--- title: "Prompt Hotbar" type: project description: "A keyboard-first workspace for retrieving reusable AI instructions and reviewed operational references at the moment they are needed." summary: "Prompt Hotbar began with nine number-key shortcuts for repeated operational instructions. Continued use turned it into a broader retrieval workspace with 72 built-in public-safe prompts, 11 preset and Workflow Sets, a 134-card Toolbox for reviewed cross-tool operational references, browser-local personal prompts and preferences, reviewed Discovery, LAB experiments, and a private local MCP pilot." ai_summary: "Keyboard-first retrieval workspace with nine configurable slots, 72 built-in public-safe prompts, 11 preset and Workflow Sets, browser-local personal prompts and positive preferences, reviewed Prompt/Workflow/LAB Discovery pages, six LAB challenges, and a 134-card Toolbox spanning Claude Code, Codex, Antigravity, Hermes Agent, and OpenCode. Toolbox references pair copyable commands or configuration recipes with official provenance, review dates, and bounded risk cues; low-frequency Install / Setup lives behind Provider utilities rather than permanent cards, and unsupported symmetric commands are omitted instead of invented. Prompt Hotbar copies instructions and references but does not execute prompts, commands, installers, or configuration changes. A private local read-only MCP pilot separately tested reviewed in-client retrieval without accessing browser-local personal state." status: public visibility: public language: en topics: [ai-workflow, productivity, developer-tools] tags: [vite, react, typescript, local-first, prompt-reuse, discovery, mcp, personalization] audience: [builders, product-managers, ai-power-users] publishedAt: 2026-07-09 updatedAt: 2026-10-26 featured: true series: null related: [articles/a-prompt-library-is-not-enough, articles/ai-product-planning-team-capability, articles/evidence-first-ai-workflow, articles/prompt-hotbar-local-mcp-pilot, articles/prompt-hotbar-local-first-preference-layer] relations: - target: "articles/a-prompt-library-is-not-enough" relation: related_to note: "Explains why Prompt Hotbar emphasizes reusable workflow context instead of a static prompt dump." - target: "articles/ai-product-planning-team-capability" relation: related_to note: "Connects personal prompt reuse to repeatable team capability and workflow design." - target: "articles/evidence-first-ai-workflow" relation: related_to note: "Explains the inspection, provenance, scope-control, and safety discipline that also informs Toolbox curation." - target: "articles/prompt-hotbar-local-mcp-pilot" relation: related_to note: "Documents the first successful private local MCP use of Prompt Hotbar inside Claude Code." - target: "articles/prompt-hotbar-local-first-preference-layer" relation: related_to note: "Documents the v0.6.0 Local-first Preference Layer and the decision to keep personal preference state browser-local until cross-device requirements justify a backend." sourceStatus: original sourceSensitivity: public monetization: free projectStatus: active role: "product designer and builder" artifactType: tool artifactUrl: "https://prompt-hotbar.vercel.app/" repoUrl: null image: "https://prompt-hotbar.vercel.app/social/prompt-hotbar-og.png" imageAlt: "Prompt Hotbar with nine numbered 퀵슬롯 for reusable AI instructions." mediaEmbed: provider: linkedin title: "Launch demo — planning to verification" embedUrl: "https://www.linkedin.com/embed/feed/update/urn:li:ugcPost:7487886929766662144?collapsed=1" sourceUrl: "https://www.linkedin.com/posts/wonbeenlee_aiproductmanagement-productmanagement-aiworkflow-ugcPost-7487886929766662144-awBj" caption: "Plan a feature in ChatGPT, retrieve a verification prompt in Prompt Hotbar, place it in slot 4, copy it with the number key, and continue in the same conversation." publicProof: - "Live product: https://prompt-hotbar.vercel.app/" - "Reviewed Discovery index: https://prompt-hotbar.vercel.app/discover/" - "Current version line: v0.8.0 plus reviewed post-v0.8 Production updates" - "Current verified Prompt Hotbar main / Production merge: 8233f786271c4e771122e4a8c86fffe55feb8e9d" - "Current Library snapshot: 72 built-in public-safe prompts and 11 preset / Workflow Sets" - "Current Toolbox snapshot: 134 unique cards across Claude Code, Codex, Antigravity, Hermes Agent, and OpenCode" - "Toolbox Product contributions: Claude Code 46, Codex 41, Antigravity 26, Hermes Agent 54, OpenCode 1" - "Toolbox