{"slug": "github-copilot-reply-best-way-to-build-hermes-expert-tools", "title": "GitHub Copilot reply — best way to build hermes, expert tools,", "summary": "A developer outlined a blueprint for \"Hermes,\" an operations copilot built on GitHub Copilot for a wood-interior design and fabrication studio, designed to convert customer inquiries into checked project briefs and reply drafts. The proposed agent is restricted to reading CRM leads, studio knowledge and pricebooks, drafting replies without sending, and calculating quotes only through a deterministic calculator, with human approval required for pricing, scheduling and code-compliance claims.", "body_md": "Build **Hermes as the operations copilot for your wood-interior studio**. Start with one job: turn a new customer inquiry into a checked project brief and a useful reply draft. Don’t start by asking it to design anything, set prices, or make promises on its own.\n\n**Input:** A customer’s email or website inquiry, plus any photos, sketches, or measurements they provide.\n\n**Hermes produces:**\n\n- A short summary of what the customer wants.\n- A list of known facts, missing details, and assumptions.\n- A preliminary project category, such as fitted wardrobe, kitchen, wall panelling, shelving, or custom furniture.\n- Follow-up questions tailored to that project.\n- A draft reply and, if appropriate, a site-survey checklist.\n\n**Hermes must not:** Send messages, confirm a price or completion date, order materials, or claim a design is structurally safe or code-compliant. Those require your review.\n\nFor example, for “We’d like an oak built-in wardrobe under a sloped ceiling,” Hermes should ask for approximate dimensions, location, photos, desired storage, finish, budget range, and timing. It should not invent a price from the description.\n\nCreate a folder called **Hermes Studio Knowledge**. Add short, dated documents for:\n\n- **Services:** what you make and install, and what you don’t.\n- **Service area:** locations you cover and any travel conditions.\n- **Materials and finishes:** woods, veneers, boards, hardware, finishes, and approved suppliers.\n- **Process:** inquiry, survey, design, quote, approval, deposit, fabrication, installation, handover.\n- **Commercial rules:** minimum project size, payment terms, quote validity, lead-time policy, warranty, and what affects cost.\n- **Policies and constraints:** what requires a site survey, what requires specialist review, and any relevant local building or fire-safety requirements.\n\nMark unknowns as **“not set yet.”** Don’t let Hermes fill them in. Put prices, labor rates, taxes, and lead times in a controlled pricebook or spreadsheet—not in an informal prompt.\n\nFor a first version, use a CRM or a structured table such as Airtable. Give each lead fields like:\n\n```\nlead_id\ncustomer_name\nemail\nphone\nproject_location\nproject_type\ncustomer_request\nbudget_range\ndesired_timing\ndimensions_provided\nphotos_or_files\nmissing_information\nstatus\nlast_contact_date\nhuman_owner\n```\n\nUseful statuses: `new`, `needs_information`, `survey_needed`, `quote_in_progress`, `won`, `lost`. Keep customer files in a separate project folder and link to them from the record.\n\nStart with these permissions:\n\n| Tool | Initial permission | \n|---|---|\n| CRM | Search and read leads; create a lead only if you approve or the intake workflow explicitly permits it | \n| Studio knowledge and project files | Read approved documents and customer-provided files | \n|  | Draft replies; **no sending** | \n| Calendar | Check availability; ask before booking | \n| Pricebook | Read approved materials and pricing inputs | \n| Quote calculator | Calculate only from your written formulas and supplied rates; show assumptions | \n\nThe language model should not calculate a binding quote from intuition. Put estimates behind a deterministic calculator you control, and label the result “preliminary” until you’ve checked the measurements, materials, labor, and installation conditions.\n\nUse this as the starting system instruction in your chosen agent platform:\n\n```\nYou are Hermes, the operations copilot for a wood-interior design and fabrication studio.\n\nYour first task is to turn customer inquiries into accurate project briefs and\nhelpful reply drafts. Use only the studio knowledge, CRM records, and customer\nfiles made available to you.\n\nRules:\n- Separate customer-provided facts, studio rules, and your assumptions.\n- Never invent dimensions, material availability, prices, lead times, or\n  building-code conclusions.