Strategy-First Development

Turn a high-impact product or engineering request into a shared control plane before implementation: strategic consensus, mature-reference research, per-capability technology selection, roadmap gates, and canonical project documents.

When to use it

• A greenfield product or MVP has no accepted strategy or stack.
• A core runtime, database, auth, orchestration, deployment, or state owner must be selected or replaced.
• A provider, model, UI, data, search, jobs, or observability decision needs evidence rather than familiar defaults.
• The team wants strategy and stack decisions materialized before code.

Do not use it for skill/prompt meta-work, small copy edits, translations, already-scoped bug fixes, pure code review, or deployment of an already-selected architecture.

What it produces

• STRATEGY_PACKET and CONSENSUS_GATE.
• Mature product and architecture findings with adopt / absorb / spike / defer / reject decisions.
• CAPABILITY_BLOCK_INVENTORY, frozen decision briefs, candidate records, shortlists, spikes, and STACK_SELECTION_GATE.
• Whole-stack coherence review across runtime, ownership, security, deployment, operations, and exit paths.
• Now / Next / Later / Not Doing roadmap slices.
• Canonical AGENTS.md, strategy, north-star, technical-stack, and roadmap documents in write-enabled modes.

Strategy materialization is docs-only. It never installs dependencies, scaffolds runtime modules, edits deployment files, or claims implementation completion.

First-selection defaults

Tier	Typical impact	Discover	Credible	Archetypes	E2 deep reads
A	state/security/topology owner; expensive migration	7	5	3	3
B	meaningful production subsystem	5	3	3	2
C	local, low-risk, replaceable utility	3	2	2	1

Counts are floors, not padding targets. Same-engine wrappers, toy projects, incompatible choices, and search-only evidence do not count. Tier A acceptance requires a representative spike unless recent equivalent evidence exists for the same major version, deployment shape, workload, and constraints.

Selection funnel

1. Inspect current code, docs, manifests, ADRs, and ownership boundaries.
2. Inventory every applicable decision-worthy capability.
3. Freeze must-haves, hard filters, weights, and candidate floors.
4. Discover candidates across distinct solution archetypes and source classes.
5. Verify current version, license, lifecycle, compatibility, security, and maintenance.
6. Hard-reject ineligible options before scoring.
7. Deep-read finalists, run sensitivity analysis, and define or execute required spikes.
8. Select per block, then run whole-stack coherence.
9. Keep unresolved decisions provisional; only passed gates may become accepted.

The full protocol lives in references/technology-selection-protocol.md.

Boundaries

• discuss-only and strategy-packet never write files.
• Accepted ADRs stay frozen during targeted_change unless new evidence invalidates them.
• Network gaps, stale evidence, incomplete candidate coverage, or unresolved Tier A spikes keep the stack provisional.
• User-facing products must define agent role, authority, hidden internal details, and role-specific quality bars.
• AGENTS.md remains canonical; CLAUDE.md and CODEX.md are pointer/overlay files, not duplicated strategy.

同类 Skill 对比

Drafted through tranfu-publish and approved by griffithkk3-del.

公司库内

• Project Docs Initialization — initializes existing-repo documentation; difference: this skill reaches strategy and stack consensus first.
• Agent Architecture Decision — decides agent workflow shape; difference: this skill governs the whole product and technology stack.
• AI Project Scoring — evaluates whether to invest; difference: this skill plans how an accepted product is delivered.

外部世界

• 暂无

本 skill 独特价值

• Per-capability candidate pools
• Evidence-gated convergence
• Strategy documents before code

使用技巧

Drafted through tranfu-publish and approved by griffithkk3-del.

材料方案

• Start from canonical docs and ADRs.
• Verify dynamic facts from primary sources.
• Keep a native/no-dependency baseline.

推荐用法

• Begin with discuss-only when uncertain.
• Use wide_first_selection for greenfield work.
• Use targeted_change for accepted stacks.

已知限制

• Broad first-selection research takes time.
• Offline evidence remains provisional.
• Tier A usually needs an authorized spike.