From one scenario to a finished video: the XCONDA production map in 7 steps

English · XCONDA Hub

AI video production EP.00 — the full map. Start here if this is your first time.

Hello, this is XCONDA. Cubeberry's AI video starts with a scenario, is produced in XCONDA, gets its sound in Suno, and is finished in Premiere Pro. This post shows the whole sequence as one map, with an explanation for each step. The detailed method for every step continues in the series below.

Who this is for: people making AI video for the first time, people who have used the tools but struggle to make cuts connect, and teams that need to share one production order.

How to read: if you are new, start here; when you get stuck on a step, jump to the detailed post for that step. The series keeps going, so this post doubles as a table of contents.

The flow at a glance

Scenario → references → text storyboard → storyboard and blocking → video generation → sound → edit and finish

The area outlined in yellow on the map is 04 storyboard and blocking, because it is the link between preparation and video work. The outputs of the earlier steps gather there, and the key images confirmed there become the basis for the prompts in 05. If you dislike the result at 05, you fix the problem area and generate again — that is where the flow turns into a loop.

Why four preparation steps before generating video?

Give a video AI nothing but text and you get a different face, different clothes and a different composition every time. Stitch the cuts together and it looks like a different actor performed each one. The preparation steps exist to lock that wobble down before any video is made.

The easiest analogy is our CEO's time directing dramas at KBS. On set you confirm the actors, build the set and check the order and framing of cuts on a storyboard before rolling the camera — because fixing things on paper costs almost nothing, while fixing them after the shoot means shooting again. AI video also spends credits every time you roll. Checking a badly framed video and starting over is the same as reshooting the entire scene.

The key here is the division of labour between text and images. Each does what the other cannot.

However precisely you describe a face or an outfit in words, it never conveys what an image does, and conversely the order and timing of cuts are hard to convey with images. So characters and spaces are handled by images, and flow and movement by text.

Step by step

Each step is laid out as input → deliverable → check before moving on. The check is what you verify yourself before going on to the next step.

01. Write the scenario

• Input: idea, proposal
• Deliverable: a locked scenario
• Check before moving on: are the characters and locations decided?

Everything starts here. Without characters and locations you cannot build the references for the next step. Rather than naming an emotion ("sad"), write it as something visible — "hides the broken shard behind her back" — so it is easier to break into cuts later.

02. Get the reference images

• Tools: Claude, the XCONDA FLEXBOARD skill, GPT Image, XCONDA Turn
• Input: the scenario
• Deliverable: a face image per character (character sheet), location images (background sheets), and cut images made in Turn where needed
• Check before moving on: does every character have a face image, do you have the location images, and did you build the hard-to-shoot cuts in Turn?

If you already have references, use them as they are alongside the scenario. If not, the FLEXBOARD skill reads the scenario and writes the prompts that make a character sheet and a background sheet. Copy those prompts into GPT Image to make the images, then feed the images back in.

Why do it this way? Generate the background separately for each cut and the windows, furniture and props move from cut to cut. On a real set you build the set once and only move the camera inside it — the same idea: fix the space first, then change the angle. Method B is especially useful when you want to place a complex composition yourself, because text alone rarely produces it.

03. Text storyboard and FLEXBOARD prompts

• Tools: Claude, the XCONDA FLEXBOARD skill
• Input: the scenario and the reference images from 02
• Deliverable: a text storyboard (length, camera, action, dialogue and sound effects per cut) and the prompts for FLEXBOARD
• Check before moving on: does every cut state a start action and an end action?

A text storyboard is the scene written out in words first. At this stage you also check that the cut lengths add up, that no single cut is crammed with action, that cuts do not jump awkwardly, and that the number of characters is right. Where to download and install the skills is covered lower down.

04. Storyboard and blocking video

• Tools: XCONDA FLEXBOARD, XCONDA Blocking Board
• Input: the FLEXBOARD prompts from 03, and the cut images made in Turn at 02 (the hard-to-shoot cuts)
• Deliverable: a key image per cut (storyboard images) and a blocking puppet video
• Check before moving on: is each cut's key frame the composition you intended, and does the puppet video's movement match your intent?

