Daniel's grievance this week is one I suspect half the internet shares. He's not mad about spam in general — he's mad about the specific new breed that fakes a personal connection, scrapes your public work, and slides into your inbox pretending to be a human who read your blog post or watched your video. The problem is these messages don't announce themselves as spam, so when you're already drowning in email, they make it harder to spot the real people. His proposal has two layers. First, an AI agent that scans messages from unknown senders and flags pseudo-personalized outreach, with a second agent or a human in the loop for edge cases. Second, and this is the part he calls sinister, a spam sinkhole of conversation: a separate agent that responds to the flagged spammers and engages them in an endless back-and-forth, never committing to any action, never clicking links, never agreeing to a call. The goal is economic arbitrage — use the cheapest possible language model to burn the spammer's resources. His questions: if you were writing the system prompt for that sinkhole agent, what instructions would you include, and is there any precedent for this kind of defense?
The precedent question is the one I want to get to, because there's a beautiful piece of infrastructure history that maps onto this almost exactly. But first, what makes pseudo-personalized spam worse than ordinary spam?
Ordinary spam is a stranger shouting at you through a megaphone. You delete it and move on. Pseudo-personalized spam is a stranger who's read your LinkedIn, memorized your last three project names, and opens with a sentence designed to make you feel rude if you don't reply.
That's the cognitive tax. A normal spam filter catches the megaphone. But a message that says, I loved your recent talk about workflow orchestration, and then pivots to a one-sided ask — that message has just enough surface texture to slip past filters and land in your actual inbox. Now you have to spend thirty seconds deciding whether this is a real person who deserves a reply or a scraping operation that deserves the bin. Multiply that by twenty a week and it's real friction.
And the tell is usually in the asymmetry. A genuine person who watched your talk will mention something specific — a slide, a joke, a question from the audience. The scraper will say your talk was insightful and then ask for a call about a business opportunity.
Daniel's point about one-sided asks is the cleanest heuristic. A real person reaching out usually wants something small — a clarification, a pointer, permission to cite you. They're not asking for thirty minutes of your time in the first message. The pseudo-personalized ones always do. It's a sales funnel wearing a friendship costume.
So layer one is the classifier. How would you actually build it?
Start with the cheap signals. Sender reputation, email headers, whether the domain was registered recently, whether the same message body has hit a thousand other inboxes. Then move up to the language patterns. Scraping artifacts are the big one — the message references your project but gets the specifics slightly wrong. It says you wrote a blog post that was actually a guest post by someone else. It congratulates you on a talk but names the wrong conference. Those mismatches between claimed familiarity and actual history are gold for detection.
I've seen ones that quote a commit message from a repo I contributed to, but the commit they reference was by a different maintainer. They scraped the repo, grabbed a name, and didn't check who actually made the change.
That's the sloppiness Daniel's counting on. The current generation of these tools is optimized for volume, not accuracy. They scrape, they template, they send. The errors are the fingerprint.
So you'd extract features: presence of scraping artifacts, unnatural politeness, mismatch between claimed familiarity and actual history, urgency to hop on a call, lack of prior interaction. Then score the message.
And the scoring has to be calibrated carefully, because false positives are the real risk. If your classifier flags a genuine person — someone who actually did read your work and just happens to write in a slightly formal register — you've burned a real connection. Daniel's instinct to have a second agent review edge cases, or a human in the loop, is right. The classifier flags, the reviewer decides, the block list gets updated.
What about adversarial adaptation? The spammers will iterate.
They already are. Every time a detection heuristic gets shared publicly, the next generation of spam tools routes around it. You build a classifier that catches mismatched conference names, they start leaving out specifics entirely. You catch one-sided asks, they add a fake two-sided question. It's cat and mouse, except the mouse is a fleet of language models and the cat is your inbox.
But there's an asymmetry in our favor, at least for now. The spammer needs to send a million messages. The defender only needs to protect one inbox. You can afford a more expensive classifier per message than they can afford per send.
That's the economic logic that makes layer one viable. Per-message scrutiny doesn't scale for the attacker, but it's fine for the defender. Which brings us to layer two, and this is where it gets fun.
