perhapsspy

Check early design or implementation-planning direction for fit with user/domain purpose, explicit constraints, current evidence, and verifiable success conditions before existing code, copied behavior, fallback/default logic, or shortcut pressure shapes the design. Use when reuse/precedent, rejected constraints, meaning-affecting fallback, or "copy now, clean later" could steer design. Skip mechanical edits, read-only owner discovery, clear specialist cases, and coding after direction is settled.

Use when a concrete document, documentation diff, or explicitly selected documentation package needs final wording or document-level cleanup such as distinct roles, reader routes, current-vs-stale separation, and one owner per current decision, contract, gate, or plan. Stay within the selected documentation scope; do not choose external source truth, design task state, or change implementation meaning.

Use when cross-layer feature, migration, integration, port, refactor, review, bug fix, or design planning needs one owner for user/domain meaning across UI, route, client state, command, API, storage, realtime, adapter, or presentation layers. Trigger for identity aliases, lifecycle/status, permissions/capabilities, route/query grammar, command payloads, result/event projection, freshness/fallback/revision semantics across representations, compatibility translation, duplicated meaning rules, or multiple representations of the same user/domain meaning. Do not use for read-only owner discovery, local flow cleanup after owners are clear, pure async responsiveness/freshness work, or scope-control alone.

Use for read-only source-of-truth audits: identify the current code, API, config, doc, or behavior to follow; compare a proposed, migrated, ported, or current change against it; and report evidence, mismatches, unresolved decisions, and owner-level recommendations.

Audit scattered repository docs and notes, then move only the right working context into the `project-context` structure.

Resume, handoff, and long-running repo work by keeping durable context across threads or agents: task briefs, logs, and reusable reference notes.

Use for code generation, feature work, bug fixes, refactoring, and code review where readable primary flow, minimal boundaries, owned decisions, isolated side effects, clear async/state behavior, or contract-focused tests improve correctness and maintainability. Skip only purely mechanical edits, trivial local changes, and throwaway experiments.

Use for non-trivial code-change planning, implementation, bug fixes, refactors, review edits, and diff reviews where scope decisions such as fields, options, abstractions, or follow-ups must be necessary, goal-proportional, evidence-backed, and verifiable without speculative work

Create complete, usable interfaces rather than polished mockups for new screens or major redesigns. Use when the user wants a landing page, dashboard, web, mobile, or desktop app screen, or a major flow and still needs decisions about content, hierarchy, layout, responsive behavior, or relevant loading, empty, error, and permission states. Preserve existing brand systems and verify the rendered result. Skip small style edits, mechanical implementation of a settled specification, general code refactoring, read-only interface review, documents, slides, email, CLI work, and standalone graphics.

Use when creating, editing, or reviewing AGENTS.md, agents.md, CLAUDE.md, Copilot instructions, Cursor rules, or any always-read repository instruction file. Keep persistent agent instructions small, durable, non-duplicative, and action-oriented; classify content that belongs in commands, onboarding, reference docs, handoff state, enforcement, or removal instead of preserving it in always-read files.

Use for high-token-risk Codex work: long sessions, broad reads/diffs/logs, browser/UI loops, subagents, always-read changes, or token audits. Guides summary-first reads, bounded delegation, compact resume state, and usage audits.

Use for interactive app code where typing, selection, navigation, search/filter, command palettes, previews, streaming/tool output, realtime updates, or file/diff/log views are laggy, stale, or race-prone because user intent/source state, derived presentation, async IO, scheduling, or background work are mixed. Keep source state prompt, move only justified expensive work off the urgent path, and commit async or presentation results through freshness-owning boundaries.