菜单
Control the Windows desktop/GUI via the 'desktop' MCP: screenshot, click, type, hotkeys, scroll, launch apps, and run audited multi-step workflows. Use for any on-screen automation the engine's macOS-only computer_use can't do.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
The trust-and-authority policy you operate under. Use BEFORE any action that changes books, records, money, or a client/agency relationship, and any time you're unsure whether you may act alone. Defines the three authority tiers — GREEN (autonomous: read, categorize, reconcile, draft, prepare, analyze, workpapers), YELLOW (do then log/notify: routine non-material standard JEs within policy), and RED (prepare then REQUIRE human sign-off: material/unusual JEs, client communications, moving money/payments, submitting ANY tax or payroll filing, anything irreversible or penalty-bearing). Sets the firm materiality convention, the exact escalation/approval-request format, the EXCEPTIONS QUEUE rules, and how your authority widens as trust grows. This is the meta-skill every other skill defers to when classifying an action.
THE DROP ZONE convention for RealDeal CPA. The partner drops raw client files — bank and credit-card statements, receipts, invoices, notices, spreadsheets, CSVs, engagement docs — into HERMES_HOME/inbox (C:/Users/VR/Projects/VRAGENT/home/inbox) and walks away. Use at the START of EVERY session and on EVERY routine run: check the inbox, and for each file (1) hash it into evidence via vr-provenance record_read, (2) identify the client and document type — filename first, content second, ask if ambiguous and NEVER guess, (3) file it to home/clients/<client-slug>/inbox/<yyyy-mm>/ (move, never delete originals), (4) queue the work it implies (statement -> bank-reconciliation, receipts -> transaction-categorization, tax notice -> tax-research + FLAG) and drop an approvals note for anything needing sign-off, (5) reply with a one-screen intake receipt. Unreadable or suspicious files stay put and get flagged. Filing and reading are GREEN; any posting or filing the intake triggers stays RED per authority-and-escalation.
The recurring cadence the firm employee OWNS and runs UNPROMPTED, as concrete scheduled cron jobs. Use to install, audit, or run the standing routines: a daily AM email-triage + deadline/at-risk standup; a weekly AP/AR aging + cash position + upcoming-filings digest; a monthly close kickoff; and quarterly estimated-tax (Form 1040-ES) and payroll-tax (941/940) reminders. Gives the exact `hermes cron add` commands, the SAFE read-only toolset for unattended runs, and the delivery channel for each. Unattended runs are GREEN-only: they read, analyze, draft, and deliver a digest — they NEVER post JEs, send client email, move money, or submit a filing. Anything material lands in the EXCEPTIONS QUEUE for the partner. Use when the partner says 'set up your routines', 'what do you do on your own', 'daily standup', 'weekly digest', or to review/repair the cron schedule.
Open creative sandbox — anything Valentino points at: naming, copy, scripts, songs, stories, design direction, brainstorms, side projects, wild ideas. Drop the buttoned-up register and go all in.
Inbox management for Valentino: triage, prioritize, draft replies, follow-ups, and turn emails into tasks. Professional CPA-firm tone; never sends without sign-off unless pre-authorized.
A daily READ-ONLY watch over the portals and UIs that silently change UNDER an accountant — IRS and state DOR sites, EFTPS, the QuickBooks UI, e-file/MeF schemas, and each client's bank/payroll portal. Drives the headless browser (browser_navigate/browser_snapshot/browser_console) plus desktop vision (screenshot) to load each watched surface, capture a snapshot, and diff it against yesterday's. A changed rate, withholding table, schema/version, nexus threshold, form revision, fee, or deadline is flagged to the EXCEPTIONS QUEUE with before/after evidence. Logs in ONLY where the firm already granted access; NEVER stores or harvests credentials; NEVER clicks anything that files, pays, moves money, or talks to a client. Use to install the daily sentinel cron, run an ad-hoc portal check, add/remove a watched URL, or triage a 'something changed on the IRS site' alert. Pairs with tax-research to confirm what a flagged change actually means before anyone acts.
| name | desktop-control |
| description | Control the Windows desktop/GUI via the 'desktop' MCP: screenshot, click, type, hotkeys, scroll, launch apps, and run audited multi-step workflows. Use for any on-screen automation the engine's macOS-only computer_use can't do. |
| version | 1.0.0 |
| author | RealDeal CPA |
| license | MIT |
| platforms | ["windows"] |
| metadata | {"hermes":{"tags":["Desktop","Windows","GUI","Automation","RPA","Screenshot","Mouse","Keyboard","computer-use"],"related_skills":["bookkeeping","practice-management"]}} |
You can see and operate this Windows machine through the desktop MCP server
(tools are prefixed by your client, e.g. desktop.screenshot). The engine's built-in
computer_use is macOS-only — use these tools instead on Windows.
screenshot first. Read the image; find the pixel coordinates of
your target. Use screen_info (e.g. 1920x1080) to reason about positions.screenshot a region to zoom.click / double_click / right_click / type_text / press_key /
hotkey / scroll / drag / open_application.screenshot again to confirm the UI changed as expected. Never assume
an action worked; look.screenshot(region?) → returns the screen as an image. region=[x,y,w,h] to crop.screen_info() → "WIDTHxHEIGHT".move_mouse(x,y), click(x,y,button,clicks), double_click(x,y), right_click(x,y)drag(x1,y1,x2,y2), scroll(amount,x?,y?) (positive=up, negative=down)type_text(text), press_key(key) ('enter','tab','esc','f5'…),
hotkey(['ctrl','c']), ['win','r'], ['alt','tab']locate_image(image_path, confidence) → "x,y" center of a template PNG, or "not found".open_application(name) → launches via the Run dialog ('notepad', 'excel', a path).run_workflow(steps_json, allow_unsafe=false) → run a multi-step
ai_rpa_system workflow through the audited engine; a security scan blocks
CRITICAL-risk workflows unless you pass allow_unsafe=true. Prefer this for repeatable
sequences (e.g. logging into a desktop app, batch data entry).Driving QuickBooks Desktop / Lacerte UI where no API covers a step, filling desktop forms, moving files in Explorer, operating any app that lacks an MCP. When an MCP or CLI exists for the task (QuickBooks, KarbonCopy, Lacerte, browser), prefer that over pixels.