reference kinds: 49 Slash Commands, 20 Keyboard Shortcuts, 26 Agent Skills, 3 Configuration cards, and 36 CLI Commands" - "Reviewed Toolbox operations cover maintenance, diagnostics, permission/autonomy modes, and low-frequency Provider Setup with official provenance" - "Browser-local Like / Love this preferences and My Picks are deployed without accounts or server-side preference storage" - "Privacy-bounded PostHog measurement covers curated preference transitions and My Picks activation without sending personal prompt bodies or local prompt IDs" - "Privacy-safe Discovery analytics and cached approximate view buckets are deployed" - "Published LinkedIn launch post: https://www.linkedin.com/posts/wonbeenlee_aiproductmanagement-productmanagement-aiworkflow-ugcPost-7487886929766662144-awBj" - "LinkedIn-native split-screen launch demo embedded on this project page" - "Private local Claude Code MCP pilot verified five read-only tools against 23 reviewed public-safe prompts and three workflows" - "Public npm publication remains intentionally withheld during the personal-use evaluation" - "Launch film (34 s, 1280×720, no narration) embedded on this project page and served from the Prompt Hotbar product domain" canonicalUrl: "https://personal-ai-ops.vercel.app/wbeen/projects/prompt-hotbar" --- ## Overview Prompt Hotbar is a keyboard-first workspace for retrieving reusable AI instructions and operational references at the moment they are needed. It began with a very small frustration: I kept typing the same short instructions. > “Looks good, proceed.” > “Stay within scope.” > “Verify the work when you’re done.” Longer instructions increasingly moved to voice input, while the keyboard was still carrying these short approval, scope-control, verification, and handoff messages. I wanted those instructions close at hand without turning them into another document I had to search. That became Prompt Hotbar. Repeated use widened the problem. Prompt retrieval was only one part of the friction. I was also repeatedly looking up the same AI-tool commands, shortcuts, permission modes, setup paths, and diagnostic references across coding agents. The product therefore grew from a Hotbar and Prompt Library into three connected surfaces: **Library**, **Toolbox**, and **LAB**. **국문 요약:** 반복해서 쓰는 AI 업무 지시를 1–9번 퀵슬롯에 넣고 필요한 순간에 바로 찾아 복사하는 데서 출발한 키보드 중심 워크스페이스입니다. 현재는 72개의 내장 Prompt와 11개의 Preset / Workflow Set을 제공하는 Library, Claude Code·Codex·Antigravity·Hermes Agent·OpenCode의 검증된 운영 레퍼런스를 다루는 134-card Toolbox, 그리고 모델·프롬프트 capability를 실험하는 LAB까지 확장되었습니다. 개인 Prompt와 선호 상태는 browser-local을 유지하며, Toolbox의 명령·설정 역시 복사만 제공하고 제품이 직접 실행하지 않습니다. - **Live product:** [prompt-hotbar.vercel.app](https://prompt-hotbar.vercel.app/) - **Reviewed Discovery:** [Prompt, Workflow, and LAB assets](https://prompt-hotbar.vercel.app/discover/) - **Preference layer case-note:** [Preference memory without accounts](/wbeen/articles/prompt-hotbar-local-first-preference-layer/) - **Evidence-first Guide:** [Inspect, verify, and audit before asking AI to change anything](/wbeen/articles/evidence-first-ai-workflow/) - **Local MCP pilot:** [Testing Prompt Hotbar inside Claude Code](/wbeen/articles/prompt-hotbar-local-mcp-pilot/) - **Source status:** private development repository - **Published LinkedIn post:** [From repeated instructions to Prompt Hotbar](https://www.linkedin.com/posts/wonbeenlee_aiproductmanagement-productmanagement-aiworkflow-ugcPost-7487886929766662144-awBj) ## Launch film A 34-second product film without narration. It shows the same approval sentence being typed again, the prompt placed in Hotbar slot 1, copied with the `1` key, and pasted back into the running conversation. The file is served from the Prompt Hotbar product domain, which keeps the product's own assets canonical. ## What it does The product still keeps retrieval separate from execution. In **Library**, the core interaction remains deliberately simple: 1. Find a Prompt or choose a reusable Set. 2. Place it into one of nine Hotbar slots when it deserves a shortcut. 3. Click the slot or press its number key. 4. Paste the copied instruction into the AI conversation already in progress. In **Toolbox**, the same retrieval principle applies to operational knowledge: 1. Filter or search by AI tool, purpose, or reference type. 2. Inspect a reviewed command, shortcut, Skill, or configuration reference. 3. Copy the exact reference when appropriate. 