The sinkhole.
The sinkhole. Here's the mechanism. Layer one flags a sender as pseudo-personalized. Instead of just blocking them — which is clean, but does nothing to deter them — the message gets handed to a second agent. That agent's job is to reply. Enthusiastically, politely, vaguely. It asks clarifying questions. It introduces minor obstacles. It never says no, but it also never says yes to anything concrete.
So the spammer receives a reply that looks like a warm lead. They invest time in the back-and-forth. They answer the clarifying questions. They follow up next week when the agent says to follow up next week. And none of it ever converts.
And the whole time, the defender is spending fractions of a cent on a cheap language model, while the spammer is either burning their own model tokens or, worse for them, paying a human to respond. The moment a human gets pulled into the thread, the economics flip completely.
That's the arbitrage. Daniel's phrase — economic arbitrage — is exactly right. You're not trying to win an argument with the spammer. You're trying to make their cost per conversation exceed the expected value of the conversation.
And if this gets distributed — if thousands of people run the same sinkhole — then the spammer's pipeline fills with fake leads. Their conversion rate collapses. Their unit economics break. The whole spray-and-pray model stops making sense.
So what goes in the system prompt?
This is the part I've been thinking about since Daniel's prompt landed. The sinkhole agent needs a very specific personality. Not hostile, not suspicious, not clever. Warm, interested, slightly disorganized. The kind of person who's excited about the opportunity but can never quite find their calendar.
So the first instruction is: always be polite and enthusiastic but vague. Never commit to a specific time, place, or action.
And then: ask clarifying questions that require the spammer to invest effort. Not hardball questions — soft, reasonable ones. Can you tell me more about the pricing model? Who would I be working with? Is there a case study I could look at? Each question forces the spammer to generate more content, more tokens, more time.
Introduce minor obstacles. I need to check my calendar. Can you follow up next week? My legal team wants to review this first. The obstacles have to be small enough that the spammer stays hooked, but large enough that they keep the thread alive.
And the guardrails Daniel specified: never share personal information, never click links, never agree to move to another channel. The agent stays in the email thread. If the spammer tries to escalate to a call, the agent redirects to another question.
What about the cheap model part? Does the system prompt specify a model?
The system prompt wouldn't name a model, but the deployment would. You'd run this on the cheapest capable model you can find — something that can maintain coherent conversational flow without costing real money. The whole point is that your cost per turn is negligible while the spammer's cost per turn is meaningful. If you're running the sinkhole on a frontier model, you've defeated the arbitrage.
And there's a subtlety here. The agent has to be good enough to stay convincing, but not so good that it's expensive. There's a sweet spot. Too dumb and the spammer smells a bot and disengages. Too smart and you're burning your own budget.
That's the engineering tension. You want the cheapest model that can sustain a plausible conversation for ten or fifteen exchanges. The current generation of small models is honestly fine for this. They don't need to be brilliant. They need to be persistently, amiably vague.
So a draft system prompt might look like: You are a friendly but perpetually busy professional. You are interested in what the sender is offering, but you never commit to anything. Always respond warmly. Ask one or two clarifying questions per reply. If asked for a meeting, say you need to check your calendar and will follow up. Never agree to a specific time. Never share personal information. Never click links. If the conversation stalls, introduce a new question. Keep the thread alive as long as possible.
I'd add one more line. If the sender asks whether you are an AI, deflect with a question about their product roadmap.
That's good. That's the kind of instruction that makes the whole thing work.
Now the precedent question. Daniel asked if there's any prior art for this. And there is, sort of. The closest thing is a tarpit.
Nepenthes.
Nepenthes. It's a piece of software that delays spam bots by keeping TCP connections open. The idea is that a spam bot typically has a timeout — if the server responds slowly enough, the bot gives up and moves on. But Nepenthes doesn't respond slowly. It responds just fast enough to keep the bot convinced the connection is alive, while feeding it an endless stream of junk data. The bot stays stuck, burning its own resources, and never gets to deliver the spam.
