# Analyzing & improving Once your agent is live, this is how you keep it sharp — check the numbers, read the real conversations, find what's not working, and make targeted fixes without guessing. --- ### See how my agent is doing (the numbers, scored) **What you get:** a scored report showing how often your agent shares the freebie, books calls, pitches too early, and how engaged leads actually are — with green / yellow / red ratings against the goals for your agent's sales style. **When to use it:** run this before making any changes, or any time you feel like the agent "isn't converting" but you're not sure why. **Just say:** > How is [agent name] performing? Give me a full breakdown of the numbers. **What happens:** Claude pulls recent conversation data, calculates key rates (freebie shares, call bookings, early-pitch frequency, engagement quality), and scores each one against the benchmarks for your sales style. You get a plain-language summary with the biggest problems flagged first. **Tips:** - You don't need to specify a date range — Claude uses a sensible recent window by default. - Ask "what does yellow on early pitching mean?" if any score is unclear. --- ### Read my worst or stuck conversations **What you get:** real chat transcripts — the bad ones, the ones that stalled, the unhappy leads, and the ones that almost booked but didn't. **When to use it:** when you want to see *exactly* what's going wrong in the words your agent actually used, not just in abstract numbers. **Just say:** > Show me the worst conversations [agent name] had this week — the ones that stalled or ended badly. **What happens:** Claude fetches recent conversations filtered for the ones that went cold, received negative replies, or never reached the key milestone (freebie sent, call booked). You see the actual message-by-message exchange so you can spot the exact moment things went wrong. **Tips:** - You can also ask for a specific type: "show me the ones that almost booked but didn't" or "show me any angry or frustrated replies." - Reading 3–5 bad conversations is usually enough to see a pattern. --- ### Get a quality check (full QA review) **What you get:** a prioritized list of problems found in real conversations, rated from most serious to least — without changing anything on the live agent. **When to use it:** before a launch, after an agent has been running for a few weeks, or any time you want an outside-eyes review without touching the live setup. **Just say:** > Do a full quality check on [agent name] — review the real conversations against best practices and tell me what's wrong, most serious first. **What happens:** Claude reads a sample of real recent conversations and checks each one against a best-practices checklist (right timing, rapport-building, no premature links, proper objection handling, etc.). You get a ranked list of issues — nothing is changed on the agent. **Tips:** - This is read-only — Claude will not make any changes during a QA review unless you ask separately. - A good cadence is a QA review once a month, or after any big change to your offer. --- ### Fix one specific problem **What you get:** one targeted change made to your live agent, verified to have taken effect. **When to use it:** when you've identified a specific behavior that's hurting results — either from the QA review, from reading conversations, or from your own observation. **Just say:** > Fix [agent name] — it's sending the link before asking any questions. Stop that. **What happens:** Claude confirms what the problem is and tells you what it plans to change before touching anything. Once you approve, it makes one focused change, puts it live, and verifies the update went through correctly. You see a confirmation. Claude always makes changes one at a time so it's easy to know what caused any shift in results. **Tips:** - Be as specific as you can about the problem: "it pitches too early," "it ignores objections," "it never asks about their goal." - Claude will not change anything without confirming with you first. - If the change makes things worse, you can undo it (see below). --- ### Undo a change that made things worse **What you get:** your agent restored to exactly how it was before the last change. **When to use it:** any time a recent fix had unexpected side effects or made results worse. **Just say:** > The last change to [agent name] made things worse — undo it and go back to how it was before. **What happens:** Claude rolls the agent back to the previous version. It confirms the rollback worked and that the agent is now running as it was before the change. No conversations are lost — only the agent's behavior reverts. **Tips:** - You can also be specific: "undo the change from yesterday" or "go back to the version from before we changed the opening message." - If you're unsure what the last change was, ask "what was the most recent change made to [agent name]?" first. --- ### Test my agent before real people see it **What you get:** scripted test conversations — transcripts showing exactly how your agent responds to typical and tricky lead messages — with zero risk to real leads. **When to use it:** before turning an agent on for the first time, after a significant fix, or any time you want to catch problems in a safe sandbox before they hit your real audience. **Just say:** > Test [agent name] — run it through a few example conversations and show me the transcripts. **What happens:** Claude runs the agent through a set of scripted example chats (a typical lead, a skeptical lead, a price-objection lead, etc.) and shows you the full message-by-message transcripts. No real leads are involved. You can read every reply your agent would give and flag anything that looks off. **Tips:** - You can guide the test: "include a lead who says they can't afford it" or "test what happens when someone asks for the price right away." - If you spot a bad response in the transcripts, go straight to "fix one specific problem" to address it. --- ### Compare my agent to similar ones **What you get:** your agent's key numbers placed next to comparable agents so you can see if you're ahead, behind, or on track. **When to use it:** when your numbers feel low (or surprisingly high) and you want context — is this normal for this type of offer and sales style, or is something genuinely off? **Just say:** > How does [agent name] compare to similar agents? Am I ahead or behind on the key numbers? **What happens:** Claude pulls your agent's metrics and benchmarks them against agents running the same sales style and offer type. You see a side-by-side view of the main rates (engagement, freebie delivery, call booking) and a plain-language read on where you stand. **Tips:** - This