4. Run it yourself in the intended tool or environment after reviewing its source and risk context. Prompt Hotbar does not perform the task for the user. It does not automatically send Prompts to an AI service, execute shell commands, run installers, or mutate tool configuration. The AI conversation or external tool keeps its own context and remains responsible for actual execution. ## The workflow shown in the launch demo The published demo connects Prompt Hotbar to an actual planning context rather than showing copy behavior in isolation. - A feature is planned with ChatGPT. - A reusable verification Prompt is found in Prompt Hotbar. - The Prompt is placed into slot 4 and copied with the number key. - It is sent back into the same conversation. - The existing implementation plan is converted into a concrete verification checklist. This remains a useful expression of the product boundary: Prompt Hotbar sits between planning and execution as a retrieval layer, while the AI conversation keeps its own context and performs the reasoning. The LinkedIn-native demo is embedded on this page. The original LinkedIn post remains the canonical source for its bilingual launch copy, reactions, comments, and publication context. If LinkedIn does not allow the player to load in a particular browser, the original-post link remains available directly below it. ## Current product experience What started as nine shortcuts now includes three main product surfaces. ### Library - **Nine configurable Hotbar slots** with click and number-key copy - **72 built-in public-safe Prompts** for approval, verification, scope control, handoff, reporting, safety, writing, release, design, and other recurring AI work - **11 built-in Preset / Workflow Sets** for switching between recurring work modes - Browser-local personal Prompts with CRUD and Hotbar placement - Board / List browsing, search, category and workflow grouping, sorting, and display controls - Browser-local positive preferences and **My Picks** for returning to useful Prompt, Toolbox, LAB, and local My Prompt cards without an account - Already-equipped Quickslot controls with a light accent state so assigned shortcuts are easier to recognize - Korean and English UI with a separate AI response-language preference - Safe sharing based on stable public IDs and slot numbers rather than Prompt bodies - Versioned local JSON backup / restore for My Prompts and Local Sets ### Toolbox The reviewed Production snapshot contains **134 unique cards**. Product contribution counts overlap where a shared reference or multi-agent Skill applies to more than one tool. - Claude Code: **46** Product contributions - Codex: **41** - Antigravity: **26** - Hermes Agent: **54** - OpenCode: **1** - 49 Slash Commands - 20 Keyboard Shortcuts - 26 Agent Skills - 3 Configuration references - 36 CLI Commands The Toolbox is not intended to be another bookmark list. Its working model is **executable knowledge + provenance**: a useful command or configuration recipe should be directly copyable, but the user should also be able to see where it came from, when it was reviewed, and whether its behavior carries additional risk. That led to several curation and interaction decisions: - Repeated-use operational references such as **Update**, **Doctor / Status / Help**, and reviewed permission/autonomy modes remain normal Toolbox cards where the provider actually supports them. - Low-frequency **Install / Setup** does not occupy the normal card grid. It lives in each Provider utility menu and appears through explicit install/setup search. - Provider utilities also link to the official site, documentation, and release notes. Desktop right-click and mobile long-press expose the same secondary menu as a shortcut, not as the only discoverable path. - Permission/autonomy references distinguish **scoped**, **auto-reviewed**, **auto-edit**, and **full-bypass** behavior instead of flattening them into one generic “automatic” mode. - OpenCode V2 is represented through its reviewed permission-rule configuration instead of inventing a YOLO command that the official reference does not support. - Commands that look symmetrical across providers are not added merely for visual completeness. If an official equivalent cannot be verified, it is omitted. - Installer and permission commands are references only. Prompt Hotbar does not execute them. This is also why Toolbox source metadata matters. The useful asset is not just a string such as a CLI flag; it is the string plus enough provenance and risk context to decide whether it