[{"data":1,"prerenderedAt":294},["ShallowReactive",2],{"blog-post-blog_en-context-engineering-fuer-ki-systeme":3},{"id":4,"title":5,"body":6,"cover":278,"date":279,"description":280,"draft":281,"extension":282,"meta":283,"navigation":284,"path":285,"seo":286,"stem":287,"tags":288,"__hash__":293},"blog_en\u002Fen\u002Fblog\u002Fcontext-engineering-fuer-ki-systeme.md","Context Engineering for AI Systems: Why Prompts Are Not Enough",{"type":7,"value":8,"toc":273},"minimark",[9,13,18,21,24,59,66,70,73,76,215,247,250,254,257,260,269],[10,11,12],"p",{},"Context engineering becomes relevant once AI systems do more than write answers. They prepare decisions, change code, or trigger actions in backend systems. The term describes which information, rules, and tools a model sees before it acts. For growing software teams, this is not prompt optimisation. It is an architecture and quality concern.",[14,15,17],"h2",{"id":16},"what-context-engineering-means-in-ai-systems","What Context Engineering Means in AI Systems",[10,19,20],{},"A prompt is only the visible task. Production AI systems operate with much more context: system instructions, user roles, knowledge sources, tool descriptions, history, permissions, and output formats.",[10,22,23],{},"Context engineering deliberately shapes this environment:",[25,26,27,35,41,47,53],"ul",{},[28,29,30,34],"li",{},[31,32,33],"strong",{},"Instructions:"," Which business rules, tone, security boundaries, and quality criteria apply?",[28,36,37,40],{},[31,38,39],{},"Knowledge:"," Which documents, tickets, code areas, or customer data may be retrieved through RAG or search?",[28,42,43,46],{},[31,44,45],{},"Tools:"," Which APIs, MCP servers, or internal services may the system read from or execute?",[28,48,49,52],{},[31,50,51],{},"State:"," Which history, user role, or process stage is actually relevant to the current task?",[28,54,55,58],{},[31,56,57],{},"Traceability:"," Which sources, versions, tool calls, and assumptions must be auditable later?",[10,60,61,62,65],{},"The important point is not to push as much context as possible into a model. Too much context increases cost and latency, weakens tool selection, and can pull contradictory or sensitive information into decisions. Good context engineering therefore also decides what ",[31,63,64],{},"must not"," enter model context.",[14,67,69],{"id":68},"where-teams-should-start-with-context-engineering","Where Teams Should Start With Context Engineering",[10,71,72],{},"The most common mistake is treating context engineering as a collection of additional prompt files. Context then grows without control, ownership, versioning, or measurable quality boundaries.",[10,74,75],{},"A useful starting point is one product-adjacent workflow, such as support summaries, internal knowledge search, or AI-assisted pull request preparation. For that workflow, the team should define a small context policy:",[77,78,83],"pre",{"className":79,"code":80,"language":81,"meta":82,"style":82},"language-yaml shiki shiki-themes github-light github-dark","ai_context_policy:\n  workflow: support_ticket_summary\n  allowed_sources: [\"ticket_text\", \"customer_plan\", \"public_docs\"]\n  forbidden_sources: [\"payment_data\", \"internal_margin_notes\"]\n  tool_access: [\"read_customer_status\"]\n  required_metadata: [\"source_id\", \"source_version\", \"tenant_id\"]\n  evaluation: [\"accuracy\", \"missing_context\", \"data_leakage\"]\n","yaml","",[84,85,86,99,112,138,156,169,192],"code",{"__ignoreMap":82},[87,88,91,95],"span",{"class":89,"line":90},"line",1,[87,92,94],{"class":93},"s9eBZ","ai_context_policy",[87,96,98],{"class":97},"sVt8B",":\n",[87,100,102,105,108],{"class":89,"line":101},2,[87,103,104],{"class":93},"  workflow",[87,106,107],{"class":97},": ",[87,109,111],{"class":110},"sZZnC","support_ticket_summary\n",[87,113,115,118,121,124,127,130,132,135],{"class":89,"line":114},3,[87,116,117],{"class":93},"  