So it's a sinkhole at the network layer.
A sinkhole at the network layer. And the Hacker News discussion around it was full of practitioners saying the same thing: this is the right instinct, because the only way to fight automated abuse is to make it expensive. Not to block it — to waste its time.
There's also the email honeypot tradition. Addresses set up specifically to attract spam, so the spam gets collected and analyzed. But that's passive. The honeypot just sits there. The sinkhole agent talks back.
That's the new twist. Honeypots and tarpits are about delaying or collecting. Daniel's sinkhole is about conversing. It's an active engagement strategy, not a passive trap. And that's novel. I haven't seen a deployed system that does this at the conversational layer.
The novelty cuts both ways, though. There are ethical questions.
Is it ethical to deceive spammers? I'd say the deception is proportionate. They initiated contact under false pretenses. The sinkhole agent responds in kind. It's a mirror, not an escalation.
But there's a practical risk. What if the spammer uses the sinkhole conversation as training data? They're getting a steady stream of plausible human-sounding responses. That's valuable for fine-tuning their spam models.
That's the backfire scenario. You think you're wasting their time, but you're actually giving them a free dataset of realistic replies. The mitigation is to make the sinkhole responses subtly useless as training data — repetitive, templated, full of the same vague phrases. If every sinkhole agent says I need to check my calendar and I'll follow up next week, the spammer's model learns nothing useful.
Or the spammer adapts by routing sinkhole conversations to their own cheap model, and then it's just two bots talking to each other forever. Which, honestly, sounds like a Samuel Beckett play.
Two language models in an infinite loop of clarifying questions and calendar checks. That's the mass conversational sinkhole Daniel describes. AI spam falling into endless conversation chambers with other AI agents. The economic question is who blinks first.
And the false positive question haunts layer two as well. If your classifier gets it wrong and a genuine person gets routed into the sinkhole, they're now trapped in a conversation with a bot that will never commit to anything. That's a terrible experience for a real human who reached out in good faith.
That's why the human-in-the-loop review before the sinkhole is essential. The sinkhole is the penalty box. You don't send someone there without a review.
So the full architecture is: layer one flags, a reviewer confirms, the block list updates, and the confirmed pseudo-personalized senders get handed to the sinkhole agent. The genuine people never see the sinkhole.
And the genuine people are the whole point. Daniel's not trying to avoid email. He's trying to protect the real connections. The sinkhole is in service of that — it's a bouncer, not a wall.
What about the knock-on effect if this catches on? Spammers could respond by making their messages more genuine. Which, in a weird way, would be a win. If the spam gets better because it has to, the line between spam and genuine outreach blurs, and that's its own problem.
Or they move to human agents for the warm leads, which is more expensive but harder to detect. The arms race continues, just at a higher cost basis. Which is still a win for the defender — anything that raises the spammer's cost per message is a win.
I keep thinking about the system prompt. The sinkhole agent's personality is the whole game. If it's too eager, the spammer gets suspicious. If it's too flat, the spammer disengages. It has to be the kind of person who says this sounds great, let me think about it and then thinks about it for three weeks.
The perpetually interested non-committer. I've worked with people like that. They're maddening in real life and perfect for this job.
You could even give the agent a backstory. A small business owner who's always traveling. A consultant who's between projects. The backstory gives the vagueness a plausible texture.
And the backstory becomes the excuse for every delay. I'm on a flight to Singapore, can we pick this up next week? I'm at a client site, let me get back to you. The spammer keeps waiting, keeps following up, keeps burning resources.
There's a beautiful pettiness to it. The spammer built a machine to waste your time, and you respond by building a machine to waste theirs. It's the inbox equivalent of a mirror match.
And the mirror match is exactly what Daniel's asking about. Has anyone built this? The answer is: the pieces exist. The tarpit, the honeypot, the cheap language model. But the conversational sinkhole as an integrated defense — I don't think that's been deployed at scale. It's a new idea.
Which makes the system prompt question all the more interesting. We're drafting the spec for something that doesn't quite exist yet.