is a benchmark, not a ranking — the goal is context, not competition. - If you're behind on one number, ask "what usually causes low [metric] in agents like mine?" to get a hypothesis before making any changes. --- ## Common fixes people ask for These are the requests that come up again and again — real, everyday tweaks to a live agent. They all work the same safe way: Claude looks at your real conversations, makes **one** focused change, puts it live, verifies it, and you can undo it if you don't like the result. ### Stop the agent repeating itself **What you get:** an agent that says something once and then waits, instead of re-sending the same message or re-asking after a lead has already replied (or already got the link). **When to use it:** when leads complain the agent is "spammy," or you notice it sending the booking link three times or re-introducing itself on every turn. **Just say:** > [agent name] keeps repeating itself — it sends the link and then keeps messaging. Make it send things once and then wait. **What happens:** Claude finds where the agent loops, adds a clear "stop and wait here" boundary after the key moments (link sent, resource delivered, handed off to a human), puts it live, and verifies. You'll see the agent go quiet at the right points instead of piling on. **Tips:** - Point to the exact moment if you can: "after it sends the WhatsApp number, it shouldn't message again." --- ### Stop sending my link before qualifying **What you get:** an agent that asks a couple of real questions and earns the moment before it drops your booking or checkout link. **When to use it:** when the agent pitches or sends the link in the first message, before it knows anything about the lead — which burns warm leads and lowers your show-up rate. **Just say:** > [agent name] sends my booking link too early. Have it ask about the lead's situation first and only share the link once they're a fit and say yes. **What happens:** Claude adds a qualification gate before the link — the agent learns to ask, listen, get a yes, *then* share. It puts the change live and verifies it. You'll see links going out later in the chat, to leads who actually wanted them. **Tips:** - This pairs well with a QA review — ask Claude to confirm the early-link rate dropped afterward. --- ### Stop interrogating leads with too many questions **What you get:** a smoother conversation that qualifies naturally instead of firing off a checklist of questions before it helps anyone. **When to use it:** the opposite problem — leads go quiet because the agent asks three or four questions in a row before giving any value. Common when qualification is too strict. **Just say:** > [agent name] asks too many questions before it gets to the point — it feels like an interrogation. Loosen it up so it qualifies more naturally. **What happens:** Claude trims the mandatory questions down to the few that truly matter and lets the rest come out of the conversation. It deploys the change and verifies. You'll see fewer leads dropping off in the first couple of messages. **Tips:** - If you know which questions are essential, say so: "it only really needs to know their goal and their budget." --- ### Make it sound more like me **What you get:** an agent whose wording, warmth, and register match your real voice — less robotic, or less slangy, or more on-brand, depending on what you want. **When to use it:** when the agent's tone drifts — too formal, too casual, wrong regional flavor, too many emojis, or just "not how I'd say it." **Just say:** > [agent name] doesn't sound like me — it's too formal and stiff. Make it warmer and more casual, the way I talk. Here are a few of my real messages: [paste 3–5]. **What happens:** Claude studies the examples you paste and adjusts the agent's voice to match, then puts it live and verifies. The cleanest way to fix tone is to show real messages in your voice rather than describe it. **Tips:** - Pasting real messages beats adjectives every time — "warmer" is vague, your actual texts are not. - You can also ban specific words: "never let it say 'amazing' or use the 🙏 emoji." --- ### Stop the agent making things up **What you get:** an agent that never invents a link, phone number, price, or detail it doesn't actually have — it sticks to what's real or stays quiet. **When to use it:** when you catch the agent giving out a made-up WhatsApp number, an old or wrong link, or describing your offer inaccurately. **Just say:** > [agent name] invented a phone number that doesn't exist. Lock it to only the real contact details and never make anything up. **What happens:** Claude pins down the correct, approved details (the one real link, number, or price) and adds a firm rule against inventing anything else, then deploys and verifies. You'll see it stop filling gaps with fiction. **Tips:** - Give Claude the correct value when you ask: "the only real link is [your link] — nothing else." --- ### Update my booking or checkout link **What you get:** every place your agent shares a link updated to the new one, so no lead ever gets the old address. **When to use it:** you switched calendars, moved your checkout, rebranded a URL, or your form changed — and the agent is still handing out the old link. **Just say:** > My booking link changed. Update [agent name] to use [your new link] everywhere it shares it. **What happens:** Claude swaps the link everywhere the agent might send it (not just one spot), puts it live, and verifies the old one is gone. You'll see the new link going out from the next conversation onward. **Tips:** - The same works for a new freebie, a new price, or a new offer — just tell Claude what changed and to what. --- ### Give the agent a better answer to a common objection **What you get:** a stronger, on-brand reply for the objection your leads keep raising — price, time, "is this legit?", "I need to think about it" — instead of a weak or dismissive one. **When to use it:** when you read the conversations and the same pushback keeps stalling deals because the agent fumbles it. **Just say:** > Leads keep telling [agent name] it's too expensive and it just deflects. Give it a better, more empathetic answer that handles the price objection well. **What happens:** Claude adds or rewrites how the agent handles that specific objection — validating the concern first, then reframing — puts it live, and verifies. You'll see the agent hold the conversation through that objection instead of losing the lead. **Tips:** - Tell Claude how *you'd* answer it if you have a line that works — it'll build on your approach. - Ask afterward: "show me a test conversation where someone raises the price objection" to check the new answer. --- ← Back to the [wiki home](index.md)