All experiments

Experiment 01 / Magimica / 2026

How fast can a serious product actually be built?

A constrained build tested whether AI could compress delivery without collapsing architecture, testing or production judgment.

Product strategyUXFull-stack engineeringAgent orchestration

01 / Hypothesis

Speed and seriousness are not opposites.

I believed a useful product could move from foundation to a production-shaped beta in days if I kept the architecture small, coached the system deliberately and held product coherence myself.

AI assistance with a human control boundary
You ownYour objectiveWhat matters to you stays editable at all times.
AI may helpSuggested missionsOptional ideas appear quietly in the background.
Nothing passes without youAccept · Edit · DeclineYou decide what belongs in the journey.
Only approved workYour focused planTrack actions and progress without losing sight of the goal.

BenefitLess planning friction. More sustained focus.

Risk controlAI cannot change an objective or add a mission without approval.

02 / Bet

Build the smallest complete loop.

Objective → missions → dependencies → schedule → completion. AI proposes and decomposes work, but a pure TypeScript engine owns points, ordering, capacity and conflicts.

  • One codebase, explicit module boundaries
  • Database migrations as schema history
  • AI output validated before persistence
  • Production deployed during discovery

03 / Division of labour

A super-capable system still needed a coach.

AI leverage / broad

Acceleration across the full application

Research, architecture exploration, mission decomposition, implementation, tests, review, debugging and copy all benefited from agent capability.

Human judgment / accountable

Three decisions stayed mine

The product concept; the design direction and generated-image feature; and the final QA, visual refinement and level of finish.

04 / Evidence

The compressed build, without smoothing the edges.

The product changed shape over five compressed days: first the model, then selective AI, then a working beta, production repair and finally hardening. The sequence matters more than commit volume.

Magimica planning screen showing a selected direction and actions to refine the plan or generate missions.
Production UI / human direction before mission generation
Magimica review screen showing generated missions, dependencies and scheduling controls.
Production UI / generated work made reviewable before scheduling
05build days reconstructed
07database changes
03test layers
02product languages
  1. 01Aug 19

    Foundation artifacts

    Schema artifacts establish users, objectives, missions, dependencies and domain events.

    Product history / data model
  2. 02Aug 20

    AI becomes selective

    Objective-image migrations add generated illustration state without coupling AI to the deterministic planning core.

    Product history / schema change
  3. 03Aug 21

    A working system arrives

    A usable beta connects the full planning loop; database coverage, secret scanning and authenticated browser checks follow the same day.

    Working beta / quality gates
  4. 04Aug 22

    Production answers back

    Localization, PWA support and production compilation arrive alongside CORS, metadata, session and mobile fixes.

    Production behavior / repair
  5. 05Aug 23

    Hardening becomes the work

    Security, form validation, deployment configuration, mobile navigation and Safari behavior dominate the final pass.

    Security / mobile / deployment

Source note / reconstructed from repository history and dated database changes.

Reality check / revised decision

Continuous automated testing added real pipeline cost during the build. My current counter-hypothesis is to make one coherent coding pass, run an independent agent audit, then add tests in a deliberate batch before release. Faster does not mean untested; it may mean sequencing rigor more efficiently.

Visit Magimica