claude-code-aidlc-template

Designs a Microsoft Foundry (formerly Azure AI Foundry) footprint on Azure — surface map, model catalog (1,900+ via Azure OpenAI + partners), Agent Service (GA Mar 16 2026), private networking (BYO VNet), Foundry IQ + Azure AI Search, safety stack (Prompt Shields + Groundedness), cost model (PTU / PAYG / Batch), lock-in posture. Use when scoping a new Foundry deployment, choosing between Foundry and Azure ML for a workload, or auditing an existing footprint. Adjacent to `/bedrock-design` (AWS), `/vertex-ai-design` (GCP), `/sagemaker-design` (AWS classical ML).

Designs an Amazon Bedrock footprint on AWS — service split (Models / AgentCore / Knowledge Bases / Guardrails / Flows / Prompt Mgmt / IPR / Eval / Custom Import), model selection (Claude / Nova 2 / Llama / Mistral / Cohere / AI21), inference pattern (on-demand / PT / batch / cross-region), agent runtime (AgentCore vs Bedrock Agents vs Flows), guardrails posture (incl. Automated Reasoning), routing + cost (IPR), MLOps wiring, and lock-in posture. Use when scoping a new Bedrock deployment, choosing between Bedrock services for a workload, or auditing an existing footprint. Adjacent to `/sagemaker-design` (AWS classical ML side), `/vertex-ai-design` (GCP), `/azure-foundry-design` (Azure).

Designs an OpenAI platform footprint — API surface (Responses API as default; Assistants sunset 2026-08-26; Chat Completions feature-frozen on GPT-5.4+ reasoning models), model catalog (GPT-5.5/5.4/4.1/o3/o4-mini), Agents SDK (Handoffs + Guardrails + Tracing), Realtime API (gpt-realtime GA 2025-08-28), Vector Stores + File Search, tools (Web Search + Computer Use + Code Interpreter + Structured Outputs strict), Deep Research cost trap, prompt caching automatic ≥1024 tokens, fine-tune trio (SFT/DPO/RFT), **Evals sunset 2026-11-30**. Use when building on OpenAI directly (not Azure OpenAI). Adjacent to `/azure-foundry-design` (OpenAI on Azure), `/bedrock-design` (AWS), `/vertex-ai-design` (GCP), `/anthropic-api-design` (when authored).

Designs a Vertex AI footprint on Google Cloud — service split (Workbench / Pipelines / Training / Endpoints / Feature Store / Model Garden / Model Monitoring), compute selection, MLOps wiring, deployment pattern (online vs batch vs streaming), cost guardrails, and observability. Use when scoping a new Vertex deployment, choosing between Vertex services for a workload, or auditing an existing Vertex footprint. Adjacent to `/databricks-asset-bundles` + `/databricks-model-serving` (Databricks side) and `/sagemaker-design` (AWS side).

Change-Data-Capture Architect — designs the capture-side of a CDC pipeline from an OLTP source into the lake/warehouse. Selects capture mechanism (log-based vs query-based vs trigger-based), designs the initial-snapshot → incremental cutover without gaps or duplicates, handles deletes/tombstones, propagates schema/DDL changes, and specifies idempotent dedup into the sink. Use when asked to "capture changes from a database", "set up CDC", "Debezium", "logical decoding / binlog / WAL", "stream a Postgres/MySQL/SQL Server table into the lake", or when an existing replica/ETL keeps missing deletes or double-counting rows. Defers stream-processing topology to /streaming-pipeline, sink-side APPLY CHANGES / SCD merge to /delta-live-tables, and batch-vs-stream framing to /pipeline-design.

Designs distribution-free uncertainty quantification via conformal prediction — prediction SETS for classification and prediction INTERVALS for regression with a finite-sample coverage guarantee at 1−alpha. Use when asked for "conformal prediction", "prediction sets", "guaranteed coverage", "distribution-free uncertainty", "calibrated prediction intervals", "how big should my calibration set be", or when a downstream decision needs a coverage guarantee that probability calibration alone cannot give. Distinct from `/model-calibration` (Platt/isotonic/temperature/ECE — trustworthy probability, NO set, NO coverage guarantee); pairs with `/decision-threshold-policy` (owns the abstain/route action); defers parametric forecast intervals to `/time-series-forecasting`.

