CtrlK
BlogDocsLog inGet started
Tessl Logo

ai-architecture

AI/LLM system architecture: software arch, pipelines, conops, design patterns, audit, implementation. Topics: ai-conops, ai-design-patterns, ai-pipeline-architecture, ai-software-architecture, audit, chatbot-design, embedding-strategy, fine-tuning-prep, implementation, prompt-engineering, rag-patterns, structured-output, voice-interface.

SKILL.md
Quality
Evals
Security

ai-architecture

Router skill: resolve topic from the request, then Read EXACTLY ONE playbook from routes:. The playbook does the work; this file only routes. [DOC]

When to use

User asks to design, audit, or implement an AI/LLM system component — software architecture, pipelines, ConOps, design patterns, RAG, prompts, embeddings, fine-tuning prep, structured output, chatbot/voice interface. If the request is not LLM/AI-system shaped, do not route here. [INFERENCIA]

Inputs → Outputs

  • In: topic (one enum value, required), depth (quick|deep, default quick).
  • Resolve topic: infer from the request; ask ONLY if two enum values fit equally. Map intent, not keywords (e.g. "make my prompt reliable" → prompt-engineering; "ground answers in our docs" → rag-patterns). [INFERENCIA]
  • Out: the artifact the chosen playbook defines, with evidence tags and any validation gate that playbook specifies — never an answer improvised here. [DOC]

Routing procedure

  1. Pick the single best topic; if none fits, say so and stop — do not force-fit. [SUPUESTO]
  2. Read its routes: playbook. Read NOTHING else from the cluster. [DOC]
  3. quick → essentials only; deep → apply the playbook exhaustively, verifying at each step. Honor the playbook's own gates over these defaults. [INFERENCIA]

Spine: Discover → Analyze → Execute → Validate. Quality gates: constitution v6.0.0 (enforcement), evidence tags, script-first rule. [CONFIG] Deterministic routing aids: assets/ (routing-rubric.json, routing-checklist.md). [CONFIG]

Validation gate (before handing off)

  • Exactly one topic resolved and present in the enum. [DOC]
  • Exactly one playbook Read; cluster not bulk-loaded. [DOC]
  • Output carries evidence tags from ONE family (Alfa core here: [DOC] [CONFIG] [CÓDIGO] [INFERENCIA] [SUPUESTO]). [CONFIG]

Anti-patterns

  • Loading several playbooks "to compare" — defeats the router; pick one. [INFERENCIA]
  • Answering AI-architecture questions inline without routing. [INFERENCIA]
  • Asking for topic when the request already makes it obvious. [SUPUESTO]
  • Mixing tag families, or emitting [CÓDIGO]/[CONFIG] you cannot point to in-repo — downgrade to [SUPUESTO]. [CONFIG]

Self-correction

If the playbook doesn't fit the ask, stop, re-resolve topic, Read the correct one — never patch with a second. If no enum fits, report the gap. [SUPUESTO]

Repository
JaviMontano/jm-adk-beta
Last updated
First committed

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.