AI can generate interfaces in seconds. But does it really understand what we're trying to solve?
1. The ultimate paradox of the AI era
There's a growing belief that AI can write code so quickly that we no longer need wireframes, sketches or much planning. Just describe what you want and ship it!
We're certainly building things faster. But are we necessarily building the right things?
AI can generate thousands of lines of code based on a misunderstanding of the original requirement. And it can do that remarkably quickly.
1.1 Jevons Paradox of software development
Ever heard of Jevons Paradox? It describes how improvements in efficiency can actually lead to increased consumption rather than less. We're seeing something similar with software development.
AI has made producing software significantly cheaper. We're now generating more prototypes, screens, features and applications than ever. But producing more software doesn't necessarily mean we're solving more problems.
As the cost of implementation comes down, knowing what to build becomes even more valuable. The bottleneck was never typing curly braces. It has always been understanding the problem and figuring out how to translate it into a meaningful experience.
1.2 High velocity demands a tighter steering wheel
Think about riding a bicycle versus flying a jet fighter. The faster you move, the more important it becomes to get your direction right. Even a small deviation can have significant consequences.
AI development is no different. A poorly understood requirement can turn into a fairly substantial implementation within minutes. This is where concept sketching comes in. It gives us an opportunity to think through the solution before asking AI to build it.
Spending a few minutes exploring layouts and information hierarchy can save hours of unnecessary rework. This is precisely why concept sketching matters more than ever.
2. Machine-readable concept sketches
Ever heard of machine-readable concept sketches or wireframes? The interesting part is that concept sketches no longer have to be hand-drawn or created in bulky design tools like Figma, only to be exported as images. Those images can be harder for AI to interpret, reference or incorporate into other documents.
Let's look at them differently. The terminal isn't new, and ASCII diagrams aren't new either. But have we ever looked at them from the perspective of building a text-based wireframe?
Instead of thinking of the terminal/ASCII purely as a place to execute commands, what if we looked at them as a way to describe interfaces? Think of an ASCII wireframe that communicates layout, hierarchy, content and relationships between components.
It gives AI a structured representation to work with before generating an actual interface. They're lightning-fast to generate and super lightweight too. And more importantly, they give us something tangible to look at, discuss and refine without having to build the entire thing first.
+--------------------------------------------------------------------------------------------+
| [DAIRY OPS] *Dashboard* Orders Fleet Hubs | [ [R] Search orders, routes... ] [JD] |
+--------------------------------------------------------------------------------------------+
[! ALERT: Chilling Vat 3 exceeded 7.2C (Sector 4 Hub) [Resolve] [X] ]
+--------------------+ +--------------------+ +--------------------+ +--------------------+
| ORDERS TODAY | | VOLUME DISPATCHED | | ACTIVE ROUTES | | DELAYED DROPS |
| 1,420 (+8%) [^] | | 14,250 LTR (+12%) | | 24 / 26 (*) Active | | 2 Routes (X) Alert |
+--------------------+ +--------------------+ +--------------------+ +--------------------+
[ *All Orders* ] ( Pending ) ( Dispatched ) ( Exceptions (2) )
================
[ [R] Search route, ID... ] [ Hub: All [v] ] [ Status: Any [v] ] [ Today [CAL] ] < Reset >
[*2 Selected*] | [ Assign Driver ] [ Export | v ] [ + Create Route ]
+-----------------------------------------------------------+ +-----------------------------+
| [X] ROUTE | DRIVER | CRATES | STATUS | ACTION | | LIVE DISPATCH FEED |
|------------|-----------|--------|------------|------------| +-----------------------------+
| [X] RT-014 | R. Sharma | 450 | (*) Active | [View] | | (OK) 06:15 Chilled (3.8C) |
| [X] RT-015 | V. Kumar | 310 | (X) Delay | [View] | | | |
| [ ] RT-016 | A. Patel | 520 | (OK) Ready | [View] | | o 06:45 Loaded DL-04-12 |
| [ ] RT-017 | M. Singh | 280 | (-) Paused | [Edit] | | | |
| [ ] RT-018 | S. Verma | 390 | (*) Active | [View] | | * 07:10 In transit [Now] |
+-----------------------------------------------------------+ | |
< Prev *[1]* [2] [3] ... [14] Next > (1-20 of 275) | BATCH SYNC PROGRESS |
| [============......] 68% |
+-----------------------------+
----------------------------------------------------------------------------------------------
TOKENS: (*) Active (OK) Ready (X) Delay/Alert (-) Paused [R] Search [CAL] Date [^] Trend
[X] Checked [ ] Unchecked [v] Dropdown < > Ghost Link [ Label ] Primary CTA
3. Here's how I build these with AI
If you ask an AI to "draw an ASCII dashboard" from scratch, it will invent a slightly different visual language every time—broken borders, inconsistent buttons, vague boxes, and clutter mixed into the drawing.
