> ## Content Index > Fetch the complete content index at: https://blog.melardenio.com/llms.txt > Use this file to discover other available public pages before exploring further. # Meet Team Clutch: The AI Agents That Run My Client Work (and Never Hit Send) - URL: https://blog.melardenio.com/meet-team-clutch-the-ai-agents-that-run-my-client-work-and-never-hit-send/ - Published: 2026-11-02T15:13:50.000Z - Updated: 2026-11-02T15:13:50.000Z - Author: Rommell Ardenio - Tags: AI & Automation, Broadcast Engineering For twenty-plus years in broadcast engineering, the rule that kept us on air was simple: nothing goes to transmission without a human hand on the switch. Automation could prepare the playlist, cue the clips, even flag the loudness problems before we did. But the moment something went *out*, someone was accountable for pressing the button. I took that rule with me when I started building my own client side-business, and it ended up shaping the whole thing. Instead of one giant "AI assistant" that does everything, I built a small team of Claude Code sub-agents, each with one job, a name, and a hard rule: **they draft, I decide.** None of them can send an email. Most of them can't even run a shell command. That's on purpose. I call them Team Clutch. Here's who's on it, what they do, and how they hand work to each other. ## Why a team instead of one big prompt I used to do client work the way most of us do: open a chat, paste the meeting notes, ask for a quotation, then ask for an invoice two weeks later and re-explain everything. Every time, the AI re-derived the context, and every time it was a little different. Prices drifted. Terms got reworded. Once, a quotation and the invoice that followed it didn't agree on the deposit. The fix wasn't a smarter prompt. It was structure. Each agent now has its own charter file, its own skills, and a short list of things it is never allowed to do. They all read the same team charter first, and they all pass work through files in one client folder, so the next agent picks up exactly what the last one produced, not a paraphrase of it. ## Pia: the questionnaire and the pitch Pia works at both ends of a deal. Before the first meeting, she drafts a short questionnaire tailored to the kind of project (a website gets about ten plain-language questions, roughly ten minutes for the client). The client's answers get saved word for word, so the meeting starts with their own words instead of my guesses. Later, once the scope and quotation exist, Pia writes the cover and the pitch and assembles one client-facing proposal: cover, narrative, scope, quotation, acceptance. Her rules are strict: she copies the scope and the numbers *verbatim*, she never promises an outcome the scope doesn't commit to ("this will double your sales" is banned), and her "why us" section can only use facts from one approved file. If that file is thin, she leaves a marked gap and asks me instead of inventing credentials. **Skills:** draft-questionnaire, draft-proposal. ## Lota: scope first, then the quotation Lota turns meeting material (notes, a Whisper transcript, chat threads, the questionnaire answers) into a scope of work. Every deliverable gets acceptance criteria, every milestone has to be something that can later be invoiced, and a scope without an exclusions list counts as incomplete. The important part is what she does with gaps. If the material doesn't say something that changes price, deliverables or timeline, she doesn't guess. She asks me, in one batch, most important first. Then she drafts a reply email to the client re-asking anything still vague. Only after I say "scope approved" does she draft the quotation, using a rate card I fill in myself. If a rate says TBD, she stops and asks. **Skills:** draft-scope, draft-quotation. ## Torney: the legal desk Torney is the newest member, and the one I thought hardest about. He drafts the mutual NDA during discovery and the Service Agreement after the client accepts the proposal, written under Philippine law, since that's where I work. He's on my side of the table. I'm the provider in every agreement, every client is the other party, and nothing gets signed without my approval. He drafts to protect the things that matter to a small provider: getting paid, keeping ownership of the work until it's fully paid for, a cap on liability, and a clean way out if things go wrong. The approved scope and quotation get attached to the agreement word for word, and Torney checks that the price and payment schedule in the contract match them exactly. Here's what I was careful about: **Torney is not a lawyer, and nothing he writes has been reviewed by one yet.