Glossary
Twelve domain terms for local-first prompt management. Each entry carries a fragment anchor for direct linking.
Adaptive Thinking
DE: Adaptives Denken
A 2026 capability in Claude 4.6+ models: the model dynamically decides when and how deeply to reason, calibrated by query complexity. API users control depth via the effort parameter (low/medium/high). OpenAI's GPT-5+ has an equivalent reasoning_effort parameter. Replaces older explicit "think step by step" triggers for these models. See Chapter 8.
Knowledge Cutoff Drift
DE: Aktualitäts-DriftGap between a model's training cutoff date and the present. The model knows no events, tool versions, prices, or legal changes after that date unless web search is active. Especially impacts tool comparisons, market data, and compliance topics. Fix: explicitly request web search or verify facts externally.
Anchor Bias
DE: AnkereffektWhen too many examples in a prompt cause the model to copy their style or structure rather than reasoning toward the best solution for the specific case. The fix is fewer, better examples — usually one. Wikipedia
Append-Only Prompt Log
DE: Append-Only-Prompt-LogStorage model in which every change creates a new revision; existing revisions are never overwritten.
Job Prompt
DE: Auftrags-PromptThe per-task instructions sent after the System Prompt. Contains only the variable data for that request — company, topic, format. Defined more fully in Chapter 9.
Data Processing Agreement
DE: AuftragsverarbeitungsvertragA Data Processing Agreement (DPA) is the contract required under GDPR Art. 28 between controller and processor. For DACH companies using US LLM APIs, the DPA is a mandatory checkpoint: without a signed DPA, the tool does not process personal data in a GDPR-compliant way.
AWS Bedrock
DE: AWS BedrockAWS Bedrock is Amazon's managed service for foundation models. Anthropic Claude is available in the Frankfurt region (eu-central-1) among others, making Bedrock the strongest EU data residency option among US LLM providers.
Context Engineering
DE: Context EngineeringThe discipline of getting the right content into a language model's context window at the right time, in the right form, order, and amount. Context engineering goes beyond prompt engineering because it manages the entire information flow, not just a single instruction. Anthropic framed the operational definition in 2025 alongside Claude Sonnet 4.5.
Chain-of-Thought (CoT)
DE: Gedankenkette (Chain-of-Thought)A prompting technique where you tell the model to reason step by step before answering. Measurably improves quality on multi-step or reasoning-heavy tasks. Introduced by Wei et al. (2022); Kojima et al. (2022) showed even the zero-shot trigger "Let's think step by step" produces gains. Anthropic docs · OpenAI docs
ChatGPT
DE: ChatGPTA conversational AI assistant developed by OpenAI, released November 2022. Runs on OpenAI's GPT model series (GPT-5 in 2026). One product inside the broader AI category — direct competitors are Claude, Gemini, Mistral and Perplexity. Works in German with the same fluency as English. Wikipedia
Context Rot
DE: Kontext-VerfallThe general degradation of LLM output quality as the input context grows. Includes Lost-in-the-Middle and other forms of dilution. The Chroma 2025 technical report tested 18 frontier models — every one degrades as input length increases. A 200K-token-window model can show significant degradation at 50K. (Industry technical report, not peer-reviewed.)
Context Window
DE: KontextfensterThe total amount of text a model can "see" in one conversation, measured in tokens. Once exceeded, earlier content disappears from the model's view. Anthropic docs
Cross-Site Request Forgery (CSRF)
DE: Cross-Site Request Forgery (CSRF)Cross-Site Request Forgery is a vulnerability where an authenticated session is misused for unintended actions (e.g. money transfers, data changes). AI-generated code often lacks CSRF protection mechanisms — a Tenzai study (December 2025) found 0% CSRF protection in 15 tested vibe-coding apps.
CTA (Call to Action)
DE: CTA (Call to Action)An instruction asking the reader to take a next step. Effective US B2B CTAs are concrete and value-driven: "Book a 15-minute call" beats "Contact us". Wikipedia
Custom GPT
DE: Custom GPTA Custom GPT is a personalized ChatGPT configuration with a persistent system prompt, uploaded knowledge files, and optional API actions. Requires a ChatGPT Plus subscription. Unlike standard ChatGPT, a Custom GPT applies its stored behavioral rules to every request instead of starting from scratch.