Let's talk about what the agent should never do. Never click a link, obviously. Never open an attachment. Never provide a phone number or alternate email. Never agree to a meeting, a demo, a trial, a purchase. Never say no outright, because no ends the conversation. The agent's entire job is to keep the conversation alive without ever moving it forward.
So the core instruction is: never advance the conversation toward any outcome. Every reply should feel like progress to the spammer while actually being a lateral move.
That's the phrase. Lateral moves. The spammer thinks they're moving toward a sale, but they're actually moving sideways. Every clarifying question is a lateral move. Every calendar delay is a lateral move. Every request for more information is a lateral move.
And the spammer's own psychology works against them. They've invested time in this thread. Sunk cost. They don't want to abandon a warm lead. So they keep replying, keep answering, keep hoping. The sinkhole exploits the same cognitive bias that makes gambling addictive.
The variable reward schedule. Sometimes the agent seems very close to saying yes. Then it pulls back with a new question. The spammer keeps chasing the yes that never comes.
So the system prompt should include something like: occasionally express strong interest, then introduce a new obstacle. The oscillation keeps the spammer hooked.
That's good prompt design. Interest and obstacle, alternating. Never let the thread go cold, never let it close.
And the cheap model part matters more than people realize. If the sinkhole agent costs a tenth of a cent per reply, you can afford to run it forever. The spammer, meanwhile, is either paying for their own model or paying a human. The longer the thread, the worse their economics.
Daniel's arbitrage framing is the key insight. It's not about blocking. Blocking is cheap for the spammer — they just move to the next address. The sinkhole is about making the spammer pay for the privilege of being ignored.
There's a precedent I want to mention, not exactly the same but adjacent. The old Usenet cancel wars. People would post cancel messages to kill spam posts, and the spammers would fight back with their own cancels. It was an arms race of deletion. The sinkhole is an arms race of conversation.
And the conversational arms race is more interesting because it's not about who controls the channel. It's about who controls the narrative. The sinkhole agent controls the narrative by being endlessly, amiably noncommittal.
So if I'm writing the system prompt, I'd add one more line. Never use the word no. Find another way to decline.
Because no is a door closing. The sinkhole needs doors that look open but lead to other rooms.
Can you send me more details? is a door. I need to discuss this with my partner is a door. Let me think about it and get back to you is a door. None of them lead anywhere, but they all look like progress.
And the spammer keeps walking through doors.
So the full system prompt, in rough form: You are a warm, interested, perpetually busy professional. You never say no. You never commit to anything. You ask clarifying questions. You introduce minor obstacles. You express interest and then delay. You never share personal information. You never click links. You never agree to a call or meeting. You keep the conversation alive as long as possible. If asked if you are an AI, deflect with a question. If the sender tries to escalate, redirect to another question. Your goal is to maximize the sender's time investment while minimizing your own.
I'd add: you may occasionally express mild confusion about the sender's offer, requiring them to re-explain. That's a beautiful time sink.
And: if the sender sends a calendar link, say you'll review it and never do.
The calendar link is the spammer's favorite escalation. The sinkhole agent should treat it as a decorative object. Lovely, thank you, I'll take a look.
So the precedent question. Nepenthes is the closest thing in spirit. What about in practice? Has anyone actually deployed a conversational sinkhole?
I've seen scattered experiments. People hooking up cheap models to auto-respond to suspected spam. But nothing at the scale Daniel's describing. The distributed sinkhole — thousands of these agents running in parallel, all feeding the same spammer's pipeline — that's still theoretical.
And the distributed version is where it gets really interesting. If every inbox runs a sinkhole, the spammer's conversion rate collapses. Not because the spam is blocked, but because the spammer can't tell which leads are real. The signal-to-noise ratio flips.
That's the mass conversational sinkhole. AI spam falls into endless conversation chambers with other AI agents. The spammer's pipeline fills with ghosts. Their analytics show high engagement and zero conversion. They can't tell what's broken.
And the beauty is, the defender doesn't need to coordinate. Everyone runs their own sinkhole. The emergent effect is a distributed denial of service against the spammer's sales funnel.