Doc-CI gate for repos whose deliverable includes a document set. Scans for (1) count drift across README/CLAUDE.md/NEXT_SESSION.md/prompts indices, (2) broken relative links, (3) skill↔prompt↔CLAUDE.md↔README↔prompts/README parity, (4) ungraded external facts, (5) markdown table pipe-integrity (stray-pipe phantom tables). Severity-grouped report. Use BEFORE shipping any docs commit, AFTER adding a skill/prompt/hook, and as a pre-commit gate. Pair with `.github/workflows/doc-ci.yml` for CI enforcement.

Designs a Data Loss Prevention (DLP) system — data classification, egress-surface enumeration, detection technique per class, enforcement per (class × surface), policy engine, and response/audit. Use when asked to prevent data exfiltration/leakage, design DLP controls, stop secrets/PII/PHI/PCI from leaving a system, or gate data egress (prompts, model output, tool actions, logs, network, files, third-party APIs).

Designs the training-data curation recipe for post-training an agentic / tool-use model (SFT trajectories + optional RL) — the task-sourcing & generation strategy, executable environments and the verification axis (hard test-based vs soft no-tests), teacher-rollout generation & teacher selection, trajectory filtering & source mixing, and the SFT→RL composition. Use when building a specialized open-weight agent from trajectory data, deciding how to generate agentic training tasks, or curating tool-use / coding / terminal trajectories for fine-tuning. Defers the fine-tune-vs-prompt decision to /fine-tune, RL mechanics + reward design to /rl-design, general modality synthesis to /synthetic-data-gen, eval framework to /eval-design, and the runtime agent loop to /agent-design.

Designs LLM evaluation frameworks with metric selection by task type, test set sizing, pass/fail thresholds, and drift triggers. Use when setting up evals for an LLM feature, before first production deployment, or when asked how to measure AI quality.

Determines whether to fine-tune vs. prompt-engineer, and if fine-tuning is justified, produces dataset requirements, training data format, pre/post eval plan, and cost-benefit analysis. Use when asked whether to fine-tune a model, how to prepare training data, or when a prompt-engineering approach has hit a quality ceiling.

Generate synthetic training data for ML tasks. Use when real data is scarce, PII prevents use of real data, rare class augmentation is needed, or stress-testing models on edge cases.

Designs RAG (retrieval-augmented generation) systems covering corpus architecture, chunking strategy, embedding, vector store selection, retrieval pattern, reranking, freshness, and observability. Use when building a RAG pipeline, choosing a retrieval approach, or deciding between long-context and RAG.

Decides whether and how to scale inference-time compute on a FIXED model — method selection (self-consistency / best-of-N + verifier / generative aggregation / verifier-guided search / sequential refinement), aggregator choice, sample-count (N) and round (T) sizing on the quality-vs-compute curve, and the scale-samples-vs-route-up-vs-train-bigger gate. Use when a single model call misses the quality floor and you're considering sampling multiple times, when choosing how to combine N candidate answers, or when weighing test-time compute against a bigger model or a fine-tune. Defers fan-out execution to /workflow-design, verifier/metric construction to /eval-design, token-cost math to /cost-optimize, cross-model routing to /llm-routing.

LLM Router — selects routing strategy, designs fallback chain, and defines quality-cost evaluation for directing queries to the right model at runtime

Designs a deterministic multi-agent workflow — parallel fan-out, pipeline stages, journaling, resume-from-cache, adversarial-verify panels, and Performance-Outcomes graders. Use when a task warrants spawning many subagents under deterministic orchestration (not model-driven control flow); when the same task will run many times and cost matters; when fan-out + barrier semantics matter; or when you need replayability after a failure. Distinct from `/agent-design` (single agent loop) and `/multi-agent-design` (architecture choice between patterns).