To stop the AI from improvising, I treat ASCII concept sketching the same way a front-end team treats a design system: with a static library of reusable UI primitives stored as plain .txt files.
Step 1: A Static UI Primitive Catalogue (.txt files)
I've created small, static component .txt files, organised into 7 core UI categories covering 46 standardised primitives. Every .txt file defines two things:
- Explicit visual states (
[DEFAULT],[FILLED],[FOCUS],[ERROR],[DISABLED]) so you can inspect happy paths and failure states inside the same lightweight sketch. - A Token Vocabulary Reference at the bottom, enforcing one strict rule: zero metadata in the sketch canvas (no story IDs, author notes, or dates cluttering the layout—just pure slots, layout, and interactive tokens).
1. Inputs — Form controls, pickers, toggles, and uploaders (inputs.txt)
1. TEXT INPUT [DEFAULT] [ Enter work email... ]
[FILLED] [ alex.smith@acme.corp ]
[ERROR] [! Invalid email format ]
5. SEARCH INPUT [FILLED] [ [R] Route-7B Milk Run [X] ]
6. DROPDOWN [FILLED] [ Central Chilling Hub - Sector 4 [v] ]
7. MULTI-SELECT [FILLED] [ (Driver: Rajesh [x]) (Van: DL-04) |.. [v] ]
8. CHECK / RADIO [FILLED] [X] Require temp logging (*) Priority (2h)
10. TOGGLE SWITCH [FILLED] (--O) ON - SMS Dispatch Notifications
11. DATE PICKER [FILLED] [ 2026-11-20 -> 2026-11-27 ] [CAL]
14. FILE UPLOAD + - - - - - - - - - - - - - - - - - - - - - - - - - - -+
' [^] Drag & drop manifest, or [Browse] '
+ - - - - - - - - - - - - - - - - - - - - - - - - - - -+
TOKEN VOCABULARY REFERENCE:
[^] Upload/Inc [v] Caret [R] Search [X] Clear [CAL] Date [X] Check (*) Radio (--O) Toggle
2. Actions & Triggers (actions.txt)
1. PRIMARY BUTTON [DEFAULT] [ Submit Order ] [FOCUS] |* Submit Order *| 2. SECONDARY BUTTON [DEFAULT] [ Cancel ] [FOCUS] | Cancel | 3. GHOST / TEXT BUTTON [DEFAULT] < View Details > [FOCUS] < *View Details* > 4. ICON-ONLY BUTTON [DEFAULT] [+] [X] [?] [R] 5. SPLIT BUTTON [DEFAULT] [ Export | v ] [FILLED] [ Export CSV | v ] 6. SEGMENTED CONTROL [DEFAULT] [ *Day* | Week | Month | Year ] TOKEN VOCABULARY REFERENCE: [ Label ] Primary CTA [ Label ] Secondary CTA < Label > Ghost/Link [+]/[X] Icons
3. Display, Status & Metadata (displays.txt)
1. STATUS PILL (*) Active (OK) Verified (X) Delayed (-) Paused
2. COUNTER BADGE [ Inbox (14) ] [ Alerts (!3) ]
3. AVATAR / USER CHIP [JD] John Doe <Admin> [!] Unverified
4. KPI STAT CARD +-----------------------+ +-----------------------+
| TOTAL DISPATCH TODAY | | TOTAL DISPATCH TODAY |
| 14,250 LTR (+12%) [^] | | [! Metric Error ] |
+-----------------------+ +-----------------------+
5. TOOLTIP / POPOVER [?] -> "Temperature must stay below 4C"
6. SECTION DIVIDER -------------------- [ OR ] --------------------
TOKEN VOCABULARY REFERENCE:
(*) Active (OK) Success (X) Alert (-) Muted [JD] Avatar [^]/[v] Trend
4. Data Display & Collections (collections.txt)
1. DATA TABLE | ROUTE ID [^] | DRIVER | VEHICLE | STATUS | ACTION |
| RT-014 | R. Sharma | DL-04-12 | (*) Active | [View] |
| RT-015 | V. Kumar | DL-04-88 | (X) Error | [View] |
