Case study2026Designed and built it: research, rules, content, app

Pencil Down

A trainer for CEED, a design entrance exam, that drills deciding on a clock you can't see. See it, decide, commit, move on.

1,494real questions from 49 past papers
11puzzle generators whose answers can't be wrong
< 1 dayresearch to live
Dailyin use since, and changed by that use
The Part A runner in Hard mode: a generated mirror-image puzzle, with the time-budget line reading Budget, hidden, 01:00.
Hard mode. The clock is running, and you can't see it.
01

The exam

CEED punishes hesitation more than ignorance.

CEED is the entrance exam for a master's in design at the IITs. Part A, the objective paper, is only a gate. Part B, the drawing paper, carries 75% of the score and is marked on ideas, not polish.

Marks leak in the same places: rereading a figure, guessing on a multiple-select question, four minutes on one puzzle, redrawing a first idea. Someone preparing for CEED doesn't need more questions. They need to practise deciding.

Part A is a gate. Clear the band, then win on ideas in Part B.
02

The decision

I cut the feature that would demo best.

My own research pointed to an AI critic that grades photos of Part B sketches. I left it out. The app runs no AI and grades no drawings; it trains the decision.

Score only what's honest.

Chose
Seconds to commit, number of ideas, and whether each idea obeys the rules.
Gave up
An AI critic of the drawing.
Because
CEED pays for ideas. A critic of line quality pulls practice toward polish.

Judge your own drawing.

Chose
A self-rating against the question's own printed criteria, shown only as a trend.
Gave up
A verdict on the drawing itself.
Because
A trainer is only useful if its feedback can be trusted every time.
03

The behaviour

Every screen takes away a crutch.

Hard mode is the default. Each Part A question gets a budget you can't see: 60 seconds for single-choice, 90 for multiple-select, 120 for numerical. At zero the answer is saved as it stands and the question goes to a revisit list.

Part B shows the prompt for two minutes, then hides it, so you draw from memory. An idea passes three checks before you commit it: it obeys every rule, it isn't your first obvious idea, and you can draw it in the time left.

No pause.

Chose
A clock that lives on the server. Closing the tab doesn't stop it; quitting marks the rest skipped.
Gave up
A pause button.
Because
The exam doesn't pause either.
04

The material

A trainer for snap decisions can't have a wrong key.

Learning to trust your first call only works if the key is right. Every answer shows its source: official key, worked out, or generated. A solved answer is never promoted to official.

Where an answer was in doubt, I ruled on each one: 4 of my worked answers corrected, 6 questions with no single fair answer dropped, 11 with a contested official key kept out of practice. A worked explanation ships only if it reaches the official answer.

1,494

Real questions from 49 past papers of CEED and UCEED, its undergraduate twin, 2010 to 2026, each one solved and checked.

  1. 01FetchPast papers and official keys
  2. 02Crop and readEvery question cut out, its text read
  3. 03JoinEach question matched to its key
  4. 04Solve and checkDisputes settled, arguable ones dropped
  5. 05ServeOnly what passed, with where its answer came from
05

The generators

Endless puzzles, and every answer checked twice.

Real papers run out; visual puzzles don't have to. I built 11 generators for the types CEED repeats: paper folding, mirror images, cube nets, rotations, views of stacked cubes.

Each generator builds a puzzle forward from a seed, a random starting number; a separate solver checks it in reverse. Wrong options are near-misses (a mirror instead of a rotation, one cell off), so you have to see the answer rather than eliminate your way to it.

  • fold-punch-unfoldMCQ · NAT
  • rotation-2dMCQ · MSQ
  • mirror-water-imageMCQ
  • matrix-completionMCQ
  • transparent-overlayMCQ
  • cube-netMCQ · MSQ
  • rotation-3dMCQ
  • cubes-from-viewsNAT · MCQ
  • iso-to-orthoMCQ
  • ortho-to-isoMCQ
  • visible-facesNAT · MCQ
  • 1,000+seeds per generator, built one way and solved the other. They agree every time.

All 11. One seed each, drawn by the app's own code. MCQ: single choice · MSQ: multiple select · NAT: a number.

Eight seconds to look. Then it's gone.

A generated puzzle under the app's rules: the figure hides, the clock runs unseen, and a commit is final.

Live · try it
01 / 12 · Paper folding · seed 6
--:--Budget · hidden 01:00Hard

A square sheet of paper is folded as shown and then punched through all layers in the last step. A punch made across a folded edge opens into a single hole. Which option shows the sheet when it is fully unfolded?

8 s to look · 60 s to decide

Look · 8 s
Choose one

+3


right +3 · wrong −0.5 · blank 0

06

Shipped

Live in under a day. Used every day since.

It's for one student, whose exam is on 17 January 2027, so every day spent building was a day not practising. I wrote the spec, with pass-or-fail acceptance criteria, before any code, then built in the order each step feeds the next: content, generators, app, ship, QA.

  1. Research

    Three reports: the exam, timed thinking, every Part B prompt since 2010

  2. Spec

    Principles, modes, scoring, acceptance criteria

  3. Live

    Deployed on push to pencil.nesh.studio, behind a login

  4. Stable

    Every criterion checked: 36 pass, 2 fixed, 1 can't pass (a 2022 prompt printed only in the answer booklet)

  5. Daily use

    Each session feeds the next change: the habit tracker, 802 more worked explanations, the disputed-key review

The Today screen: 63 days to CEED, a seven-day habit grid, a running drill that kept its clock, and the next session.
Today. Days to the exam, the habit grid, and the next session. The clock didn't stop while you were away.
07

Also in this work

Part B prompt bank

115 real prompts from 17 CEED years, plus 312 written to the same pattern: Indian everyday settings, about one in three with a twist, banned topics checked by script.

In the app
Knowledge cards

215 design-knowledge cards on spaced repetition (each returns just before you'd forget it), every sourced image with its licence recorded.

In the app
Progress

Mock history against the cutoff band, seconds to commit per Part B question, and a 7-day habit tracker that only shows what was done, never what was missed.

In the app
Design rules, enforced

One accent per screen, one large number per screen, time drawn as a dimension line, as on a technical drawing. Checked by script on 160 screen captures, phone and laptop, light and dark.

Checked in QA
Next.jsSQLite (node:sqlite)Tailwind v4 tokensRadixts-fsrsRechartsPythonPyMuPDFTesseractVitestPlaywrightForgejo CIClaude Code