belongs in real work. ### LAB and Discovery LAB currently contains six reviewed capability challenges. It copies Prompt material for experiments but does not execute a Prompt Hotbar model backend. Discovery publishes a deliberately small reviewed proof set rather than generating public pages for every built-in asset. Prompt Hotbar owns the canonical Prompt, Workflow, and LAB pages and their copy/open actions; this site keeps the project story, methodology, and proof-of-work context without duplicating their full bodies. ## Preference memory and My Picks Repeated use exposed another retrieval problem. Finding an item once did not make it easier to return to the items that mattered most. Version 0.6.0 introduced a deliberately small positive-preference model: neutral, **Like**, or **Love this**. The states are mutually exclusive, and choosing the selected state again returns the card to neutral. There is no dislike state because no reaction is not reliable negative evidence; an item may simply not have been seen or needed yet. The preference map stays in the browser behind a versioned local abstraction. **My Picks** is derived from those positive states and turns preference into a useful retrieval surface rather than a decorative reaction control. No account, Supabase database, or cloud profile is required for the same browser to remember the state. That local boundary is intentional. A backend becomes justified when the requirement itself leaves the browser — for example PC↔mobile synchronization, recovery after local data is cleared, account-level personalization, or server-enforced public aggregation. Product learning is separated from personal state. Production PostHog receives only allowlisted behavioral metadata for curated assets, such as a canonical preference transition and My Picks activation. Personal Prompt bodies and local prompt identifiers are not analytics dimensions. The analytics setup also avoids persistent person identity, so aggregate behavioral signals should not be interpreted as durable cross-day user profiles. See [From Prompt Reuse to Personal Retrieval](/wbeen/articles/prompt-hotbar-local-first-preference-layer/) for the product and architecture reasoning behind this milestone. ## Local MCP pilot A private local pilot tested whether Prompt Hotbar could provide the same reviewed retrieval layer directly inside an AI coding client. Claude Code connected to a local `stdio` MCP server and exposed five read-only tools for Prompt search, Prompt retrieval, contextual recommendation, Workflow search, and Workflow retrieval. The runtime used a generated public-safe snapshot containing 23 reviewed Prompts and three Workflows. The pilot successfully handled natural-language requests to: - recommend a Prompt for resuming interrupted work - suggest verification-stage Prompts and explain why they fit - retrieve the nine-step `coding-agent` Workflow - return the canonical body of a requested Prompt The MCP did not read browser-local personal Prompts, clipboard data, Hotbar assignments, search history, accounts, or cloud state. It introduced no write tool and did not execute the retrieved Prompt. This creates a complementary product boundary: - **Web product:** visual discovery, organization, browser-local personalization, Prompt shortcuts, Toolbox references, and copy - **Local MCP pilot:** read-only search, recommendation, and retrieval inside a compatible AI client - **AI client / external tool:** task context, reasoning, and any actual execution The package is prepared as a standalone artifact, but public npm publication remains intentionally withheld while personal-use value is evaluated. See the [local MCP pilot proof note](/wbeen/articles/prompt-hotbar-local-mcp-pilot/) for the exact evidence and limitations. ## Reviewed Discovery proof set The Discovery Foundation publishes a small reviewed proof set rather than generating public pages for every built-in asset. ### Prompts - [Inspect the current state before acting](https://prompt-hotbar.vercel.app/prompts/inspect-current-state/) - [Keep the work within the minimum required scope](https://prompt-hotbar.vercel.app/prompts/keep-scope-minimal/) - [Run a public/private safety check](https://prompt-hotbar.vercel.app/prompts/public-private-safety-check/) ### Workflows - [Safe implementation workflow for a coding agent](https://prompt-hotbar.vercel.app/workflows/coding-agent-safe-implementation/) - [Business-model design and red-team