Then read the transcript. Voice models swallow names. “Maya” becomes “Maria” and you send the wrong thank-you.

Short video is garnish until the still is right. A six-second Ken Burns of a correct title can help a social post. A six-second Ken Burns of Analtyics Nite just advertises the typo. If you make motion, specify duration, subject, and camera in the brief. Do not ask for a feature film. The Grok Imagine tutorial walks that budget in more pages than this hallway allows.

Live answers without becoming a search engine

Live or browsing modes fail in boring ways: stale pages, wrong people with the same name, a PDF from 2019 dressed as this morning. Your check is a journalist’s check. Open it. If the claim is a number, find the table. If the claim is “the vendor announced X,” find the post. Grok’s access to X can be useful for “what are people yelling about.” It is a terrible way to close a $4,250 credit. Omar’s Sheet still wins. For analysis numbers, Practical AI already said the model is not a signed-off analyst. Same rule, different surface.

Common mistakes

Eleven regenerations instead of overlaying type.
A fake executive face because “we needed a person.”
Shipping a live sentence you did not click.
Using Imagine to write the thank-you from Part 1. Wrong door.
Video of a still that already failed the spelling check.

Rights in one uncomfortable page

Generated stills are easy to steal from and easy to over-claim. If the meetup is a school event, ask whether generated art is allowed on the flyer. If it is a client deliverable, say it is generated. Do not invent a photographer. Do not drop a brand mark you do not have rights to “because the model knew the shape.” AMS product posts composite official logos for identification; your flyer is not that article. When in doubt, leave the logo off and print the name in type you control.

Deepfakes of colleagues, students, or a VP who “could not make the shoot” are a hard no in this hallway. Illustration, objects, rooms, charts: yes. A photoreal person who does not exist, labeled as staff: no. Voice clones of a person who did not agree are the same no with a microphone. If a vendor offers a likeness tool, that is a legal conversation, not a Friday experiment. The later safety series will go longer. You do not need that series to reject the fake VP on the Analtyics Nite poster.

Stock-looking hands and extra fingers are a quality tell. They are also a reminder to inspect. Zoom. Print a proof. Have a second person read the title out loud. Dana’s four-app problem was comparison theater. The flyer problem is inspection theater that never happens. One still, one overlay, one proof. Then you can talk about video. If the second person still reads Analtyics, the overlay failed, not the model. Fix the type file. Do not regenerate the room.

How to practice this week

One still. No on-image words. Overlay a three-word title yourself. Open one live citation on a question you actually need. Then go to Part 6 and pick defaults. Grok’s image path in depth: the Grok series.

Quick recap

Pixels, voice, and live are different surfaces.
Read the letters. Open the link. No fake VP.
A still with real type beats a clip of Analtyics Nite.

Sources

Grok Imagine
ChatGPT and Gemini image help pages (names and models move)
AMS: Grok series (Imagine tutorial)
Part 4: privacy

Keep going

Same lessons in your feed

Short diagrams and hooks on Instagram, X, and Facebook.

@analyticsmadesimple Follow @AnalyticsSimply Follow Analytics Made Simple Follow