Stack Scout

Stack Scout

Run a short research pass only for the stack areas that matter.

Keep the output practical. The goal is not to produce a literature review. The goal is to reduce implementation risk and hand back a brief that another Codex instance can apply immediately in the repo.

Workflow

1. Identify which stack pieces are involved.
2. Research only the pieces that are likely to be time-sensitive or easy to remember incorrectly.
3. Inspect only the minimum relevant local files needed to map the recommendation into this repo.
4. Return a short implementation brief.
5. If the user asked for implementation, switch from research to local coding work and verify the change.

Research Pass

Prefer a small number of focused questions over broad searching.

Research only what is relevant to the current task, for example:

• current best pattern for AI SDK streaming in the exact framework version
• current Gemini tool-calling or structured-output patterns
• Next.js App Router caveats for the feature being touched
• Drizzle migration or Bun workspace guidance when schema or tooling work is involved
• Docker or local Postgres setup caveats when the change touches local infra

If DOCKER MCP is available, use it as the default research sidecar instead of making the user re-explain the setup each session.

DOCKER MCP Bootstrap

When this skill runs in a fresh session:

1. Add exa and context7 through the DOCKER MCP gateway.
2. Create a code-mode helper named code-mode-stack-research if it does not already exist.
3. Use that helper for the external research pass.

Treat code-mode-stack-research as the default reusable research tool for this repo. The user should not need to ask for MCP setup again.

Recommended shape:

• servers: exa, context7
• purpose: combine current web results with official library docs into one short research packet
• output: one string with practical sections, not raw dump unless debugging the helper

Important details:

• code-mode tools are session-scoped, so recreate code-mode-stack-research when needed.
• For hyphenated tool names inside code mode, use globalThis['resolve-library-id'] and globalThis['get-library-docs'].
• Prefer exact official library IDs when known:
  • Next.js: /vercel/next.js
  • AI SDK: /vercel/ai

Fallback rules:

• If DOCKER MCP is unavailable, or code-mode-stack-research cannot be created, use the best available primary-source documentation tools and web search directly.
• Keep the same output contract even when the code-mode helper is unavailable.
• Do not block the task on recreating the helper if equivalent research can proceed without it.

Starter script pattern:

const getLibraryDocs = globalThis['get-library-docs'];

const nextDocs = getLibraryDocs({
  context7CompatibleLibraryID: '/vercel/next.js',
  topic: 'the exact Next.js feature involved in this task',
  tokens: 1800,
});

const aiDocs = getLibraryDocs({
  context7CompatibleLibraryID: '/vercel/ai',
  topic: 'the exact AI SDK feature involved in this task',
  tokens: 1800,
});

const web = web_search_exa({
  query: 'official docs, release notes, or migration guidance for the exact feature and version involved',
  numResults: 3,
});

return ['DOCS', nextDocs, aiDocs, 'WEB', web].join('\n\n');

Avoid:

• generic tutorials
• unrelated libraries
• long summaries with no implementation consequence
• researching the whole stack when only one layer is changing

Output Contract

Return a concise implementation brief with these sections:

1. Recommended pattern State the best fit for this repo and feature.

2. Caveats Call out version-specific or provider-specific constraints.

3. Avoid Name the patterns likely to cause churn, incompatibility, or unnecessary complexity.

4. Apply here After the research pass, point to the local files or modules where the pattern should be implemented.

5. Verify Suggest the local commands or checks that should confirm the change.

Keep the brief short enough that it can be read once and acted on immediately.

Implementation Hand-off

When the user asks to implement after the research pass:

• follow the recommended pattern unless the local codebase strongly suggests a better fit
• preserve existing repo conventions
• make the code changes locally rather than trying to do implementation through research tools
• verify with the smallest relevant set of local commands

Example Requests

The following user requests should trigger this skill:

• Use stack-scout before adding tool calling to our chat route
• Research the best current AI SDK pattern for this feature, then implement it
• Check the latest Drizzle migration approach before we touch the schema
• Scout Next.js 16 caveats for this UI change before coding