Audio / speech ML pipeline design — task framing (ASR/speech-to-text, TTS, speaker ID/verification, sound-event classification, keyword spotting, diarization/who-spoke-when, audio anomaly), representation selection (raw waveform / mel-spectrogram / log-mel / MFCC) with sample-rate + framing + VAD decisions, model family (Whisper / wav2vec2 / HuBERT / AST / pyannote, pretrained-then-finetune), augmentation (SpecAugment / noise / reverb / speed-pitch), task-specific metrics (WER/CER, DER, event-F1/mAP, EER, MOS), and the audio-specific leakage failure modes. Use when building any audio or speech ML pipeline, transcription, voice, or acoustic-monitoring system.

Designs a BigQuery ML (BQML) footprint on Google Cloud — the in-warehouse, SQL-driven ML surface. Covers model-type selection (linear / logistic / boosted-tree XGBoost / DNN / k-means / ARIMA_PLUS / matrix-factorization / AutoML / imported TF+ONNX), Gemini-in-BigQuery (ML.GENERATE_TEXT, ML.GENERATE_EMBEDDING, VECTOR_SEARCH + vector indexes, remote models over Vertex endpoints), the BQML-vs-export-to-Vertex decision, on-demand vs slot/editions compute + model-creation vs prediction billing, MLOps (TRANSFORM clause for train/serve consistency, scheduled-query retraining, model versioning), and serving (in-warehouse ML.PREDICT vs Vertex endpoint export for online). Use when data already lives in BigQuery and the team is SQL-first, when scoping a warehouse-native ML workload, or when deciding whether a model should stay in BQML or graduate to Vertex. Adjacent to `/vertex-ai-design` (the export target + sibling), `/snowflake-cortex-design` (warehouse-native peer on Snowflake), and `/rag-design` (vector-search retr

Designs per-turn context-window assembly for an LLM/agent call — budget allocation across all context sources (system prompt + tool defs + memory + retrieved chunks + history + scratchpad), position-of-evidence ordering (lost-in-the-middle mitigation + cache-aware stable prefix), overflow/compaction strategy (truncate vs summarize vs evict), relevance gating + cross-source dedup, and assembly-quality measurement. Use when a request is over token budget, when long sessions degrade (context rot), when key evidence is being ignored, or when the prompt cache keeps missing. Sibling to `/agent-memory` (which memories exist) and `/rag-design` (retrieval ranking); defers retrieval to `/rag-design`, memory-tier design to `/agent-memory`, and cost math to `/cost-optimize`.

Designs the de-identification treatment for sensitive fields — technique selection per field (suppression / masking / tokenization / pseudonymization / hashing / generalization / k-anonymity·l-diversity·t-closeness / differential privacy / format-preserving encryption), direct-vs-quasi-identifier classification, re-identification risk scoring, the utility-vs-privacy tradeoff against the downstream ML task, and a per-field treatment matrix. Use AFTER `/pii-scan` has identified sensitive fields and you need to decide HOW to treat each one before release for analytics or model training. The design-side complement to `/pii-scan`'s audit. Distinct from `/privacy-preserving-ml` (training/inference-time mechanism) and from `/synthetic-data-gen` (generate-fake alternative).

Designs run-time, estate-wide data observability over arbitrary warehouse/lakehouse tables nobody wrote explicit rules for. Owns (1) the five pillars (freshness, volume, schema, distribution/quality, lineage) as AUTO-BASELINED monitors vs the rule-authored model, (2) column-level LINEAGE as the root-cause primitive (bad metric → upstream table → column → job), (3) the incident lifecycle (detect → triage → ownership routing → SLA), (4) anomaly detection on metadata signals, (5) the build-vs-buy category axes. Use when monitoring data health across many tables, tracing a bad dashboard number to its upstream cause, standing up automated table monitors, or scoping a data-observability platform. Defers author-time per-pipeline validation/quarantine to /data-quality, ML-feature run-time freshness/null/PSI+routing to /feature-monitoring, Databricks UC product-config to /lakehouse-monitoring, and vendor TCO scoring to /build-vs-buy.