workflow](https://prompt-hotbar.vercel.app/workflows/business-model-design-red-team/) ### Guide - [Before Asking AI to Change Anything: An Evidence-first Inspect, Verify, and Audit Workflow](/wbeen/articles/evidence-first-ai-workflow/) ### LAB - [Pelican on a Bicycle — SVG capability check](https://prompt-hotbar.vercel.app/lab/pelican-on-a-bicycle-svg-check/) The Toolbox itself remains a product surface rather than a duplicated PAOS catalog. Its individual commands, shortcuts, Skills, and configuration references stay canonical in Prompt Hotbar. ## Privacy and product boundary Personal Prompt content, clipboard contents, search text, Hotbar assignments, local settings, and the stored preference map stay on the device. Shared links do not contain Prompt bodies or personal configuration. Production measurement is limited to a small allowlisted set of non-personal events; automatic pageviews, session recording, autocapture, and persistent person profiles are disabled. For curated assets, preference analytics uses canonical public IDs and bounded transition metadata; local personal-prompt identifiers and bodies are rejected from that transport. Public detail-page view labels use cached, nonlinear aggregate buckets rather than raw counts or unique-person claims. The Toolbox adds another explicit boundary: copied commands, installers, configuration recipes, permission flags, and Agent Skill references are information products. Prompt Hotbar does **not** execute them, collect credentials for them, or silently change the external environment. The public web product intentionally does not include: - Automatic Prompt or command execution - Accounts or cloud synchronization - A server-side Prompt or preference database - Browser-wide global shortcuts - Public reaction counts or social ranking - In-product payments, teams, or a Prompt marketplace The local MCP pilot does not change those boundaries. It reads only the reviewed public-safe snapshot and does not access browser-local personal state. Prompt Hotbar does not process payments inside the product. Future support links, if activated, remain an external Support Layer and do not change access to the Core Utility. ## What changed through use The first version treated the problem as Prompt storage. Using it repeatedly showed that retrieval was the harder part. That changed the product direction. Similar Prompts needed to be grouped. The Library needed search. Sets needed to support different modes of work. New users needed to understand the placement and copy model quickly. Reviewed public asset pages became useful so that selected Prompt, Workflow, and LAB artifacts could be understood before opening the product. The local MCP pilot extended retrieval outward: when the user is already inside an AI client, natural-language search or recommendation can be more useful than another browser round trip. The Preference Layer extended retrieval inward: when the user stays in the browser, the product can remember which items were explicitly useful without turning that memory into an account system. Toolbox extended the same retrieval idea sideways across the operating environment. The repeated thing was no longer always a sentence to send to an AI. Sometimes it was a command, a permission mode, a diagnostic action, or a Skill that needed to be found again with enough evidence to trust it. The product evolution is therefore less about accumulating features than about making retrieval more situated: ```text prompt storage → structured retrieval → in-client retrieval → preference-aware retrieval → cross-tool operational retrieval ``` The useful lesson was not that every repeated sentence or command should become a product feature. It was that a small, persistent friction can reveal a clearer product boundary once it is observed in real work, shipped narrowly, reviewed against evidence, and expanded only when the requirement justifies the added complexity. --- Related reading: [A Prompt Library Is Not Enough](/wbeen/articles/a-prompt-library-is-not-enough/) · [Preference memory without accounts](/wbeen/articles/prompt-hotbar-local-first-preference-layer/) · [Evidence-first AI Workflow](/wbeen/articles/evidence-first-ai-workflow/) · [Local MCP pilot](/wbeen/articles/prompt-hotbar-local-mcp-pilot/) · [From personal prompts to team capability](/wbeen/articles/ai-product-planning-team-capability/) --- Canonical page: https://personal-ai-ops.vercel.app/wbeen/projects/prompt-hotbar This is the AI-readable Markdown mirror. Public, curated content only.