ai-architecture-advisor
Chooses among AI/ML approaches: classical ML, LLM, RAG, fine-tuning, agents, multimodal, embeddings/recsys, dense/MoE/SSM/diffusion. Use when picking or scaling an architecture.
vasilyu1983 · AI review on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add vasilyu1983/AI-Agents-public --skill ai-architecture-advisorAbout this skill
name: ai-architecture-advisor description: "Chooses among AI/ML approaches: classical ML, LLM, RAG, fine-tuning, agents, multimodal, embeddings/recsys, dense/MoE/SSM/diffusion. Use when picking or scaling an architecture." compatibility: Portable core. Works on Claude Code and Codex. version: "1.7" last_validated: 2026-07-11
AI Architecture Advisor
The front-door decision skill for "I have problem X — what should I build with?" It owns the choice — which approach fits, when to promote complexity (and when not to), how to transfer knowledge, how to scale — then hands off to the deep skill that owns the depth. It spans the whole modeling space in one place so you can compare options that normally live in separate skills side by side:
tabular GBDT · deep net · Transformer/LLM · RAG · fine-tuning · agents
multimodal/omni · embeddings & retrieval · recsys/ranking · model architecture (dense/MoE/SSM/diffusion)
No theory dumps — decision tables, elimination logic, tradeoffs, and a pointer to the deep skill.
The architect's move is to ask before answering. The amateur hears "build an AI feature" and reaches for the model they know ("we'll fine-tune Kimi"). The architect first asks: what data type? what volume? what task? what's the success metric? do you even need a Transformer? The skill that distinguishes an architect is the willingness to say "for this, CatBoost wins," "here you need a Transformer," or "LoRA is enough here" — and to refuse to name an approach until the problem is classified. Never jump to a model before the Intake questions below are answered.
ASCII Flow
problem + data + constraints
|
v
0. INTAKE — ask before answering (see questions below)
| do NOT name a model until task + data + metric + constraints are known
v
1. classify the problem (tabular? text? generation? decision/action? retrieval?)
|
v
2. eliminate ineligible options (with a reason each — never silently drop)
|
v
3. score survivors independently (accuracy, latency, cost, data need, interpretability, ops)
|
v
4. pick the SIMPLEST that clears the bar (start simple, promote only on evidence)
|
v
5. hand off to the deep skill (ai-ml-data-science / ai-llm / ai-rag / ai-agents ...)
Intake: Ask Before You Answer
A request like "let's fine-tune model X" is a proposed solution, not a problem statement. Do not accept it at face value. Surface the six questions that decide the architecture, and hold any model name until they're answered. If the user can't answer one, that gap is itself a finding (most often: no success metric, or no labeled data).
| Ask | Why it changes the answer |
|---|---|
| What's the task? classify / rank / generate / extract / retrieve / act | Picks the lane before anything else — tabular-classify and open-ended-generate share no architecture |
| What's the data? type (tabular/text/image/audio), volume, labeled? private? fresh? | Tabular+small → trees; private/fresh knowledge → RAG not fine-tune; no labels → no SFT |
| What's the success metric & bar? accuracy / latency / cost-per-outcome / interpretability | "Make it work" can't be scored; the binding constraint eliminates most options |
| What are the constraints? p99 latency, $/query at volume, on-prem/cloud, no-hallucination | A <50ms or AWS-only or no-hallucination constraint forces the design more than the task does |
| Is there a regulatory/compliance driver? high-risk classification (EU AI Act), data residency, right-to-explanation, audit trail | Can force an inherently-interpretable model (trees, not a black-box net), ban sending data to a third-party API, or mandate human-in-the-loop — overriding an accuracy-only ranking |
| Do you even need an LLM/Transformer? | The cheapest thing that clears the bar wins — often a tree, a regex, or a single API call beats a fine-tune |
Breaking ties among survivors: hard constraints (a p99 ceiling, an on-prem requirement, a regulatory bar) are gates — fail one and the option is out, no matter how it scores elsewhere. Only among options that clear every gate do the soft criteria (cost, developer familiarity, marginal accuracy) act as tie-breakers. Don't let a soft criterion overrule a failed hard constraint, and don't eliminate an option on a soft criterion before checking whether a rival actually clears every gate.
Quick Reference
| You're choosing between | Default first move | Promote when | Deep skill |
|---|---|---|---|
| Tabular model family | Gradient-boosted trees (XGBoost) | Within a tabular foundation model's envelope (TabPFN-3 ≤~1M rows; TabICL v2 if open-source) it now beats tuned GBDTs there | ai-ml-data-science |
| Trees vs deep learning (tabular) | Trees | Huge data + raw signals (images/text/audio) embedded in features | ai-ml-data-science |
| Thinking effort on one model | Low/no thinking | Gap is multi-step logic/planning: raise the thinking budget (it's a dial, not a separate model) before fine-tuning | ai-llm |
| LLM prompt vs RAG | Prompt-only | Needs current/private knowledge or citations (and the corpus won't fit affordably in a long-context window) | ai-rag |
| RAG vs fine-tuning | RAG | Behavior/format gap persists after RAG+prompt fixes; combine both (RAFT) for domain-specific RAG | ai-llm |
| Adapt vs post-train (RLHF/DPO/GRPO) | Prompt → RAG → SFT first | Behavior/preference/reasoning gap survives SFT and you own preference or verifiable-reward data | ai-post-training |
| Single LLM call vs agent | Single call / workflow | Needs tools (via MCP), multi-step planning, or external actions | ai-agents |
| Omni model vs specialist pipeline | Omni/VLM to validate | Accuracy KPI, cost-at-volume, or no-hallucination ne |