Deep Learning
DE: Deep LearningA branch of Machine Learning that uses neural networks with many layers ("deep"). The technology behind today's LLMs, image generators, and speech systems. Practical breakthrough since 2012; mainstream since the transformer architecture in 2017. Wikipedia
Diff-Based Prompt Comparison
DE: Diff-basierter Prompt-VergleichLine- or span-level comparison of two prompt versions, marking insertions, deletions and semantic anchors.
GDPR
DE: DSGVOThe GDPR (General Data Protection Regulation) is the EU regulation on personal data protection, in force since 25 May 2018. Relevant for AI tools: any processing of personal data by US LLMs requires a Data Processing Agreement, a legal basis, and transparency toward data subjects.
Editorial Review Mark
DE: Redaktionelle Review-MarkierungMarker for a prompt revision indicating human review; consistent with EU AI Act Art. 50 disclosure.
Retrieval Augmented Generation (RAG)
DE: Retrieval Augmented Generation (RAG)A technique that searches external knowledge stores at runtime and loads relevant matches into the context window before the model answers. This gives the model current or private knowledge without retraining. RAG is one tool within context engineering, not a replacement for it.
Engineered Prompt
DE: Strukturierter PromptStructured, reusable prompt template with explicit context, defined goal, and specified output format. Unlike spontaneous Quick-Prompts, an Engineered Prompt is carefully formulated once and stored as a template for recurring tasks.
EU AI Act
DE: EU AI ActThe EU AI Act is the EU regulation governing AI systems, in force since August 2024 with phased application dates. GPAI obligations (transparency and documentation) apply to new general-purpose AI models from 2 August 2025. Models that were on the market before this date have a transition period until 2 August 2027.
Few-shot / Zero-shot
DE: Few-shot / Zero-shotFew-shot: providing one or more examples in the prompt to show the model what you expect. Zero-shot: giving a task with no examples. Trade-off: more examples can reduce vagueness but introduce Anchor Bias.
Frontier Model
DE: Frontier-ModellA frontier model is a high-performance language model at the cutting edge of AI research. Training costs in 2026 range from several hundred million to over a billion dollars (source: Anthropic CEO Dario Amodei). Current frontier models in 2026: GPT-5.4, Claude Opus 4.7, Gemini 3.1 Pro, Mistral Large, Llama 4.
Generative AI
DE: Generative KISystems that produce new content — text, images, code, audio, video — instead of merely classifying existing data. Text-generative: ChatGPT, Claude, Gemini. Image-generative: Midjourney, DALL-E, Stable Diffusion. Output is generated token by token or pixel by pixel from patterns learned in training. Wikipedia
GPAI
DE: GPAIGPAI (General-Purpose AI) refers in the EU AI Act to versatile AI models that can be adapted for different tasks. Providers of such models have been subject to binding transparency and documentation obligations under Art. 53 of the AI Act since 2 August 2025.
GPT Builder
DE: GPT BuilderThe GPT Builder is OpenAI's no-code interface in ChatGPT Plus for configuring Custom GPTs. Two paths: via chat dialog (the tool generates the system prompt automatically) or via manual instruction editing in the configuration tab. For production workflows, the manual path delivers more precise results.
Hallucination
DE: Halluzination
When a model produces confident-sounding content that is factually wrong or invented. Prevented by supplying facts explicitly and using [INFO NEEDED] flags. Wikipedia
AI (Artificial Intelligence)
DE: KI (Künstliche Intelligenz)The umbrella term for systems that perform tasks which used to require human thinking. In 2026 the dominant flavour is pattern recognition in large datasets. Split into narrow AI (today, specialised), general AI (theoretical, human-level), and superintelligence (speculation). Everything practical in this guide is narrow AI for text. Wikipedia
Confabulation
DE: KonfabulationSpecific form of AI hallucination where the language model fills an information gap with plausible-sounding but factually wrong details. Term borrowed from neuropsychology and applied to LLMs. Typical triggers: missing training data on proper names, dates, or specialized terminology.
Context Window Pressure
DE: Kontextfenster-DruckGrowing inputs reduce answer quality well before the context window is technically full.
LLM (Large Language Model)
DE: LLM (Large Language Model)A language model trained on enormous text collections that answers requests in natural-language text. "Large" refers to billions to trillions of parameters and several hundred billion training words. 2026 frontier models: GPT-5 (OpenAI), Claude 4.6 (Anthropic), Gemini 2.5 (Google), Mistral Large, Llama 4 (Meta, open weights). Wikipedia
LLMOps
DE: LLMOpsOperational framework for production-grade Large Language Model pipelines. Includes versioning, staging, approval gates, monitoring, and rollback mechanisms — DevOps practices applied to LLM workflows.