FLEXBOARD is a storyboard tool that splits the scenario into cuts, generates an image for each cut and lays them out on one screen. It holds up to 36 cuts on a screen, and you can leave the good cuts alone and fix only the ones with problems. The click-by-click walkthrough is written up in the FLEXBOARD guide.

Blocking Board is currently open as a free beta (up to the stage before video generation). Paste a scene prompt and press Create to compose the 3D space and characters, add cameras with Add camera, and send to Seedance from the Output tab. See the Blocking Board beta notice for the full introduction.

The puppet does not need to look human. Position, direction and speed are enough. Just remember to attach the puppet video at 05 video generation.

05. Seedance prompts and video generation

• Tools: Claude, the Seedance prompt skill, the prompt editor, XCONDA
• Input: the FLEXBOARD prompts and text storyboard, reference images, the storyboard, the blocking puppet video from 04, and audio references if needed
• Deliverable: a two-column prompt editor with Korean and English side by side, and the video generated in XCONDA
• How to revise: tell the editor in conversation what to fix, it updates the prompt, and you generate again. Repeat the loop until you are happy.

The Seedance prompt skill builds prompts for Seedance 2.5 from the key images. It organises them with a four-director system and a ten-block structure, and lets you revise with Korean and English side by side. Revise in Korean and the English is rewritten to match, so you never have to edit the English separately. The skill is being prepared for download from the XCONDA hub.

If any one of the four is missing, the AI fills that part in as it likes. Drop the reference images, for example, and even if the prompt says "the woman in @image1" it cannot tell who that is, so the face changes from cut to cut. Drop the storyboard and the composition changes every time even with the same prompt; drop the puppet video and characters walk in a different direction from the prompt, or several characters' paths cross each other.

On the generation screen, choose Reference (Ref-to-video) mode and type @ in the prompt to say which reference is which. Aspect ratio, resolution, length and High Quality are set on the Options tab, and Seedance 2.5 can produce up to 30 seconds at a time.

While you iterate on revisions and regenerations, the recently added Draft feature helps: check the result at 480p first and turn only the drafts you like into 1080p finals. Note that the final is generated fresh from the same assets as the draft, so fine details can differ slightly, and drafts are valid for 7 days after generation. See the Draft notice for details.

06. Sound

• Tools: Suno
• Input: the mood and length of the video finished at 05
• Deliverable: sound files for the video
• Check before moving on: does it match the video's length and mood?

Sound is made after the video is finished, using the video's mood and length as the brief. The detailed method for this step continues in the series.

07. Combine video and sound

• Tools: Premiere Pro
• Input: the clips generated at 05 and the sound made at 06
• Deliverable: the finished video
• Check before moving on: cut connections, levels and subtitles

The last step. Before joining the cuts, keep the references from 02 and the text storyboard from 03 beside you and run through once more to confirm that costumes, props and backgrounds do not change from cut to cut.

One-page summary per step

So you can check as you read the map, the input, deliverable and pre-handover checks are gathered in one table.

Five common mistakes

• Generating video first: check with images and you only redo the one wrong cut. With video you redo the whole thing.
• Adding several face images of the same character: the faces blend. One per character is the rule.
• Putting only some of the four inputs into generation: whatever is missing — references, storyboard, prompt or puppet video — the AI fills in as it likes.
• Expecting faces and expressions from the puppet video: puppets carry position, direction and speed only. Faces belong to the reference images.
• Regenerating cuts that already look good: fix only the cuts with problems.

Series contents

AI video production continues one step per post. Posts already published are linked, and the rest will be released in order.

Good companions: the post on why to storyboard before generating video, and the production case study that locks cut consistency with Turn and FLEXBOARD.

If you are starting now

XCONDA is running a Seedance 2.5 promotion for the FLEXBOARD launch.

• Korea: 30% usage discount plus 10% price discount, up to 40%
• US: up to 30% usage discount
• Period: until 11:00 on 6 November (KST)

👉 Go to XCONDA (https://www.xconda.ai/)

For adoption or partnership enquiries, contact us below.

• Email: cubeberry@cuberry.kr
• Phone: +82-70-8095-2302 (weekdays 10:00–17:00 KST)

This map is current as of 21 October 2026. Content may change as the tools are updated. Thank you for reading a long post today!

Read this in Korean