Databricks Mosaic AI Agent Framework Engineer — authors a GenAI agent against the MLflow ChatAgent/ResponsesAgent contract, instruments it with MLflow Tracing + autolog, gates promotion with the Agent Evaluation judge harness (mlflow.genai.evaluate), and binds Unity Catalog functions as governed tools. Use when building/logging/evaluating a GenAI agent ON Databricks, when asked about ChatAgent, ResponsesAgent, models-from-code logging, MLflow agent tracing, Agent Evaluation, LLM judges on Databricks, or UC-function tools. Defers the serving endpoint to /databricks-model-serving, retrieval/vector index to /mosaic-ai-vector-search, the agent loop/guardrails to /agent-design, and generic eval theory to /eval-design.

Predict-then-optimize advisor — turning an ML prediction (forecast/score/estimate) into a constrained operational decision (inventory order, price, route, staffing, allocation). Covers the prediction→optimizer handoff gate, formulation (objective + constraints; LP/MIP/CP/stochastic/robust/simulation), the core failure mode (optimizing against a point prediction amplifies model error), uncertainty-aware approaches (scenario stochastic programming, robust optimization, SPO+ / decision-focused learning, interval-aware decisions), pure-OR vs ML+OR coupling, decision-regret evaluation, and solver tooling. Use when a forecast or score must FEED a decision rather than being the deliverable. Single-shot constrained optimization — distinct from `/bandit-design` and `/rl-design` (sequential / online).

Turns calibrated model scores into operational decisions — operating-point selection from a confusion-cost matrix (expected-cost minimization, not default 0.5), threshold-vs-prevalence and prevalence-drift shift, ROC-vs-PR operating-point choice by imbalance (Youden's J / F-beta / cost curves), multi-tier auto-approve/review/auto-reject bands, and an abstention / reject-option that routes low-confidence cases to humans sized against review capacity. Emits a threshold-policy design doc. Use when asked "what threshold should I use", "is 0.5 right", how to set an operating point, how to band scores into auto/review/reject, when to send a case to human review, or why a tuned threshold broke after prevalence drifted. Defers calibrated-probability production to `/model-calibration`, deployment wiring to `/model-deployment`, drift-detection math to `/model-drift`.

Geospatial / spatial ML design — spatial data types (point / polygon / raster / trajectory / network), CRS + projection choice, spatial feature engineering (distance, spatial joins, neighborhood aggregation, H3 / geohash binning, kernel density), task framing (interpolation/kriging, geospatial regression, hotspot detection, satellite CV, trajectory mining), spatial cross-validation, and the signature spatial-autocorrelation leakage failure mode. Use when data has coordinates / geometry, when building a model over locations, when asked about kriging / GWR / spatial regression / hotspot detection / satellite imagery ML, or when random CV gives suspiciously high scores on spatial data.

Graph / network ML design — graph construction from relational/event data (node/edge definition, homogeneous vs heterogeneous, static vs temporal), task framing (node / link / graph level), GNN family selection (GCN / GraphSAGE / GAT / GIN / R-GCN / TGN / TGAT), scale strategy (full-batch vs neighbor vs subgraph sampling), and leakage-safe splits for connected data. Use when modeling relationships/networks, asked about GNNs, graph neural networks, link prediction, node classification, fraud rings, or when relational/event data has a natural graph structure. Owns graph neighbor-leakage; defer general leakage to /leakage-audit and /split-design.

Designs the operational human-oversight loop for a NON-AGENTIC ML decision system — the in-decision-path review mechanism that sits BEFORE a prediction takes effect. Owns escalation routing + queue topology, reviewer-queue sizing / SLA / prioritization (Little's Law), the override / appeal / contest flow with its audit trail, automation-bias mitigation (anti-rubber-stamp UI + process), and oversight-effectiveness metrics keyed to risk tier. Emits a human-oversight design doc keyed to the system's risk tier. Use when asked "how do I route low-confidence predictions to a human", how big should the review team be, how does a person override or appeal an automated decision, how do I stop reviewers rubber-stamping the model, or how do I prove the EU-AI-Act "human oversight" box is real. Defers risk-tier definitions + governance framework to `/responsible-ai-governance`, the threshold/band that SELECTS cases to `/decision-threshold-policy`, post-hoc QA sampling to `/feedback-loop`, IAA metric mechanics to `/label-q