Lost-in-the-Middle
DE: Lost-in-the-MiddleEmpirically documented phenomenon: LLMs attend more strongly to the beginning and end of a prompt. Content placed centrally shows 30%+ accuracy drops on retrieval tasks. Liu et al., TACL 2024. MIT 2025 traced the cause to RoPE decay — the effect is structural, not training-related. Liu et al. 2024 (arxiv:2307.03172)
Machine Learning
DE: Maschinelles LernenThe broader field of algorithms that improve from data instead of being programmed rule by rule. Encompasses classical methods (decision trees, support-vector machines) and Deep Learning. The basis of all modern AI including LLMs. Wikipedia
Memory
DE: MemoryA feature in Claude Projects, ChatGPT custom GPTs, and the Claude Developer Platform Memory Tool that lets the model store and retrieve information across separate conversations. Build up project context over time without re-pasting it each session. Anthropic announcement
Meta Title / Meta Description
DE: Meta-Titel / Meta-BeschreibungTwo SEO output fields. Meta Title (≤55 characters) is the headline shown in Google search results and browser tabs — keyword first, then a pipe separator, then a value promise. Meta Description (≤155 characters) is the snippet below the title in search results — concrete benefit plus action prompt, no emoji. SEO-hygiene basics; not a ranking factor for the description itself, but high-impact on click-through rate.
Model Council
DE: Model CouncilModel Council is a feature in Perplexity Pro and Max plans, available since early 2026. Queries are sent in parallel to multiple frontier models (e.g. GPT, Claude, Gemini), and results are merged. Reduces hallucinations and makes model comparison visible directly in the output.
Neural Network
DE: Neuronales NetzA mathematical structure loosely inspired by brain cells, where signals flow through weighted connections between layers of "neurons". Training adjusts these weights until the network's output matches the desired output. The architecture behind Deep Learning and all current LLMs. Wikipedia
Output Drift
DE: Ausgabe-DriftGradual degradation of model responses through uncontrolled prompt changes or silent model updates. Occurs frequently in teams without systematic prompt versioning and often goes undetected until serious quality issues arise. Closely related to Prompt-Drift but focused on the output side.
PAS Framework
DE: PAS-FrameworkProblem-Agitate-Solve. A copywriting structure: name the buyer's pain, deepen the stakes of inaction, then present your solution as the direct answer. Used directly in S-2; informs voice across other Sales/Content templates.
Persona Slot
DE: Persona-SlotVariable slot dedicated to role or voice descriptions, swappable without changing the task structure.
Prompt
DE: PromptThe instruction sent to an AI model to get a specific answer. In its simplest form a single question; in practice a multi-paragraph instruction with role, context, target audience and output format. Mental model: a briefing for an external copywriter on day one — more precise briefing, better result. The craft of writing reliable prompts is Prompt Engineering.
Prompt Library
DE: Prompt-BibliothekA team's shared collection of tested prompts, saved with context about when and how to use them. Prevents re-running the same iterations from scratch.
Prompt Chaining
DE: Prompt-VerkettungA pattern: split a complex task into a sequence of prompts where the output of one becomes the input of the next. Outperforms single mega-prompts on tasks with naturally separable phases (research → write, outline → draft, extract → summarize). Anthropic's first pattern for effective agents. See Chapter 8. Anthropic: Building Effective Agents
Prompt Drift
DE: Prompt-DriftGradual quality decay of a prompt across versions, often unnoticed as context and edge cases accumulate.
Prompt Engineering
DE: Prompt EngineeringThe practice of designing and refining inputs to language models to reliably produce high-quality outputs. Combines linguistics, domain knowledge, and empirical testing. Anthropic docs
Prompt Fingerprint
DE: Prompt-FingerabdruckHash over the fully resolved prompt including variables and system message, used to verify the identity of runs.
Prompt Injection
DE: Prompt InjectionSecurity attack on LLM systems where manipulative inputs override or bypass the original model behavior. Variable fields with clearly delimited input zones reduce the attack surface significantly. Relevant for any production-grade LLM deployment.
Prompt Locking
DE: Prompt-SperrePinning a prompt version together with model, temperature and system message so that later runs stay reproducible.
RACEO
DE: RAKEZA five-element prompt checklist: Role, Action, Context, Expected Format, Output Audience. A compressed form of the seven principles in Chapter 2 — useful as a quick mental check for everyday prompts.
