Archives

18 months of running Suzune in production. Actual month-by-month API bills, the 3 costs that surprised me, and what breaks first when you scale from 10 users to 100.

Every AI companion app builds characters whose entire universe is the user. I broke that pattern — Suzune's characters have boyfriends, husbands, and crushes the user can't touch. Here's why a little jealousy makes the bot feel alive.

Every AI girlfriend app ships with affinity bars, milestone unlocks, and relationship arcs. I designed all of that for Suzune — then ripped it out before it shipped. Here's why static relationships beat progressing ones.

Every AI chatbot I've used works the same way: pick a character, start a chat. I'm tearing that out of Suzune and replacing it with a world simulator — characters who exist whether you talk to them or not. Here's the design.

After deleting half my AI bot's features, I still had a bloated system prompt. Splitting it into persona.md, rules.md, and nsfw.md — loaded selectively per context — lifted instruction compliance from ~65% to ~88%. Here's the architecture.

I spent months building GM Directives, Proactive Messaging, and an ever-growing system prompt for Suzune. Then I deleted all three. Here's why killing them made the bot dramatically better.

Copy-paste negative prompts for anime AI image generation. Fix hands, anatomy, artifacts, and quality issues. SDXL, Pony, Illustrious — tested in production.

Flux vs SDXL for anime character generation in 2026. LoRA ecosystem, VRAM, NSFW support, RunPod cost — and why SDXL still wins for AI roleplay bots.

How we added I2V to Suzune — provider comparison, prompt engineering for motion, and the two-step workflow that actually works.

Same character, same outfit — completely different photo. A shooting style preset system that controls camera angle, lens, lighting, and mood independently.

How we built an 8-step character creation wizard that lets you pick a different LLM for each field — and why NSFW content needs its own model.

How we generate a single cafe scene with 4 AI characters in their daily outfits using one API call instead of four — cutting image generation costs by 75%.

How we use keyword-triggered lorebooks in Suzune to build living worlds. Covers dynamic context injection, AND/OR keyword matching, timed effects, token budgets, and the architecture behind 27 lorebook files.

Anime AI art beat photorealistic on cost (5–7× cheaper per image), character consistency, and emotional expressiveness. 6 months of side-by-side data from building Suzune.

A deep dive into how we built a LoRA-based character portrait pipeline using Flux on RunPod — from YAML config to wardrobe presets to crash-resilient job tracking.

How we built two proactive messaging systems for Suzune — reach-outs for idle users and follow-ups during active conversations — so characters feel alive even when nobody's typing.

The full technical architecture of Suzune — every subsystem, how they connect, and why each design decision was made. From message arrival to response delivery.

How we give AI characters persistent memory across sessions. Covers tiered compression, character diaries, chat summaries, and the context management that makes characters feel like they actually remember.

How Suzune's two-pass generation works — DeepSeek V3.2 writes the uncensored draft, Claude Haiku polishes the prose, and censorship detection decides which version to serve.

We build AI roleplay bots for a living. Our honest breakdown of every major NSFW AI chatbot platform — features, pricing, and what's worth paying for.

A practical guide to building an AI roleplay bot on Telegram. Covers the tech stack, bot setup, message handling, LLM integration, and the architecture decisions that matter.

How we built Suzune's relationship scoring system — 5 emotional axes, milestone unlocks, and behavior gating that makes AI characters feel like they're genuinely warming up to you.

How we use img2img base image switching to make AI characters transform their appearance — from plain office worker to glamorous queen — while staying recognizably the same person.

The full story of building Suzune — a production NSFW AI roleplay bot. From the first prototype to a multi-model architecture handling thousands of messages. What worked, what didn't, and what I wish I knew earlier.

How we define AI characters using YAML + Markdown in Suzune. Covers persona structure, speech rules, capability flags, and the modular architecture that makes adding new characters fast.

We run 5 LLMs in production for roleplay — not a synthetic benchmark. How DeepSeek, Claude, Gemini, and GLM-5 really compare on quality, NSFW tolerance, and cost.

Practical guide to LLM API providers for NSFW AI. Covers OpenRouter, pricing, content policies, and multi-model routing — without getting banned.

How to build NSFW AI systems without jailbreaks. Covers multi-model routing, censorship detection, fallback chains, and the architecture patterns that make adult content work reliably.

How we engineer system prompts for our production AI roleplay bot. Covers character personas, dynamic context injection, tone enforcement, and the techniques that make AI characters feel alive.

How we keep Suzune under $50/month with multi-model routing, prompt caching, and smart infrastructure choices. Full cost breakdown.

You're building an NSFW project but want Claude's coding help. Here's the architecture pattern we use — separation of concerns at the file level, selective context loading, and a GM Console that lets Claude work on adult projects safely.

Introducing WaifuStack — a technical blog about building AI roleplay bots. What to expect, why we started, and what's coming next.