2. LIST ROW [*] Order #1042 - 50 Crates - 06:30 AM [>]
3. TIMELINE FEED (OK) 06:15 AM Milk chilling confirmed
|
o 06:45 AM Loaded DL-04-12
|
* 07:10 AM In transit to Hub 2 [Current]
4. TREE VIEW [-] Sector 4 Hub |-- Unit A (OK) |-- Unit B (Alarm)
TOKEN VOCABULARY REFERENCE:
[^]/[v] Sort [>] Row Drilldown o/|/* Timeline Nodes [+]/[-] Tree Nodes
5. Overlays, Dialogs & Feedback (overlays.txt)
1. MODAL DIALOG +======================================================+
| Confirm Dispatch Manifest [X] |
+------------------------------------------------------+
| Assign 450 crates to Vehicle DL-04-12? |
| [ Cancel ] [ Confirm Now ] |
+======================================================+
2. DRAWER SHEET |+ Route #RT-014 Details [X] |+ Driver: R. Sharma |+ [Save] |
3. ALERT BANNER [! ALERT: Chilling Vat 3 exceeded 7.2C [Resolve] [X] ]
4. TOAST SNACKBAR [OK: Manifest saved successfully [Undo] [X] ]
5. PROGRESS BAR [==================................] 50% (Syncing)
6. EMPTY STATE CARD +-- ( ! ) No Active Dispatches [+ Create Route] ------+
TOKEN VOCABULARY REFERENCE:
+====+ Modal Frame |+--+ Drawer Edge [X] Dismiss [! ..] Alert [===...] Progress
6. Navigation & Wayfinding (navigation.txt)
1. TOP NAVBAR [DAIRY OPS] *Dashboard* Orders Fleet | [R Search...] [JD] 2. SIDEBAR NAVIGATION | [=] DASHBOARD | [*] DISPATCH RUNS (3) | [o] FLEET 3. TABS [ *All Orders* ] ( Pending ) ( Dispatched ) ( Exceptions ) 4. BREADCRUMBS Hubs > Sector 4 > *Chilling Unit 2* 5. PAGINATION < Prev *[1]* [2] [3] ... [14] Next > (1-20 of 275) 6. WIZARD STEPPER (OK 1. Tenant Info) ---> (*2. Route Setup*) ---> ( 3. Review ) TOKEN VOCABULARY REFERENCE: *Label* Active Item === Active Underline > Crumb Separator < > Pager Links
7. Containers, Layout & Grouping (containers.txt)
1. CARD / PANEL +------------------------------------------------------+
| Route Summary - Assigned: DL-04-12 - Crates: 450 |
| [ Edit Route ] |
+------------------------------------------------------+
2. ACCORDION [>] Driver Contacts (Collapsed) / [v] Expanded View
3. FILTER BAR [ [R] Search... ] [ Hub: All [v] ] [ Status: Any [v] ] < Reset >
4. TOOLBAR [*4 Selected*] | [ Export CSV ] [ Assign ] | [+ New Route ]
TOKEN VOCABULARY REFERENCE:
+---+ Card Border [>]/[v] Accordion |...| Divider < Reset > Reset Filters
Step 2: How the Folder Structure is Organised
Instead of burying UI rules inside sprawling prose prompts, I keep every static primitive file inside a predictable workspace folder—cleanly separated into reusable Components/ and the generated feature sketches sitting side by side at the root:
HyperPrototypes/ ├── Components/ │ ├── Inputs/ │ │ └── inputs.txt # 14 form & input primitives + state variants │ ├── Actions/ │ │ └── actions.txt # 6 buttons, split actions & segmented controls │ ├── Displays/ │ │ └── displays.txt # 6 status pills, user chips, KPI cards & dividers │ ├── Collections/ │ │ └── collections.txt # 4 data tables, list rows, timelines & trees │ ├── Overlays/ │ │ └── overlays.txt # 6 modals, drawers, alerts, toasts & progress bars │ ├── Navigation/ │ │ └── navigation.txt # 6 navbars, sidebars, tabs, pagers & steppers │ └── Containers/ │ └── containers.txt # 4 cards, accordions, filter bars & toolbars │ ├── daily-dispatch-dashboard.txt # Pure ASCII concept sketch (zero metadata) ├── daily-dispatch-dashboard.html # Standalone review hub (sketch + discussion notes) ├── route-assignment-drawer.txt └── route-assignment-drawer.html
