ProdaZhang/aidd

---

name: aidd description: AIDD (AI-assisted Design→Dev) — turn a design discussion into a platform-agnostic handoff package another AI can build from, gated by deterministic checkers. Use for designing any config/rule/UI-heavy system from scratch — game systems, app & SaaS features (plans, quotas, permissions, feature flags, forms, onboarding), tools — or to advance / sync the project spine. One skill: it reads the spine, judges progress, and routes you to the right phase playbook in references/. Domain-neutral; ships with references/.

AIDD · orchestrator (one skill)

AIDD = AI-assisted Design→Dev. Turn a design discussion into a platform-agnostic handoff package — rules / config / contract / acceptance / UI prototype / spine — that another AI (or person) can develop directly from, with deterministic checkers gating it to "0 major issues" before handoff. Domain-neutral: the same engine designs a game system or a SaaS billing feature. Landing the implementation (picking a tech stack to write the client/server) is downstream, not in this package.

This is one skill covering the whole pipeline. The detailed playbook for each phase lives in references/ and is read on demand — keeping context lean without splitting into many skills.

Core model

• Two layers: ① the live project spine — project-charter (concept / platform / target users / domain / naming conventions) + manifest (system list + dependency graph + cross-layer index + number-range registry + freeze ledger), continuously revised; ② the per-system production loop.
• Centralized source of truth: enums / numbering / contracts / config tables live in the project layer; prose references them by name, position-independent (table[key].field / screenID.elementID / R-code) — never inline a bare number.
• Every operation: read the spine → do the work → write back to the spine.
• Canonical method: references/methodology.md. Pitfalls: references/gotchas.md. Per-harness specifics: references/harness-adapt.md.

On entry — self-check, then route (proactively drive forward)

1. Does the spine exist? (project-charter + manifest) — No → start at the concept phase to build it. Yes → read both.
2. Which systems are not yet Final? List them; ask which to push this time.
3. Any Recheck systems? (a shared item changed) → resolve via the finalize/sync phase first.
4. Number-range conflicts / dependency cycles? Check and report.

Phases (read the matching playbook, then do the work)

What you want to do	Read
Concept: core loop / platform / users / split into systems / start the spine	references/phases/concept.md
Design a system: rules (R-codes) + config tables (test data) + UI prototype	references/methodology.md — Steps 0–5, part A
Iterate after a dry-run: tweak rules / config / prototype (cheap, repeatable)	references/methodology.md — re-run part A
Finalize & hand off: contract + acceptance + (as needed) backend + checklist; then sync the spine	references/methodology.md part B + references/handoff-checklist.md
UI screenshot → UI DSL (capture a reference screen into the knowledge base)	references/phases/ui-capture.md

Typical path: concept (once) → per system { design → iterate… (repeat) → finalize & hand off } → sync (continuous). Ordering isn't enforced — route as the work needs. A re-split / bounce records a rollback-log entry in the manifest and marks finalized downstream systems Recheck.

Deterministic checkers (the gate — run before handoff)

references/scripts/ (pure-Python, mostly stdlib): config_check (config-spec ↔ xlsx drift) · value_check (broken FKs / coverage / monotonic) · manifest_check (spine self-consistency) · ref_graph --check (dangling R-codes / refs) · gherkin_to_checklist · ui_render (UI-DSL → wireframe). 0 major = handoffable. Runnable example: references/examples/subscription-plans/.

Boundary (what it is not)

• Manages structure & consistency, not quality judgment: the checkers catch broken links / coverage / monotonicity / schema drift; they don't judge whether the design is good (game balance, UX, business correctness) — that's for people / dedicated tools.
• The HTML prototype validates information architecture & flow, not feel / timing / networking. Good for UI-dense surfaces (inventory, shop, settings, dashboards, onboarding); leave real-time feel to an engineering prototype.
• Doesn't decide your numbers / conventions: anything undecided is tagged [to confirm] and handed to a person.
• Goes as far as a handoff package; tech-stack choice and client/server implementation = downstream.

Package (one folder)

Install the single aidd/ skill folder into the host's skills directory (e.g. Claude Code's .claude/skills/). Everything — methodology, checkers, templates, examples — ships under aidd/references/. Domain packs (app / game patterns + examples) drop into references/patterns/ + references/examples/.

Hard constraints (don't violate)

• Files are UTF-8 without BOM.
• Reference by name (position-independent); per-myriad uniformly /10000 where used.
• Centralized config: tables pooled globally, no per-system private tables; a cross-system reference is an ordinary reference, not a dependency edge. Broadcast-type shared sources (id namespace / enum / text) are append-only, never break.
• Tool calls must use your harness's function-call format or they may silently fail — verify the artifact actually landed (read-file / ls). Per-harness specifics in references/harness-adapt.md.
• After each change, append to the project change ledger (CHANGELOG.md); init from references/templates/CHANGELOG.tpl.md if absent.