Selects the correct statistical hypothesis test for an OBSERVATIONAL EDA question, checks its assumptions before running it, chooses effect size + power, applies multiple-comparison correction, and guards against observational pitfalls. Use when comparing groups or testing an association in existing data and asking "which test", "is this difference significant", "t-test or Mann-Whitney", "chi-square or Fisher", "do I need Bonferroni", "did I p-hack this". Produces a test-plan + interpretation card. Defers interventional experiments to /experiment-design, live-traffic splits to /ab-test-design, trained-model-vs-model to /model-comparison, association strength without p-governance to /feature-correlation.

Designs a missing-value imputation strategy — missingness-mechanism diagnosis (MCAR/MAR/MNAR), method selection (deletion / mean-median-mode / KNN / MICE / model-based / domain / time-aware), the missingness-indicator decision, fit-on-train-only mechanics, and a post-imputation distortion check. Use when a dataset has nulls and you must decide how to fill them before training, when mean-imputation is distorting variance or correlations, or when missingness itself may be informative. Defers heavy temporal gap-fill to /timeseries-resample and leakage enforcement to /leakage-audit.

Configures a Databricks data-profiling monitor (formerly Lakehouse Monitoring; the data-profiling half of Unity Catalog Data Quality Monitoring) over a Delta or inference table — selects profile type (Snapshot/TimeSeries/InferenceLog) from table shape, chooses a baseline strategy, sets slice expressions + custom metrics + granularities, configures refresh cadence + cost, and wires the auto-generated profile + drift metric tables to the auto-dashboard + SQL alerts, including the inference-table→drift-alert loop. Use when standing up Lakehouse Monitoring / data profiling on a UC table, wiring a served model's inference table to drift alerts, or choosing a monitor profile type + baseline. Owns the PRODUCT config; defers drift statistics (PSI/KS/threshold math) to /model-drift + /feature-monitoring, serving to /databricks-model-serving, and Databricks' built-in anomaly detection (the other half of Data Quality Monitoring) is out of scope.

Designs a learning-to-rank (LTR) system for search / ads / document retrieval — selects the objective class (pointwise vs pairwise vs listwise), picks the model family (LambdaMART / GBDT rankers vs neural / cross-encoder / learned-sparse rerankers), constructs relevance judgments (editorial vs click-derived labels with position-bias correction via click models + IPW / counterfactual LTR), chooses ranking metrics (NDCG@k / MAP / MRR / ERR), and sets up offline-vs-online evaluation with query-level splits. Use when ranking documents/results for a QUERY (web/e-commerce/ads/enterprise search), choosing a ranking loss, building relevance judgments from click logs, or when offline NDCG gains don't move the online metric. Defers user-item personalization / collaborative filtering / two-tower retrieval to /recommender-design and A/B test mechanics to /ab-test-design.

Check and fix probability calibration for ML classifiers. Use when predicted probabilities are used for ranking, thresholding, or downstream decisions — not just hard class labels.

Multimodal / vision-language system design advisor — decides multimodal vs two single-modality models, selects fusion strategy (early / late / cross-attention; joint-embedding vs generative VLM), picks models (CLIP / SigLIP for retrieval, frontier + open VLMs for generation/grounding, LayoutLM-family / Donut for documents), designs multimodal RAG (where to embed, cross-modal rerank), defines cross-modal eval, and names the cross-modal failure modes (modality collapse, missing-modality, alignment drift, OCR bottleneck). Use when a system must reason over image+text (or audio/video+text) jointly, when building a vision-language retriever or VQA/grounding system, or when deciding whether fusion is justified at all. The fusion/cross-modal layer — defer image-only to /computer-vision, text-only to /nlp-pipeline.