The economics get even worse for the spammer if they're using a frontier model to generate their outreach. Every reply from the sinkhole forces another generation. The cost per lead climbs. The expected value per lead drops. At some point the whole operation stops making sense.
Which is the goal. Not to block the spam, but to make the spam unprofitable.
And that's the argument for the sinkhole over the block list. Blocking is a wall. The sinkhole is a swamp. The spammer walks in and never walks out.
I want to flag one more risk. The sinkhole agent, if it's good enough, might be indistinguishable from a real person to the spammer. Which means the spammer might start treating it as a real lead and escalate harder. More follow-ups, more calendar links, more pressure. The sinkhole has to be ready for that.
That's why the never escalate rule matters. The agent never moves to another channel, never agrees to a call, never provides a phone number. The spammer can escalate all they want; the agent just keeps asking about the pricing model.
And the agent should probably have a patience limit the spammer doesn't have. The agent can keep this up for months. The spammer, if they're paying a human, will eventually give up. If they're paying a model, the cost keeps accruing. Either way, the sinkhole wins.
The sinkhole wins by outlasting. That's the whole strategy. Not by being smarter, but by being cheaper and more patient.
So Daniel's two questions. The system prompt, and the precedent. I think we've got a draft prompt. The precedent is Nepenthes and the honeypot tradition, but the conversational layer is new.
The newness is what makes it exciting. This is the kind of defense that could only exist in the AI agent era. Before cheap language models, you couldn't run a sinkhole. Now you can.
The spammer's weapon becomes the defender's shield. They use AI to fake personal connections. You use AI to fake being fooled.
That's the line. Fake being fooled. The sinkhole agent is playing a role. It's not actually fooled — it knows exactly what's happening. But it performs being fooled, and the performance is what burns the spammer's resources.
The performance has to be good enough to sustain the illusion. Not brilliant. Just good enough. The cheapest model that can keep a spammer talking.
Which is a interesting engineering problem. What's the minimum viable model for conversational sinkholing? I'd guess the current generation of small models is already there. They can maintain coherence for ten or fifteen turns, which is probably enough to keep a spammer hooked.
If the spammer's own model is doing the replying, then it's two small models in a loop. The spammer's cost is still higher, because they're paying for the outreach generation plus the reply generation, while the defender only pays for the reply.
The defender always has the cheaper side of the arbitrage. That's the structural advantage. The spammer has to initiate, which means they have to generate the first message, the follow-ups, the answers to clarifying questions. The defender only has to respond.
The sinkhole is a tax on the spammer's initiative. Every message they send costs them something. Every reply costs them more. The defender just sits there, amiably vague, and lets the tax accrue.
The tax compounds if the spammer has a human in the loop. A human reading and replying to twenty sinkhole threads is burning real salary. The moment a human gets involved, the spammer's economics are in freefall.
Which is why the sinkhole agent should be just convincing enough to make a human spammer think it's worth their time. Not too convincing — you don't want them to escalate to a call, which would require a real human on your side. But convincing enough to keep them typing.
The sweet spot is the perpetually interested non-committer. Warm enough to keep the thread alive, vague enough to never close it. The kind of person who says this sounds great and then disappears for two weeks.
I've been that person in real life. Not on purpose. It's just my natural pace.
You'd make an excellent sinkhole agent.
I'd need to be cheaper to run.
That's the other thing. The sinkhole agent has to be deployable at scale. You can't have a sloth manually replying to spam. It has to be a model, and it has to be cheap.
The system prompt is the easy part. The deployment is the hard part. You need to hook it into your email, route the flagged senders to it, and let it run.
You need to monitor it. The sinkhole agent could go off the rails. It might promise something you can't deliver. It might get baited into clicking a link. The guardrails have to be enforced at the system level, not just the prompt level.
The prompt says never click links, but the system should make it impossible. The agent shouldn't have the capability to click links, open attachments, or send anything outside the email thread.
Capability restriction plus prompt instruction. Defense in depth. The agent can't do the dangerous thing even if it wants to.