Keeping the primitive library as plain .txt files gives AI agents an instant advantage:
- Zero parsing overhead: There is no framework syntax, CSS cascade, or build tool for the model to reason about. It reads raw text and immediately sees the exact bounding characters and spacing.
- Tiny context footprint: Each category file is only 60–150 lines long. Pointing an agent at all seven files consumes a fraction of the context window while locking down the entire visual vocabulary.
- Deterministic tokens: Because
[R],[CAL],(X),(*), and[ Label ]are defined once in the static files, every new screen sketch across the project looks like it belongs to the same product family.
Step 3: Prompting AI to Assemble a Larger Concept Sketch (Like a Dashboard)
Once the primitive .txt files are sitting in Components/, prompting AI to assemble a full screen is just a matter of snapping blocks together—like Lego bricks. A reliable dashboard prompt has three parts:
- Point to the static primitive
.txtfiles as the only building blocks: Tell the AI not to invent new ASCII symbols or box styles, and specify which states to use ([DEFAULT],[FILLED], or[ERROR]if you want to stress-test edge cases). - Set the canvas rules: Lock the character width (94–96 column compact width), require zero metadata inside the drawing, and ask for a compact shorthand token legend at the bottom.
- Describe the screen hierarchy & domain slots: Outline the top-to-bottom regions (header, alerts, KPI strip, filter/action bar, main table, and side detail panel) in plain English.
Compose a pure ASCII concept sketch for a "Daily Order Processing & Dispatch Dashboard"
(Operations Manager view) and save it to:
HyperPrototypes/daily-dispatch-dashboard.txt
CONSTRAINTS:
- Compose strictly using the UI primitives and token vocabulary defined in:
- HyperPrototypes/Components/Navigation/navigation.txt
- HyperPrototypes/Components/Overlays/overlays.txt
- HyperPrototypes/Components/Displays/displays.txt
- HyperPrototypes/Components/Containers/containers.txt
- HyperPrototypes/Components/Collections/collections.txt
- HyperPrototypes/Components/Actions/actions.txt
- HyperPrototypes/Components/Inputs/inputs.txt
- Strict 94–96 column compact width, ASCII-safe characters only.
- Pure visual canvas: ZERO metadata (no story titles, author, version, or dates)
inside the wireframe drawing. Put a short token legend at the bottom.
LAYOUT HIERARCHY:
1. Top Navbar (nav:top-navbar) with brand [DAIRY OPS], active *Dashboard* link,
search input, and [JD] user chip.
2. Alert Banner (overlay:alert-banner) warning that Chilling Vat 3 exceeded 7.2C.
3. 4-column KPI Strip (display:kpi-card): Orders Today, Volume Dispatched,
Active Routes, and Delayed Drops.
