Wonbeen Lee wbeen / AI Product Ops notebook
Career AI-readable
← All articles

case-note · article

From Prompt Clipboard to Local MCP: Testing Prompt Hotbar Inside Claude Code

Prompt Hotbar began as a keyboard-first copy tool. This pilot tested a second retrieval path: Claude Code could search, recommend, and retrieve reviewed Prompt Hotbar prompts and workflows through a local read-only MCP server without publishing the package or sending personal prompt data to a backend.

published 2026-08-05 updated 2026-08-05 5 min read public public
#ai-workflow#mcp#prompt-reuse

Key takeaways

  • The local MCP pilot worked in Claude Code with five read-only tools.
  • Natural-language requests successfully triggered prompt recommendation, workflow retrieval, and exact prompt lookup.
  • The MCP runtime used a reviewed public-safe snapshot rather than browser-local personal data.
  • The npm package was prepared but not published; the pilot remained private and local.

What changed

Prompt Hotbar began as a keyboard-first retrieval tool: find a repeated instruction, place it into one of nine slots, copy it, and paste it into the AI conversation already in progress.

That interaction still works, but it leaves one question open:

Can the active AI client retrieve the same reviewed Prompt Hotbar assets directly, without turning Prompt Hotbar into an execution agent or uploading personal prompt data?

The first local MCP pilot answered that question positively.

국문 요약: Prompt Hotbar의 검토된 프롬프트와 워크플로를 로컬 MCP로 Claude Code에 연결해, AI 클라이언트 안에서 직접 검색·추천·조회하는 개인 파일럿에 성공했습니다. npm에는 아직 공개하지 않았고, 개인 프롬프트나 브라우저 로컬 데이터도 MCP에 포함하지 않았습니다.

The pilot boundary

The experiment used a local stdio MCP server connected to Claude Code from the Prompt Hotbar development repository.

It exposed exactly five read-only tools:

  1. search_prompts
  2. get_prompt
  3. recommend_prompts
  4. search_workflows
  5. get_workflow

The MCP runtime used a generated public-safe snapshot containing 23 reviewed prompts and 3 workflows. It did not read My Prompts, clipboard contents, browser localStorage, Hotbar assignments, search history, accounts, analytics profiles, or a remote prompt database.

It also had no write tool, no automatic prompt execution, and no remote MCP endpoint.

What worked in Claude Code

The test moved beyond a connection check. Claude Code used the MCP tools in response to ordinary natural-language requests.

1. Contextual recommendation

When asked to find a Prompt Hotbar prompt for resuming interrupted work, Claude Code called the recommendation tool and selected resume-interrupted-work as the best match, while also showing an alternative for a new-session handoff.

2. Stage-aware suggestions

When told that the current task was in verification, Claude Code recommended a sequence including build-and-test execution, verification checklist preparation, and a final audit pass. The result was not just a keyword match; it used workflow position and task stage to explain why each prompt fit.

3. Workflow retrieval

Claude Code retrieved the nine-step coding-agent workflow and identified which steps had already been completed and which verification steps remained relevant to the current task.

4. Exact prompt lookup

It also retrieved the canonical body of a requested prompt directly from the reviewed catalog.

Together, these checks covered the core MCP path:

Natural-language intent
→ Prompt Hotbar search or recommendation
→ reviewed Prompt or Workflow retrieval
→ explanation inside the active AI client

Why this matters for the product

This does not replace the Prompt Hotbar web product.

The browser interface still owns visual discovery, nine-slot placement, personal browser-local prompts, preset switching, LAB experiments, and direct copy interactions. The MCP adds a different retrieval surface for environments where the user is already working with an AI client.

The product boundary becomes clearer:

  • Prompt Hotbar web: visual browsing, organization, placement, and copy
  • Prompt Hotbar MCP: read-only search, recommendation, and retrieval inside a compatible AI client
  • AI client: reasoning, task context, and any actual execution

Prompt Hotbar remains the retrieval layer. It does not become the agent that performs the work.

Why the package is not public yet

The MCP package has been prepared and verified as a standalone artifact, but public npm publication is intentionally on hold.

The current goal is to use it privately in real work and observe whether the in-client retrieval path is consistently useful before making it globally installable. This avoids turning a technical milestone into a release claim before the personal workflow value is clear.

The local pilot therefore proves:

  • the reviewed catalog can be packaged independently
  • Claude Code can discover and use the five read-only tools
  • prompt recommendation and workflow retrieval work in a real host
  • the runtime can stay separate from personal browser-local state

It does not claim:

  • a public npm release
  • account or cloud synchronization
  • personal prompt access through MCP
  • remote hosting
  • automatic execution of retrieved prompts
  • general availability across every MCP host

Next evaluation

The useful next step is not broader distribution. It is repeated use.

The pilot should now answer product questions that implementation tests cannot:

  • Does natural-language retrieval save meaningful context switching?
  • Are recommendation scores and explanations intuitive enough?
  • Which prompts are repeatedly useful inside an AI client?
  • Do users prefer direct retrieval, the visual web interface, or a combination?
  • Is the public-safe catalog sufficient before any personal-data integration is considered?

Only after that evidence is available does a public npm release become a product decision rather than a packaging milestone.

Related: Prompt Hotbar project · A Prompt Library Is Not Enough · Evidence-first AI Workflow

Related

Markdown mirror md