Descript AI Prompts: Templates for Underlord, Editing and Video

TL;DR: Descript's AI splits into two groups: prompt-driven features (Underlord, Generate Media, Edit with AI, Generate Audio and Custom Avatars) and one-click features with no prompt anywhere, like Studio Sound and Filler Word removal. This guide gives copy-paste templates for the first group and states plainly which features never take a prompt.

Which Descript features actually take a prompt?

Six do, verified against Descript's features page and help center on 2 September 2026: Underlord, the AI co-editor chat; Generate Media, for images and video; Edit with AI, for retouching part of an existing image; Generate Audio, for music and sound effects; Custom Avatars, for a presenter's face; and AI Video Maker, which can start from a prompt, a script, or uploaded footage.

Everything else in the AI panel is either one click with no input field, or takes a script rather than a prompt. That second category matters as much as the first, because it's where most confusion happens. Regenerate and its underlying re-recording tooling take the exact words you want spoken, character for character, up to a limit. That's editing, not prompting, even though both use natural language.

The table further down lists every AI feature Descript documents and which bucket it falls into. For now, the six above are what the rest of this guide covers.

How do you write a good Descript prompt?

Descript's own help center gives a five-part structure, and it holds up well in testing: action, context, tone, format and constraints. You don't need all five in every prompt: start with what matters most, then refine.

• Action: what job needs doing
• Context: who it's for, where it'll be used
• Tone: how it should feel
• Format: what the output should look like
• Constraints: rules it has to follow

Here's Descript's own worked example, built for a clip-splitting job in Underlord:

Turn this 60-minute podcast into 5-6 short clips for Instagram and TikTok.
Keep them punchy and attention-grabbing, in vertical format with bold
captions. Each clip should be 30-60 seconds and focus on one key insight.

Action is the opening clause. Context is the platform names. Tone is "punchy and attention-grabbing". Format is vertical with bold captions. Constraints are the duration range and the one-insight rule. None of it is exotic; it's a paragraph a colleague could act on, which is exactly the bar a prompt needs to clear.

Two habits improve results further. First, attach real material as context instead of describing it: type @ in the Underlord prompt box, or right-click a script passage, scene, layer or timestamp and choose Add to Underlord. A prompt that points at the actual clip needs to guess less than one that describes it. Second, iterate in small steps rather than one giant instruction. Descript's own examples of that back-and-forth include "Make it a little more casual", "Keep the captions, remove the music.", and "Now add captions." after an earlier step — Underlord is explicitly designed for that kind of follow-up, and lets you undo a step that didn't land.

What can you ask Underlord to do?

Underlord is Descript's AI co-editor, currently in beta, and it works as a chat panel inside a project rather than a fixed input field. Descript's own documentation frames the range as open-ended: try nearly any editing goal and see what happens. A few categories with real, documented example prompts:

Captions

Add viral, social media-style captions to my video. Emphasize the active
spoken word in my brand color yellow #ffd02b.

Update this clip so the captions follow this format: Manrope font, 100pt,
extra bold, centered white text with a 30pt black shadow. Active word in
yellow with a 10pt black outline.

Clips

Can you split this into 5 clips that are max 1 minute long?

Find me clips for social media that are between 60–90 seconds long and
reformat them for vertical aspect ratio.

Animations and transitions

Add zooms to moments of emphasis.

Add a simple fade-out effect to the end of this video.

A few things worth knowing before you rely on Underlord for anything client-facing. Each prompt spends AI credits on its own reasoning, separately from whatever tool it decides to run, and Descript documents this as non-deterministic — the same prompt can cost a different amount on different runs. You also choose which model powers Underlord from a model selector, with a mix of Claude, GPT and Gemini options on paid plans; the choice changes reasoning quality but not the prompt syntax above it. And if a result goes sideways, Underlord edits can be undone or rolled back to a previous version, which is the safety net that makes iterating on a prompt low-risk.

How do you prompt AI Video Maker and Generate Media?

These are two different entry points into the same underlying models, and the distinction matters for where you start.

AI Video Maker builds a whole video (script, voiceover, visuals, and an optional avatar) from one of three starting points: a prompt ("Create a video about the origins of the Welsh Corgi"), a rough pasted script, or uploaded footage. It's reachable from the Home tab, and separately, through a Descript-published GPT inside ChatGPT, where you just describe your idea and the finished project lands in your Drive. The one catch worth knowing: pasting a script written by a chatbot often carries stage directions and bracketed notes that trip up the narration step, so strip those before pasting.

Generate Media is the more granular tool, for producing a single image or video clip inside an existing project. Add an empty scene, click Generate, and type a prompt:

Generate images showing teams collaborating in modern office settings for
a product demo video about our new project management software.
Professional but approachable, not too corporate or staged. Cinematic-style
realism, with wide-angle shots, natural lighting, 16:9 format. Show diverse
teams, avoid clichés like high-fives or fist bumps.

After the prompt, you separately choose a model, aspect ratio, and duration or batch size: parameters sit outside the text field, not inside it. One documented limitation catches people every time: switching from the Image tab to the Video tab starts a fresh prompt, and generation history is tracked separately per tab, so copy your prompt text somewhere else before you switch.