4. Order Status Tabs (nav:tabs), Filter Bar (container:filter-bar), and
Batch Action Toolbar (container:action-toolbar) showing 2 selected rows.
5. Two-column main workspace:
- Left (61 cols): Dispatch Routes Data Table (collection:data-table) with
multi-select checkboxes and Pagination Control (nav:pagination) underneath.
- Right (31 cols): Card Panel (container:card-panel) containing a Live Dispatch
Timeline Feed (collection:timeline-feed) and Batch Sync Progress Bar.
Step 4: The Assembled Dashboard Output
Because the AI reads the static .txt primitives directly before drawing, it doesn't guess widths or invent weird borders. Every section snaps cleanly into place:
+--------------------------------------------------------------------------------------------+
| [DAIRY OPS] *Dashboard* Orders Fleet Hubs | [ [R] Search orders, routes... ] [JD] |
+--------------------------------------------------------------------------------------------+
[! ALERT: Chilling Vat 3 exceeded 7.2C (Sector 4 Hub) [Resolve] [X] ]
+--------------------+ +--------------------+ +--------------------+ +--------------------+
| ORDERS TODAY | | VOLUME DISPATCHED | | ACTIVE ROUTES | | DELAYED DROPS |
| 1,420 (+8%) [^] | | 14,250 LTR (+12%) | | 24 / 26 (*) Active | | 2 Routes (X) Alert |
+--------------------+ +--------------------+ +--------------------+ +--------------------+
[ *All Orders* ] ( Pending ) ( Dispatched ) ( Exceptions (2) )
================
[ [R] Search route, ID... ] [ Hub: All [v] ] [ Status: Any [v] ] [ Today [CAL] ] < Reset >
[*2 Selected*] | [ Assign Driver ] [ Export | v ] [ + Create Route ]
+-----------------------------------------------------------+ +-----------------------------+
| [X] ROUTE | DRIVER | CRATES | STATUS | ACTION | | LIVE DISPATCH FEED |
|------------|-----------|--------|------------|------------| +-----------------------------+
| [X] RT-014 | R. Sharma | 450 | (*) Active | [View] | | (OK) 06:15 Chilled (3.8C) |
| [X] RT-015 | V. Kumar | 310 | (X) Delay | [View] | | | |
| [ ] RT-016 | A. Patel | 520 | (OK) Ready | [View] | | o 06:45 Loaded DL-04-12 |
| [ ] RT-017 | M. Singh | 280 | (-) Paused | [Edit] | | | |
| [ ] RT-018 | S. Verma | 390 | (*) Active | [View] | | * 07:10 In transit [Now] |
+-----------------------------------------------------------+ | |
< Prev *[1]* [2] [3] ... [14] Next > (1-20 of 275) | BATCH SYNC PROGRESS |
| [============......] 68% |
+-----------------------------+
----------------------------------------------------------------------------------------------
TOKENS: (*) Active (OK) Ready (X) Delay/Alert (-) Paused [R] Search [CAL] Date [^] Trend
[X] Checked [ ] Unchecked [v] Dropdown < > Ghost Link [ Label ] Primary CTA
Notice what is happening in that single block of text: every interactive affordance is unmistakable ([ *All Orders* ] shows the active tab, [*2 Selected*] explains why [ Assign Driver ] is enabled, [X] shows checked rows, and (X) Delay surfaces the exception right next to the alert banner). There is zero visual fluff, zero CSS distraction, and zero ambiguity about information hierarchy.
Step 5: Iterating at the Speed of Thought (Before Writing Any Code)
Notice that alongside daily-dispatch-dashboard.txt, my workflow uses a companion daily-dispatch-dashboard.html review hub that places the raw <pre> ASCII canvas on one side and a simple discussion card on the other (holding the user stories, open questions, and review notes that we intentionally kept out of the pure sketch).
⚡ Tighter Steering Wheel
When the hierarchy, density, and edge cases finally feel right in plain ASCII, then we earn the right to move fast with code—because we finally know we're flying in the right direction.