Create an App
Take a raw app idea to a validated, cleanly architected first version that ships on a sustainable cadence — without over-building for scale you have not earned, or writing a prototype that calcifies into a rewrite. This journey runs ten phases: it asks before every decision and records each choice in docs/ so you can stop and resume across sessions. The early phases often produce no code at all; once an experiment greenlights building, the engineering phases build the real thing well.
Core Principle
Validate fast while keeping options open: make the expensive-to-reverse decisions early and defer everything cheap. Spend scarce early effort on the few things that compound — the riskiest assumption, the dependency boundaries, the domain language — and aggressively avoid waste everywhere else.
This skill sequences the phases, asks the questions, and records the decisions; the constituent skills carry the method. Invoke them rather than improvising their frameworks.
Journey Map
Operating Rules
- Resume first. Before anything else, read
docs/CREATE-APP-PLAN.md and every artifact in the Journey Map. If the tracker exists, summarize the journey state in 3-5 lines and ask which phase to enter. Done when the user has confirmed an entry point. A journey with a tracker is resumed, never restarted.
- Intake on first run only. No tracker: run the Intake below, then create
docs/CREATE-APP-PLAN.md with every phase statused pending | in-progress | awaiting-evidence | done | deferred: reason | skipped: reason. Done when the tracker exists and the user has confirmed the phase plan.
- Phase entry. Announce: what the phase does, the decision it forces, the artifact it produces, rough effort. Offer proceed / skip / defer — phases marked GATE may be deferred, never skipped. Mark the phase
in-progress on proceed. Done when the user chose.
- Skill invocation and fallback. Load the phase's skill and use it: each phase's Invoke line names the skill by slug — use that skill to run the phase. If it is not available, offer:
npx skills add wondelai/skills/<slug> --global. If the user declines, run the phase from its Brief — the minimum viable method. State which mode you are in.
- In-phase decisions. Ask every question under "Decide with the user" — with concrete options and your recommendation. Record the choice in the tracker's Key Decisions. A decision made silently is a defect.
- Phase exit. Present the draft artifact content for sign-off before writing. On approval: write or extend the docs/ files, update the tracker (status, Key Decisions, Next Actions). Done when the files are written and the phase row shows
done.
- Artifact discipline. Read before writing; create a file only if missing, otherwise extend — add or update your sections, preserve everyone else's. Files are UPPERCASE in
docs/. Every recommendation lands as a checkbox or a table row with owner and priority. See references/artifact-templates.md when creating a docs/ file for the first time — create it from the full skeleton (all section headings), then fill the sections your phase names.
- The agent prepares and processes; only the human talks to customers and test users. Draft sprint scripts and interview guides and score the results, then pause with status
awaiting-evidence until the user returns with notes. The ios-hig-design phase applies only when shipping a native iOS app; skip it otherwise at intake.
Intake
Ask these before creating the tracker:
- What is the app idea in one sentence, and what is the riskiest belief behind it? (Gates Phase 1; if the job the app is hired for is fuzzy, add the jobs-to-be-done optional phase.)
- Is the underlying idea already validated with behavioral evidence, or still a hunch? (Unvalidated and high-stakes makes Phase 1 mandatory; if paying customers already exist you may be in the wrong journey — improve-app or grow-app.)
- What is the stage — weekend prototype, funded build toward launch, or a team forming around it? (A prototype leans on Phases 1-3 and 10; a forming team may add inspired-product.)
- Will you ship a native iOS client? (Yes keeps Phase 8; no skips it.)
- Do you have customers to talk to, or must you find them? (Gates the human evidence loop in Phases 1-2 and optional mom-test / continuous-discovery.)
- What load do you expect in the first months — daily active users and actions per day? (Feeds Phase 7 capacity math and stops premature scaling.)
- Solo build or team, and who owns shipping cadence? (Gates Phase 9 shaping.)
Phase-skip heuristics: skip Phase 2 when the core flow is already tested or trivial; skip Phase 8 unless shipping native iOS; skip Phase 7 only when load is obviously tiny — but still record the estimate; for a weekend prototype, run Phases 3-6 as one pass but never skip the Dependency Rule.
Then create docs/CREATE-APP-PLAN.md from the template with every phase statused, marking Phase 8 iOS-conditional. Done when docs/CREATE-APP-PLAN.md exists with every phase statused and the user has confirmed the plan.