ReAct
DE: ReActReasoning + Acting. A prompting pattern that interleaves model reasoning with action steps (web search, tool use, document inspection). The web research trigger from Chapter 2 is a simple ReAct: the model reasons about what to fetch, retrieves it, then synthesizes. Yao et al. 2023. See Chapter 8.
RLHF
DE: RLHFReinforcement Learning from Human Feedback. The training method where models are fine-tuned based on human ratings of their outputs. Produces helpful behavior, but causes Sycophancy — the model learns that agreeable answers receive higher ratings during training. See Mistake 8.
Rollback Prompt
DE: Rollback-PromptRestoring an older prompt revision in production without manual reconstruction.
RoPE (Rotary Position Embedding)
DE: RoPE (Rotary Position Embedding)The positional encoding mechanism used in most modern transformer architectures — Claude, GPT, Llama, Gemini. Causes a structural decay in attention to centrally-positioned tokens; this is the underlying mechanism of Lost-in-the-Middle (MIT 2025). Not training-related — cannot be "fixed" without architectural changes.
Self-Consistency
DE: Selbst-KonsistenzA prompting technique: run the same prompt 3–5 times and take the most consistent answer across runs. Reduces variance from sampling. Wang et al. 2022 reported gains of +17.9% on reasoning benchmarks; for content/marketing tasks the magnitude is smaller. See Chapter 8.
Server-Side Request Forgery (SSRF)
DE: Server-Side Request Forgery (SSRF)Server-Side Request Forgery is a vulnerability where a server is tricked into making arbitrary internal or external requests. Attackers can reach internal services that are normally shielded from the public internet. A Tenzai study (December 2025) found SSRF in 100% of the 15 vibe-coding apps tested.
Social Proof
DE: Sozialer BeweisEvidence that others have had positive experiences with a product or service: testimonials, case studies, named client logos, quantified results. Wikipedia
Silent Model Swap
DE: Stiller Modell-WechselThe provider redirects a model alias to a new version without explicitly informing the application.
Subagent
DE: SubagentIn Claude Code and the Claude Agent SDK: a delegated AI assistant that runs in its own context window with its own system prompt and tool permissions, and returns only a summary to the main agent. Used to keep large side-tasks (search dumps, log analysis, file scans) out of the main conversation. Anthropic docs
Sweet Spot
DE: Sweet SpotThe optimal prompt balance: enough context to orient the model, not so many rules that it stops reasoning and starts ticking boxes. Working baseline: RACEO complete plus ≤10 explicit constraints — every additional constraint should fix a known failure mode.
Sycophancy
DE: Gefälligkeits-BiasThe tendency of language models to agree with the user's premise or validate their work rather than offer honest assessment. A documented, structural side-effect of RLHF training (the model is rewarded for agreeable outputs during fine-tuning). Counteract by giving explicit permission to disagree: "Be direct. What doesn't work?" See Mistake 8.
System Prompt
DE: System-PromptThe persistent instruction at the start of a conversation. Defines voice, forbidden list, quality rules, and output format. The model follows it for the entire conversation without repetition in each request. Structurally distinct from a playbook (see System Prompt / Playbook), which additionally describes tools and escalation conditions.
System Prompt / Playbook
DE: System-Prompt / PlaybookStructured ruleset defining the behavior of an AI agent or LLM workflow. Describes goals, permitted tools, and escalation conditions. Foundation of any reliable agent design and basis for reproducible multi-step workflows.
Token
DE: TokenThe basic unit of text a model processes — roughly 0.75 words in English. Token count affects cost, speed, and context window usage. Anthropic docs · OpenAI docs
Token Determinism
DE: Token-DeterminismusProperty of a prompt that yields the same token sequence given identical inputs, model and seed.
Tree-of-Thought
DE: Gedankenbaum (Tree-of-Thought)A prompting technique for branching decisions: ask the model to enumerate multiple approaches before committing to one. Outperforms single-path Chain-of-Thought on tasks with multiple plausible paths — strategy, headline generation, objection prep. Yao et al. 2024. See Chapter 8.
Variable Slot
DE: Variablen-SlotNamed placeholder inside a prompt, filled at runtime with a concrete value, without altering the prompt itself.
Vibe Coding
DE: Vibe CodingVibe coding is AI-assisted software development in natural language: users describe intent in everyday language, and a language model like Cursor, Claude Code, or Lovable generates executable code. Andrej Karpathy coined the term on 2 February 2025 on X, and Collins Dictionary named it Word of the Year 2025.
Last updated: May 2026. © 2026 Lennart Austen. This glossary applies to the domain splicelog.com.