** So every draft opens with a disclaimer saying exactly that, and every law he cites (the Civil Code, the E-Commerce Act for e-signatures, the IP Code, the Data Privacy Act) is marked UNVERIFIED until someone checks it against the official text. He also flags risky terms instead of quietly "fixing" them. The first flag he raised was on my own rate card: a 2% per month late charge is 24% a year, and Philippine courts can reduce penalties they find unconscionable. He didn't change it. He told me, and the decision stayed mine. Two rules changed the whole pipeline. The NDA is now required during discovery, and the proposal doesn't go out until it's signed by both sides. And it's contract first: Bill can't draft a deposit invoice until the Service Agreement is signed. **Skill:** draft-agreement. ## Bill: invoices that can't drift Bill does deposits, milestone invoices and the final invoice, and keeps a payment tracker. He only bills what the accepted quotation's payment schedule says. Before every draft he checks that what's already invoiced plus this invoice never exceeds the quotation total. Milestone invoices only happen when I say the milestone is done, and he never assumes a payment arrived until I report it. **Skills:** draft-invoice, draft-followup (for overdue invoices). ## Sir Chief: the one who reads everything and touches nothing Sir Chief is the overseer. Ask him "where are we with this client?" and he reads the client folder, checks my inbox and calendar, and reports the stage, what each agent has produced, what's blocked, and which agent should run next. Then he asks "go?" He is completely read-only. He can't send, can't draft an email, can't create a calendar event, can't edit a file, and can't start the other agents. That's the point. I wanted one agent whose only job is to tell me the truth about the state of things, with no ability to "helpfully" change that state while reporting on it. If a file is missing, he says it's missing. If a price is TBD, he quotes it as TBD. ## How a deal actually flows Here's the path a client takes through the team: New lead → **Pia** drafts the questionnaire → I send it, the answers are saved untouched → **Torney** drafts the NDA, both sides e-sign → **Lota** drafts the scope, asks me about gaps, I approve → Lota drafts the quotation → **Pia** assembles the proposal → I review it once and send it → the client accepts → **Torney** drafts the Service Agreement, both sides e-sign → **Bill** drafts the deposit invoice → milestone invoices as work is done → final invoice. **Sir Chief** can be asked at any point where things stand. Every arrow that leaves my machine, every email, every signature request, goes through me. ## What the first test run taught me I ran the whole thing end to end on a test coffee-shop website, with me playing the client from a second email address. It worked, and it was also humbling. The questionnaire went out with *\[meeting date\]* and *\[reply-by date\]* still sitting in it, and a wrong name. The agent had done its job; I hadn't done mine on the review. That's now a written rule: fill the placeholders before sending. The client's reply also exposed everything my standard website package didn't cover: a menu admin screen, help with domain and hosting, a "future-ready" architecture, more than five pages. Without Lota's gap list, those would have quietly become free work. With it, each one became either a priced line, an explicit exclusion, or a question back to the client. And I learned that my Gmail connector can only read one mailbox, so the agents save files and the main session creates the email drafts from them. I treated it as a constraint first, then realized it was exactly the boundary I wanted anyway. As a side project during that test, I also built a spec demo site to show the client the kind of thing we could make: [Cafe Flory](https://cafe-flory.vercel.app/?ref=blog.melardenio.com), a made-up café at the foot of Mt. Pinatubo. It isn't part of the scope and doesn't change a single price. Sir Chief knows that too: he reports a demo site's link, but never treats it as a stage. ## Who's joining next The team isn't finished. A **Follow-up** workflow (a scheduled n8n job on my homelab) will draft polite nudges when a quotation or invoice goes quiet, still for my approval. **Frontend** and **Backend** agents, names still to be decided, will build approved projects from the client-approved scope plus a shared API contract. And the two I'm most looking forward to: a **secretary** and an **admin assistant**, to handle the scheduling, inbox triage and paperwork that sit around the client work. They're not built yet. When they are, they'll follow the same rules as everyone else on the team: one job each, a name, a charter, and a human hand on the switch. Same principle I learned in the control room, really. Let the automation prepare everything. Keep the button for yourself.