[{"_1":2,"_68":-5,"_69":-5},"loaderData",{"_3":4,"_38":39,"_66":-5,"_67":-5},"root",{"_5":-5,"_6":-5,"_7":8,"_22":23,"_24":25,"_26":27,"_30":31,"_36":37},"user","shoppingCartId","shoppingCart",{"_9":10,"_11":12,"_13":14,"_15":16,"_17":18,"_19":20,"_21":14},"id","no-cart","count",0,"free",true,"updatedAt","","paymentGateway","STRIPE","cartItems",[],"loading","locale","en","readerTheme","light","constants",{"_28":29},"paypalClientId","AX53Cwz8jR6rVVzzZiZNWWXl0dGagi-GZrjG-S3Ua14E5nrKMn7DxDQ730O4R8LusLPp49ALn6aRVaEy","publishers",{"_32":33,"_34":35},"companies",[],"universities",[],"needsLocaleCookieSync",false,"routes/_store+/_readers+/email_author.$slug/route",{"_40":41,"_65":-5},"book",{"_9":42,"_43":44,"_45":46,"_47":48,"_49":50,"_51":52},"Z2lkOi8vbGVhbnB1Yi9TdG9yZTo6Qm9va1Byb2R1Y3QvNzYxODg","title","Practical Patterns for LLM Pipelines","state","published","contactAuthorMethod","email","aboutText","

These patterns weren't designed. They were extracted -- from the practical constraints of building an LLM-powered personality modeling system on a single developer's budget, where every API call costs real money and \"close enough\" isn't good enough for a model that's supposed to sound like a specific person.

The Gamma Corpus project ingests roughly 28 years of written communication and uses it to power voice-faithful content generation. Building the extraction and generation infrastructure on top of that corpus is where six recurring structural patterns emerged: Tiered Model Cascade, Dual-Channel Retrieval, Corpus-as-Byproduct, Behavioral Archaeology, Behavioral Encoding in Tool Context, and the Two-Analyst Workflow. Each one solved a specific problem -- cost optimization, hallucination prevention, corpus bootstrapping, self-knowledge grounding, habit enforcement, and strategic-tactical task routing.

The patterns are presented in a format loosely inspired by the Gang of Four: intent, motivation, structure, known uses, and trade-offs. Every pattern has at least two concrete implementations with specific cost figures, line counts, and failure modes drawn from production code. The \"when it breaks down\" sections are probably the most valuable parts. Anyone can describe a technique that works; the interesting question is always where the edges are.

","authors",[53],{"_9":54,"_55":56,"_5":57},"Z2lkOi8vbGVhbnB1Yi9TdG9yZTo6UHJvZHVjdEF1dGhvci84MTU4NA","name","Alan McIntyre",{"_58":59},"avatar",{"_60":61,"_62":63,"_64":14},"url","https://secure.gravatar.com/avatar/f3da7510e3e58c054653bf3e13899c25?r=pg&d=mm","styleName","medium","exists","currentUser","routes/_store+/_layout","routes/_store+/_readers+/_layout","actionData","errors"]

Email the Author