allowed_sources",[87,119,120],{"class":97},": [",[87,122,123],{"class":110},"\"ticket_text\"",[87,125,126],{"class":97},", ",[87,128,129],{"class":110},"\"customer_plan\"",[87,131,126],{"class":97},[87,133,134],{"class":110},"\"public_docs\"",[87,136,137],{"class":97},"]\n",[87,139,141,144,146,149,151,154],{"class":89,"line":140},4,[87,142,143],{"class":93},"  forbidden_sources",[87,145,120],{"class":97},[87,147,148],{"class":110},"\"payment_data\"",[87,150,126],{"class":97},[87,152,153],{"class":110},"\"internal_margin_notes\"",[87,155,137],{"class":97},[87,157,159,162,164,167],{"class":89,"line":158},5,[87,160,161],{"class":93},"  tool_access",[87,163,120],{"class":97},[87,165,166],{"class":110},"\"read_customer_status\"",[87,168,137],{"class":97},[87,170,172,175,177,180,182,185,187,190],{"class":89,"line":171},6,[87,173,174],{"class":93},"  required_metadata",[87,176,120],{"class":97},[87,178,179],{"class":110},"\"source_id\"",[87,181,126],{"class":97},[87,183,184],{"class":110},"\"source_version\"",[87,186,126],{"class":97},[87,188,189],{"class":110},"\"tenant_id\"",[87,191,137],{"class":97},[87,193,195,198,200,203,205,208,210,213],{"class":89,"line":194},7,[87,196,197],{"class":93},"  evaluation",[87,199,120],{"class":97},[87,201,202],{"class":110},"\"accuracy\"",[87,204,126],{"class":97},[87,206,207],{"class":110},"\"missing_context\"",[87,209,126],{"class":97},[87,211,212],{"class":110},"\"data_leakage\"",[87,214,137],{"class":97},[25,216,217,223,229,235,241],{},[28,218,219,222],{},[31,220,221],{},"Context need:"," Which information does the model need to solve the task reliably?",[28,224,225,228],{},[31,226,227],{},"Context quality:"," Who owns the sources, how current are they, and how are errors corrected?",[28,230,231,234],{},[31,232,233],{},"Context budget:"," Which data is summarised, selected, or intentionally left out?",[28,236,237,240],{},[31,238,239],{},"Context boundaries:"," Which data must never land in prompts, logs, or model provider APIs?",[28,242,243,246],{},[31,244,245],{},"Context tests:"," Which example tasks show whether a change to retrieval, tools, or rules makes behaviour worse?",[10,248,249],{},"This turns context engineering into a normal part of backend architecture: data flows, permissions, interfaces, tests, and observability are considered together.",[14,251,253],{"id":252},"why-this-matters","Why This Matters",[10,255,256],{},"AI systems rarely fail because of one bad prompt alone. In production workflows, expensive failures happen when wrong or missing information leads to wrong actions: unsuitable customer responses, risky code changes, breached privacy boundaries, or decisions nobody can explain.",[10,258,259],{},"Context engineering reduces these risks because teams shape the model's decision space technically. This improves quality and maintainability: new models can be tested without every feature hiding its own context logic.",[10,261,262,263,268],{},"For decision-makers, the economic point is clear. Without controlled context, token cost, support effort, and compliance risk grow faster than the value of AI features. An ",[264,265,267],"a",{"href":266},"\u002Fen\u002F#packages","Architecture & AI Review"," can assess whether context, tool access, and quality measurement fit together before an AI workflow scales.",[270,271,272],"style",{},"html pre.shiki code .s9eBZ, html code.shiki .s9eBZ{--shiki-default:#22863A;--shiki-dark:#85E89D}html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":82,"searchDepth":101,"depth":101,"links":274},[275,276,277],{"id":16,"depth":101,"text":17},{"id":68,"depth":101,"text":69},{"id":252,"depth":101,"text":253},null,"2026-06-10","Context engineering makes AI systems more reliable by controlling context, data sources, tools and ownership before production use.",false,"md",{},true,"\u002Fen\u002Fblog\u002Fcontext-engineering-fuer-ki-systeme",{"title":5,"description":280},"en\u002Fblog\u002Fcontext-engineering-fuer-ki-systeme",[289,290,291,292],"AI","Software Architecture","Backend Development","Software Quality","LdAAE9RFdAJIG3sLUYIixLNX1mkJdRwucXzlXhUclfw",1781596426438]