Edit with AI solves a narrower job: changing part of an existing image rather than generating a new one. Select the pencil tool, draw around the area, and type a short prompt describing the change, such as adding an object onto an empty part of the frame. The eraser tool needs no prompt at all; it just removes the selected area and fills it from the surrounding content.

How do you prompt a custom avatar or AI speech?

An avatar's face and its voice are prompted through completely different mechanisms, and treating them the same is the second most common mistake after the Regenerate one above.

The face takes a genuine text prompt. Assign a speaker, choose Create a custom avatar, and in the Text prompt tab, describe a realistic, adult, non-celebrity, human-like figure, or click Inspire for a suggested prompt if you're stuck. Descript generates three options; you can keep iterating on the same prompt, and previous generations stay visible while the modal is open. Best-practice guidance is specific: face forward, mouth unobstructed, eyes clearly visible, soft even lighting, no animals or objects in a prompt or reference photo, since none of that animates reliably.

The voice is not prompted the same way. A stock AI speaker is simply selected from a list. A cloned custom speaker is built from your own recorded audio and a consent step, not from a text description of how you want it to sound. The closest thing to a voice "prompt" is the script itself, written or edited in Write mode, which gives you a place to draft dialogue before any audio exists.

How do you turn a prompt into a reusable Underlord template?

Once a prompt earns its keep, Descript lets you save it rather than retyping it per project. Open the template gallery from the Home tab, click New template, and enter your prompt — following the same action/context/tone/format/constraints structure — plus a title, a description under 280 characters, and optional tags.

Two limitations are worth knowing before you build a library of these. Published templates can't be edited or deleted; changing one means writing a new template from scratch, since Descript documents that editing existing templates is "coming in a future update." And templates you create are private to your account only — sharing one with a Drive or publishing it publicly isn't available yet. That second limit is the reason a lot of teams end up keeping the actual prompt text somewhere outside Descript as well, since a private-only template can't be handed to a new hire the way a shared prompt library entry can.

Which Descript AI features never take a prompt?

Roughly half the toolbox, and this table is the reference to bookmark. Verified against Descript's features page and help center on 2 September 2026.

Two of these deserve a second look because the name misleads. Automatic Multicam sounds like it might take direction, but it runs entirely on its own, choosing whichever camera has the active speaker. And Regenerate takes text — the exact replacement words — which is why it lands in the "no" column despite looking like a prompt field at a glance. Understanding the difference between a genuine prompt template and a literal-text edit field saves a wasted round trip on every one of these.

What goes wrong when you prompt Descript, and how do you fix it?

Descript's own troubleshooting guidance for Generate Media and AI Video Maker covers four recurring failures, and each has a specific fix rather than a generic retry:

"Custom prompt required" or a rejected image. A generated visual shows a checkered placeholder because the underlying provider's safety filter rejected the prompt. Select the placeholder, choose Add Media, and generate again with a more neutral, descriptive prompt, or upload your own file instead.

Content flagged by a safety checker. Either the prompt or a reference image tripped the model provider's filters. Remove sensitive or policy-adjacent language, switch to plain descriptive wording, or drop the reference image and generate from text alone.

An internal server error. Usually means the selected model couldn't process the prompt or reference image at all. Wait a few minutes and retry, simplify the wording, remove reference images, or switch to a different available model.

Video generation blocked by content restrictions. If you're using AI Video Maker specifically, Descript's own fix is to simplify the script first: strip section headers, speaker labels, bracketed notes, and extra spacing, and translate non-English scripts to English before generating, since only supported TTS languages work with that tool.

The pattern across all four: when Descript rejects a prompt, the fix is almost always to make the language plainer and more literal, not more elaborate. That's the opposite instinct from a lot of image-prompt advice, which pushes toward more descriptive detail — worth remembering the next time a Descript generation stalls.

Where this leaves Descript prompts, and us

Descript earned the "prompt" framing honestly — Underlord, Generate Media, Edit with AI and Generate Audio are real, documented free-text surfaces, not marketing gloss over a fixed feature set. But half the app is still one click by design, and the fastest way to waste a session is assuming otherwise: typing a mood board into Regenerate, or hunting for a prompt field on Automatic Multicam that was never going to exist.

Prompt Architects generates and stores the prompt text itself — the paragraph you'd paste into Underlord, a Generate Media panel, or a saved custom template. It renders no video, retouches no images, and clones no voices; there's no Descript integration here, and there isn't meant to be one. What it's for is keeping the version of a prompt that actually worked, structured around the same action/context/tone/format/constraints shape Descript's own docs recommend, so the fifth clip of the week isn't rewritten from a half-remembered draft. The free plan covers 5 prompt enhancements per day, forever, per the FAQ page at prompt-architects.com, checked 3 September 2026.

For the underlying skill rather than this one tool, what prompt engineering actually is covers the fundamentals, and the prompt hygiene checklist is worth a pass before any prompt you're about to reuse across projects. If avatars specifically are the part you're building around, HeyGen's avatar and script templates cover a tool that draws the script-versus-prompt line even more sharply, and ElevenLabs' voice and direction templates cover the performance-direction vocabulary Descript's own voice tools don't expose. If you want the rest of the free generators we've built in this same shape, the index is here.