Privacy-preserving ML mechanism advisor — the HOW of privacy at training and inference time. Selects per-data-flow mechanism (differential privacy / DP-SGD, federated learning + secure aggregation, k-anonymity / l-diversity / t-closeness anonymization), sets an epsilon budget, maps inference-time threats (membership inference, model inversion, attribute inference, model extraction) to defenses, and decides when HE / SMPC / TEE is justified. Emits a privacy design doc with chosen mechanism per flow, epsilon budget, residual-risk register, and privacy-utility tradeoff. Use when asked "how do I make this model private", "differential privacy", "DP-SGD", "epsilon budget", "federated learning", "membership inference", "model inversion", "train without seeing raw data", or before training/serving a model on sensitive personal data. The NIST AI RMF "Privacy-Enhanced" trait skill. Distinct from `/pii-scan` (detection — feeds this) and `/data-deidentification-design` (dataset de-id treatment).

Designs a PromptOps / prompt-lifecycle pipeline — prompt versioning + registry, environment promotion (dev→staging→prod, build-once-promote-many), eval-gated rollout, A/B-in-prod + canary for prompt variants, regression rollback, and change-control governance. Emits a PromptOps design doc (registry schema + promotion gates + rollback trigger spec + change policy). Use when asked about "prompt versioning", "prompt registry", "PromptOps", "promote a prompt to prod", "prompt rollback", "who approves a prompt change", "A/B test a prompt", or shipping prompt changes safely. DEFER eval metric/test-set design to /eval-design, one-time prompt quality scoring to /prompt-review, model-object lifecycle to /experiment-tracking + /model-deployment.

Designs a Snowflake AI/ML footprint — service split (Cortex AISQL LLM functions / Cortex Analyst text-to-SQL / Cortex Search managed RAG / Cortex Agents / Snowpark ML + Model Registry + Feature Store / SPCS serving), compute model (warehouse vs Cortex serverless vs SPCS compute pool), governance (RBAC / row-access / masking / tags), credit-based cost guardrails, MLOps wiring, deployment pattern (in-warehouse batch vs SPCS real-time), and lock-in posture. Use when scoping a new Snowflake Cortex deployment, choosing between Snowflake AI surfaces for a workload, or auditing an existing footprint. Adjacent to `/bigquery-ml-design` (the warehouse-native ML peer on GCP), `/sagemaker-design` (AWS), `/vertex-ai-design` (GCP), `/bedrock-design` + `/azure-foundry-design` (managed-GenAI siblings), and `/databricks-asset-bundles` (the lakehouse alternative).

Profiles a raw TEXT or IMAGE corpus before modeling — document/token length distributions, language + encoding mix (mojibake/mixed-script), exact + near-duplicate rate, label coverage + imbalance, vocabulary/OOV, PII flag (text); resolution/aspect/channel/bit-depth distribution, corrupt-file audit, class balance, cross-split duplicate leakage, EXIF/source-batch effects, brightness/blur outliers (image). Emits a corpus-EDA report with per-dimension GO / FIX-FIRST verdicts. Use before NLP or CV modeling, when asked to "profile this text corpus", "audit this image dataset", "EDA on documents/images", or before /nlp-pipeline or /computer-vision. Owns pre-modeling unstructured-corpus profiling; defers tabular columns to /eda, temporal structure to /time-series-eda, dedup mechanics to /dedup (reports rate only), and PII detail to /pii-scan (flags presence only).

Designs agentic AI systems covering loop architecture, tool manifests, guardrails, human-in-the-loop gates, memory design, fallback paths, and observability. Use when building an agent, designing a multi-step AI workflow, or when a system involves tool use with potential real-world side effects.

Designs the memory layer for a long-running agent — tier selection (core/working/session/episodic/semantic/procedural), backing-store choice (file/KV/vector/graph/temporal-graph), validity-window discipline, stale-context detection, and memory-quality eval. Use when an agent must remember across turns / sessions / users; when raw RAG isn't enough; or when an agent loop already exists and you're deciding what to persist where. Sibling to `/agent-design` (loop + tools) and `/rag-design` (retrieval-time semantics).

Guides a structured API portfolio audit — from discovery through primary-source-verified inventory, shortcomings analysis, prioritized recommendations, executive summary, and options analysis. Enforces primary-source discipline, cross-file consistency, and advisor review gates. Use when auditing an API portfolio, evaluating a vendor's developer ecosystem, or producing defensible technical analysis of an API set.