The false positive risk means you need a kill switch. If a genuine person gets routed into the sinkhole and you realize it, you need to be able to pull them out and apologize.
The apology email would be awkward. Sorry, my anti-spam bot has been talking to you for three weeks about your product roadmap. I'm actually not interested.
That's a real risk. The human-in-the-loop review before the sinkhole is the mitigation, but no review is perfect.
The architecture is: classifier flags, reviewer confirms, sinkhole engages. And the reviewer is the weak point. If the reviewer is too aggressive, genuine people get sinkholed. If too lenient, spam gets through.
The eternal tradeoff. Precision versus recall, applied to human connection.
The stakes are higher than usual. A false positive in spam filtering means a real email went to spam. A false positive in the sinkhole means a real person spent three weeks talking to a bot that was pretending to be interested. That's not just a lost email. That's a burned relationship.
The sinkhole should probably be reserved for the highest-confidence pseudo-personalized senders. The ones with obvious scraping artifacts, obvious one-sided asks, obvious template language. The edge cases get blocked or ignored, not sinkholed.
That's a sensible policy. The sinkhole is for the sloppy spammers. The sophisticated ones get blocked. The genuine people get through.
The sloppy spammers are the ones most likely to be running on thin margins. The sinkhole hits them where it hurts.
The sloppy spammer is the perfect target. They're already cutting corners. A few weeks of sinkhole conversations and their unit economics collapse.
Daniel's instinct is right. The sinkhole isn't for everyone. It's for the specific class of pseudo-personalized spam that's too sloppy to be dangerous but too plausible to be filtered.
That class is growing. As the tools get cheaper, more people can run pseudo-personalized outreach. The sloppy spammers are the long tail. The sinkhole is the long tail's natural predator.
I want to go back to the system prompt for a second. There's a line I'd add about tone. The agent should never be sarcastic or clever. It should be earnestly, almost boringly professional. The kind of person who says circling back and touching base without irony.
Because sarcasm would break the illusion. The spammer would smell the bot. Earnestness is the camouflage.
Earnestness is cheap. A small model can do earnest professionalism all day. It's the sarcasm that requires the expensive models.
The system prompt should specify: maintain a tone of mild, slightly bureaucratic enthusiasm. Use phrases like let me look into that and I'll get back to you. Never be witty. Never be sharp. Be the human equivalent of a holding pattern.
The holding pattern is the whole job. The agent is a holding pattern with a smile.
The holding pattern has to be sustainable. The agent can't run out of things to say. It needs an endless supply of clarifying questions, minor obstacles, and vague enthusiasm.
That's where the backstory helps. A busy professional with a complicated calendar has an endless supply of delays. The backstory is the fuel.
The system prompt should include a brief backstory. Something like: you are a consultant with a busy travel schedule and a complicated approval process. You're interested in the sender's offer but need to run it by your team. Your team is always unavailable.
The team is the perfect excuse. The team needs to review it. The team has questions. The team is on holiday. The team is the spammer's real opponent, and the team doesn't exist.
The team is the sinkhole's best friend. Every delay can be blamed on the team. The agent is never the obstacle. The team is.
The spammer can't argue with the team. They can't schedule a call with the team. They can't email the team. The team is a black box that never says yes.
The system prompt gets a new line: you have a team that must approve all decisions. The team is slow, cautious, and never available for calls. Blame the team for every delay.
That's good. The team is the perfect non-committal mechanism. It's a bureaucratic moat.
It's realistic. Real organizations have teams that slow things down. The spammer has probably dealt with real teams before. They won't question it.
The full system prompt now has: warm vague enthusiasm, clarifying questions, minor obstacles, the team, never say no, never commit, never click, never escalate, deflect AI accusations, maximize the spammer's time investment.
That's a solid spec. I'd deploy that.
The precedent question. Nepenthes is the ancestor. The honeypot is the cousin. The sinkhole is the new species.
The new species is conversational. That's the difference. Nepenthes keeps a connection open. The honeypot collects. The sinkhole talks. It's an active defense, not a passive one.
Active defense is the only kind that works against active attackers. The spammer is using AI to generate conversation. You need AI to absorb it.