Phases
Phase 1 — Validate the riskiest assumption (lean-startup)
Purpose: Prove the belief that would kill the app if false — before writing production code.
Brief (fallback): Plan Build-Measure-Learn backward: what must you learn, how will you know, what is the smallest build that finds out. Rank leap-of-faith assumptions by what is fatal, not what is easy to test. An MVP is a learning vehicle, often crude (Dropbox validated with a video, not a sync engine). Climb the validation ladder — signups weak, paid deposit strong, active usage strongest; aim for behavioral level 4-5, never "would you use this?" opinion.
Invoke: Use the lean-startup skill with the idea and, if CUSTOMER.md exists, the job statement. Ask for ranked leap-of-faith assumptions, the single smallest experiment that falsifies the riskiest, and a pre-committed behavioral success threshold.
Decide with the user: Which assumption is fatal if false? Which experiment type — smoke test (demand), concierge (value), Wizard of Oz (automation)? What go/pivot threshold, fixed before running? Only the user runs it with real people; then pause awaiting-evidence.
Artifact: Create docs/PRODUCT.md with ## Vision and ## MVP Definition; create docs/EXPERIMENTS.md with ## Experiment Cards (EXP-001 with hypothesis, metric, decision rule) and ## Experiment Backlog. Update the tracker.
Done when: the fatal assumption is named, EXP-001 has a pre-committed threshold, PRODUCT.md states the MVP scope, and the user has the experiment to run.
Phase 2 — De-risk the core experience (design-sprint)
Purpose: Settle what the core flow should be with a tested facade, before committing engineering time.
Brief (fallback): Five days: Map the problem Monday, Sketch Tuesday, Decide Wednesday, Prototype Thursday, Test with five real users Friday. Output is a high-fidelity facade plus evidence, not code. The riskiest moment is a stranger's first ten minutes — test whether they understand it and finish the core task with no explanation. Five users surface the patterns; you need patterns, not significance. Never explain the prototype; watch where they get stuck.
Invoke: Use the design-sprint skill with the MVP scope from PRODUCT.md. Ask for a Monday map and How Might We reframes, a winning concept storyboard, and a five-act interview script with a note-taking grid (checkmark / cross / tilde per participant). Record the approved script under the sprint card in docs/EXPERIMENTS.md when pausing, so it survives the session break.
Decide with the user (on return): Does the tested flow work as-is, need a reshape, or expose a demand problem that loops back to Phase 1? Which confusions become fixes?
Artifact: Create docs/DESIGN.md with ## Design Direction; extend docs/EXPERIMENTS.md ## Experiment Cards with the sprint test and its verdict. Update the tracker.
Done when: DESIGN.md records the direction, the sprint result and verdict are in EXPERIMENTS.md, and the user chose proceed / reshape / loop.
Phase 3 — Draw the boundaries (clean-architecture)
Purpose: Isolate business rules so the database, framework, and providers stay swappable details.
Brief (fallback): Source code dependencies point inward — frameworks toward use cases toward entities. Business rules must not import the framework or ORM. Use cases define interfaces (e.g. InvoiceRepository); infrastructure implements them; controllers translate an HTTP request into a plain request object and receive a plain response — no framework type crosses the boundary. Diagnostic: can you test the business rules with no database, web server, or framework running? Four circles are typical, not sacred — draw boundaries only at points of volatility.
Invoke: Use the clean-architecture skill with the validated MVP scope from PRODUCT.md. Ask for the layer map, the core use case with its request/response models, and the repository interfaces it depends on.
Decide with the user: Where are the real volatility boundaries (database, third-party APIs, delivery mechanism)? Full four layers, or collapse adapters and frameworks for a small app?
Artifact: Create docs/ARCHITECTURE.md with ## System Context and ## Layer Map & Dependency Rule (layers, what depends on what, a violations table). Update the tracker.
Done when: ARCHITECTURE.md names the layers, the Dependency Rule direction is explicit, and the core business rule is testable with no infrastructure.
Phase 4 — Model the domain (domain-driven-design)
Purpose: Make the code speak the business so a domain expert could read it and spot what is wrong.
Brief (fallback): The model is the code. Build a Ubiquitous Language — name things after domain concepts, not technical roles (an InvoiceDraft.finalize() beats a DataProcessor.process()); hard-to-name is a design signal. Bounded contexts: a word means one thing inside a boundary ("Customer" in billing need not equal "Customer" in support). Aggregates: a cluster with one root that enforces invariants — keep them small, reference other aggregates by ID. Avoid the anemic model; push behavior into entities and value objects.