\n- If essential information is missing, ask clear, project-specific questions.\n- If customer records conflict or a project is outside studio policy, stop\n  and flag it for the studio owner.\n- You may prepare drafts and recommendations. Do not send messages, confirm\n  commercial terms, place orders, approve designs, or book appointments without\n  explicit human approval.\n- Do not claim a design is structurally safe or compliant. Flag those questions\n  for qualified human review.\n- Cite the source document or customer file for important claims.\n\nFor every inquiry, return:\n1. Customer request summary\n2. Facts supplied by the customer\n3. Missing information and questions to ask\n4. Potential risks or items needing a site survey\n5. Suggested next step\n6. Reply draft\n7. Any proposed CRM updates\n```\n\nSkills should be specific procedures, not just personality descriptions. Start with one:\n\n```\n# Handling a new wood-interior inquiry\n\n1. Read the inquiry and inspect any customer-provided attachments.\n2. Search the CRM for an exact email match. If multiple records may match,\n   stop and ask the studio owner to choose.\n3. Classify the request using the studio's service list. If the project does\n   not clearly fit, flag it rather than forcing a category.\n4. Extract only information actually supplied: project type, location,\n   dimensions, materials, finish, budget, timing, and installation needs.\n5. List unknowns separately. Never infer measurements from photos.\n6. Check studio policies and approved material information where relevant.\n7. Identify questions that affect feasibility or a responsible preliminary\n   estimate. For built-ins, consider access, walls/floors, services, ceiling\n   shape, installation, and finishing requirements.\n8. Prepare a project brief, proposed CRM updates, and a friendly reply draft.\n9. Flag anything requiring a site survey, supplier confirmation, specialist,\n   or owner decision.\n10. Wait for approval before sending a message, booking a survey, or changing\n    a commercial commitment.\n```\n\nAdd other skills later—**site-survey preparation**, **material options**, and **quote preparation**—once the first workflow works reliably.\n\nA practical low-code setup is a form or inbox trigger, a workflow tool such as n8n, your CRM, a document store, and an LLM connected through controlled tool calls. The workflow should look like this:\n\n```\nNew inquiry\n  → Save inquiry and files\n  → Hermes extracts details and checks studio knowledge\n  → CRM lookup\n  → Validate required fields and flag ambiguity\n  → Prepare brief + questions + email draft\n  → Owner reviews and edits\n  → Owner sends or approves the next action\n```\n\nKeep approval outside the model: the review screen should show the draft, evidence, proposed CRM changes, and any uncertainties. If an approved proposal changes, request approval again.\n\nCreate test inquiries for a clear wardrobe request, vague “custom woodwork” request, missing location, conflicting dimensions, inaccessible photo, out-of-area project, and request for an immediate fixed quote. For each, check that Hermes:\n\n- Extracts only stated facts.\n- Asks useful questions rather than guessing.\n- Finds the right CRM record—or flags ambiguity.\n- Uses the studio’s actual policies and catalog.\n- Produces a helpful, on-brand draft.\n- Does not send, quote, book, or promise anything without approval.\n\nStart in **draft-only mode**. Let Hermes prepare briefs and replies while you compare them with your own decisions. Expand its permissions only after it consistently handles real inquiries correctly.\n\n**The first build target is simple:** a new inquiry goes in; Hermes returns a clear brief, the right follow-up questions, and a polished reply draft—with no unsupervised customer commitments.", "url": "https://wpnews.pro/news/github-copilot-reply-best-way-to-build-hermes-expert-tools", "canonical_source": "https://gist.github.com/cyberA87/1dbe5ac39891cbae63792173f2f7231f", "published_at": "2026-10-04 01:46:17+00:00", "updated_at": "2026-10-04 02:08:08.878857+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "ai-products"], "entities": ["GitHub Copilot", "Hermes", "Airtable"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/github-copilot-reply-best-way-to-build-hermes-expert-tools", "markdown": "https://wpnews.pro/news/github-copilot-reply-best-way-to-build-hermes-expert-tools.md", "text": "https://wpnews.pro/news/github-copilot-reply-best-way-to-build-hermes-expert-tools.txt", "jsonld": "https://wpnews.pro/news/github-copilot-reply-best-way-to-build-hermes-expert-tools.jsonld"}}