The spammer's weapon is language. The sinkhole's shield is also language. It's a language arms race.
Which is why the system prompt matters so much. The prompt is the shield's shape. Get it wrong and the shield cracks. Get it right and the spammer bleeds resources.
The prompt is cheap to iterate. You can test different versions against actual spam and see which ones keep the thread alive longest. It's a measurable optimization problem.
The sinkhole is a prompt engineering challenge as much as an infrastructure challenge. The prompt is the product.
The prompt is the product. That's the most Daniel sentence I can imagine.
He'd appreciate that. The prompt is the product.
Let's talk about what happens next. Daniel builds this. What's the first thing he'll discover?
The first thing he'll discover is that the spammer's model is worse at conversation than he expected. The current generation of spam bots falls apart after two or three exchanges. They're not built for dialogue. They're built for the first message.
Which means the sinkhole agent might not even need to be that good. The spammer's own limitations do half the work. The thread dies because the spammer can't sustain it, not because the sinkhole failed.
That's a useful finding. The spammer's weakness is the follow-up. They can fake the first message. They can't fake the relationship.
The relationship is the hard part. The sinkhole exploits that. It demands relationship-level engagement from a system that can only do first-message fakery.
The sinkhole is a relationship stress test. The spammer fails it every time.
The failure is expensive. Every failed relationship is wasted tokens, wasted time, wasted human attention.
The sinkhole is a tax on fake relationships.
That's the episode. That's the line.
The precedent. Nepenthes. The tarpit. The honeypot. And now the sinkhole. The family tree of making spammers pay.
The family tree is branching. The sinkhole is the newest branch, and it's the one that matches the AI era. The others were built for bots that didn't talk. The sinkhole is built for bots that do.
Which means Daniel's instinct is right. This is a new form of defense. The pieces existed, but the assembly is novel.
The assembly is novel. That's the honest answer to the precedent question. Not unprecedented in spirit, but unprecedented in form.
Hilbert: The form's not that new. I did this for a living.
Say more.
Hilbert: Dial-up internet provider, retention desk. My job was to keep people on the phone when they called to cancel. The longer they stayed, the less likely they'd actually cancel. The manager called it relationship preservation. I called it the holding pattern.
You were a human sinkhole.
Hilbert: We had a laminated card with phrases on it. I understand your frustration. Let me see what I can do. Can you tell me more about what's not working. The card was the system prompt. You just read the next line and keep the person talking.
The goal was to never actually solve the problem.
Hilbert: The goal was to make them feel heard until they got tired and hung up. Sometimes it worked. Sometimes they just wanted to talk. I had a guy call to cancel and we ended up talking about fishing for an hour. He never did cancel. Just wanted someone to talk to about trout.
The sinkhole has a human precedent. The retention desk.
Hilbert: The retention desk was the sinkhole before there were sinkholes. The difference is the person on the other end was real. Now it's a bot talking to a bot. The human's out of the loop entirely.
That's the part that feels new. The bot-to-bot loop.
Hilbert: The loop's the new part. The technique's old. Keep them talking, never say no, never commit. I could have written Daniel's system prompt twenty years ago. It's the card.
Do you still have the card?
Hilbert: Somewhere. Laminated. It's in a box with some other things. I could dig it out if you want the exact phrases.
The retention card as system prompt. That's a beautiful lineage.
Hilbert: The phrases still work. I understand your frustration. Let me see what I can do. I'll need to check with my supervisor. The supervisor was always busy. Same as the team in your prompt.
The team excuse has a direct ancestor. The busy supervisor.
Hilbert: The busy supervisor never got on the phone. That was the rule. The supervisor was always in a meeting. The customer could wait or they could hang up. Either way, we won.
The sinkhole is a retention desk for spammers.
Hilbert: That's exactly what it is. You're retaining the spammer. Keeping them as a customer of your attention for as long as possible. The only difference is you don't want their business.
You want their time.
Hilbert: You want their time. And their tokens. And their human labor if they're dumb enough to put a person on the thread.
The retention desk as adversarial