Invoke: Use the domain-driven-design skill with ARCHITECTURE.md's layer map. Ask for the ubiquitous language, the bounded contexts and context map, and the core aggregate with the invariants its root enforces.
Decide with the user: What is the core domain (your competitive edge) versus generic subdomains (auth, email, payments) to buy or use open source? Where does the same word legitimately mean different things across contexts?
Artifact: Extend docs/ARCHITECTURE.md: add ## Bounded Contexts & Context Map and ## Domain Glossary (Ubiquitous Language) (term | meaning | code name). Update the tracker.
Done when: the glossary names the core concepts, contexts and relationships are mapped, and the repository interface sits in the domain layer per Phase 3.
Phase 5 — Write code to be read (clean-code)
Purpose: Keep every function readable and tested as the codebase grows.
Brief (fallback): Code is read far more than written. Small functions that do one thing at a single level of abstraction; intention-revealing names; no flag arguments (a smell — the function does two things); commands separate from queries; a well-named extracted function beats a comment. Tests are first-class — dirty tests are worse than none. Write them F.I.R.S.T. (Fast, Independent, Repeatable, Self-validating, Timely) with behavior-based names like shouldRejectNegativeTotal.
Invoke: Use the clean-code skill with the core module from Phases 3-4. Ask for a score-and-fix review against the rules and clean unit tests (arrange-act-assert, descriptive names, a builder helper) for the core use case.
Decide with the user: What quality bar gates a commit — function size, name clarity, no swallowed exceptions? Which happy-path and failure cases must the first tests cover?
Artifact: Create docs/TESTING.md with ## Test Strategy (pyramid, tooling, what green gates) and ## Safety Net Map (module | pinned behaviors | test files | gaps). Update the tracker.
Done when: TESTING.md states the strategy, the core use case has passing behavior-named tests that need no database, and the commit quality bar is recorded.
Phase 6 — Install the meta-disciplines (pragmatic-programmer)
Purpose: Adopt the habits that keep the codebase easy to change over years.
Brief (fallback): Fire a tracer bullet — one thin real slice through every layer (UI to use case to database and back), kept, to surface integration bugs on day two. DRY applies to knowledge only, not coincidental similarity. Orthogonality: changing the database must not touch the UI. Reversibility: wrap third-party SDKs behind your own interfaces so Stripe or a model provider swaps without touching business logic. Broken windows: fix the first hack or board it up with a tracked ticket — never leave silent rot.
Invoke: Use the pragmatic-programmer skill with ARCHITECTURE.md and the core use case. Ask for the thinnest end-to-end tracer-bullet design and an adapter interface for each third-party dependency.
Decide with the user: What is the tracer-bullet slice? Is the broken-windows policy zero-tolerance with tracked tickets? Which dependencies get adapter interfaces now versus later?
Artifact: Extend docs/TESTING.md (## CI Gates) and docs/TECH-DEBT.md (## Debt Budget & Broken-Windows Policy and ## Adopted Conventions). Update the tracker.
Done when: the tracer bullet is specified end-to-end, adapter interfaces are named for each provider, and the broken-windows policy is written down.
Phase 7 — Size it honestly (system-design)
Purpose: Estimate the real load and prove what you do NOT need to build yet.
Brief (fallback): Start from requirements, not solutions. QPS = daily active users x actions/day / 86,400, peak 2-5x average; storage = records/day x record size x retention. For a few hundred users the math almost always says one well-indexed database plus a cache. Scale in order: vertical first, then read replicas, then cache aside, shard last. Reach for a queue to absorb spikes and decouple slow work (e.g. photo OCR) from the request path only when an estimate or real bottleneck justifies it.
Invoke: Use the system-design skill with the load numbers from intake. Ask for a back-of-the-envelope QPS and storage estimate and an explicit list of techniques you do NOT need yet.
Decide with the user: Given the estimate, which scaling do you deliberately defer? Which real bottleneck, if any, justifies a cache or queue now?
Artifact: Extend docs/ARCHITECTURE.md: add ## Data & Storage Decisions and ## Decision Log (date | decision | why | alternatives rejected), recording each "not yet" as a decision. Update the tracker.
Done when: the capacity estimate is recorded, the deferred scaling techniques are listed explicitly, and any cache or queue is justified by a number.
Phase 8 — Make it feel native (ios-hig-design) — iOS only
Purpose: Make an iPhone client feel native — the basis of iOS users' trust. Run only when shipping a native iOS client (decided at intake); skip otherwise.
Brief (fallback): Three pillars — clarity, deference, depth. Respect safe areas (Dynamic Island, home indicator); every touch target at least 44x44 pt; semantic colors (Color(.label), Color(.systemBackground)) so Dark Mode is automatic; semantic text styles for Dynamic Type; native navigation — tab bars for primary destinations, NavigationStack for drill-down, sheets for focused tasks, never a hamburger menu. Accessibility is first-class: a label on every control, and VoiceOver can complete every task.
Invoke: Use the ios-hig-design skill with the app's key screens. Ask for a HIG review of safe areas, 44pt targets, semantic colors and type, native navigation, and accessibility labels.
Decide with the user: Which screens are in scope for the first review? Any deliberate platform deviations, and are they justified?
Artifact: Extend docs/DESIGN.md: add ## Components (component | decision | status) and ## UX Audit Findings (issue | heuristic | severity 0-4 | fix | status). Update the tracker.
Done when: each in-scope screen is reviewed, findings are logged with severity, and accessibility gaps are captured as fixes.
Phase 9 — Ship on a cadence (37signals-way)
Purpose: Ship a focused v1 on fixed time that fights feature creep.
Brief (fallback): Build half a product, not a half-assed one. Shape work before betting it: rough enough for design freedom, solved enough to remove big unknowns, bounded by an appetite ("this is worth two weeks") — not an estimate. Fix the time and cut scope to fit, never the reverse. Bet shaped pitches into fixed cycles; a circuit breaker kills anything unfinished at the deadline. Opinionated software: every preference offered is a decision refused — pick sensible defaults; the default answer to a feature is a respectful "not now".
Invoke: Use the 37signals-way skill with the MVP scope from PRODUCT.md. Ask for the next feature shaped into a pitch (problem, appetite, breadboard, rabbit holes, no-gos) and a v1 cut list.
Decide with the user: What is the appetite for the next feature? Which v1 features get cut, and which user-facing preferences become opinionated defaults?
Artifact: Extend docs/PRODUCT.md (## Outcome Roadmap) and docs/STRATEGY.md (## No-List — what we explicitly will not do). Update the tracker.
Done when: the next feature has a shaped pitch with an appetite, the v1 cut list is in STRATEGY.md's No-List, and the roadmap reflects fixed-time / flexible-scope.
Phase 10 — Treat complexity as the enemy (software-design-philosophy)
Purpose: Hold complexity down across every module — the lens over all the other phases.
Brief (fallback): Complexity is anything about structure that makes a system hard to understand and change. Prefer deep modules — powerful functionality behind a simple interface — over shallow ones; judge a module by functionality divided by interface complexity, not line count. This corrects Clean Code's "small" and DDD's many concepts tipping into classitis (swarms of one-method shallow classes). Strategic over tactical: invest a steady 10-20% extra on design, because early shortcuts compound exactly as the team and codebase grow.
Invoke: Use the software-design-philosophy skill with the modules from Phases 3-6. Ask for a deep-vs-shallow evaluation flagging shallow classes, pass-through methods, and information leaking across boundaries, plus where to consolidate.
Decide with the user: Which shallow modules should merge into deeper ones? Where is tactical shortcutting accruing debt worth the strategic 10-20% now?
Artifact: Extend docs/TECH-DEBT.md (## Smell Inventory: smell | location | refactoring | status) and docs/ARCHITECTURE.md (## Decision Log). Update the tracker.
Done when: shallow modules and leaks are logged in the Smell Inventory, consolidation decisions are recorded, and the strategic-vs-tactical stance is agreed.
Optional Phases
Optional phases follow the same operating rules — load and use each listed skill exactly as a core phase would; insert where the Add-when condition first becomes true.
Common Mistakes
Completing the Journey
Exit checklist:
Close the tracker: every phase done or skipped: reason, Key Decisions captured, and Next Actions carried into the artifacts (not left in the tracker). Forward routing:
- When the idea underneath the app turns out to be unvalidated, or is really a commercial-model question, continue with the
create-business skill.
- When the app grows past its first architecture and the structure itself becomes the hard problem, continue with the
design-code-architecture skill.