--- title: "AI Coding Assistant Domain Knowledge: Grounding in Schemas - AstroDev" description: "AI coding assistant domain knowledge fails when it is memorized. Google's Ads API Developer Assistant v4.0.0 shows why schema grounding beats guessing." url: "https://astrodev.carlosarias.com/blog/guides/grounding-ai-coding-assistants-domain-knowledge" --- [Guides](/categories/guides) # AI Coding Assistant Domain Knowledge: Grounding in Schemas AI coding assistant domain knowledge fails when it is memorized. Google's Ads API Developer Assistant v4.0.0 shows why schema grounding beats guessing. [Carlos Arias](/authors/carlos-arias) · September 25, 2026 · 5 min read ![An AI coding assistant reading a live API schema instead of guessing.](/_astro/cover.rR3xfXZq_Z2nOVUU.webp) *An AI coding assistant reading a live API schema instead of guessing. AI-generated illustration by Carlos Arias .* On-brand editorial cover for an article titled "AI Coding Assistant Domain Knowledge: Grounding in Schemas". Sophisticated, minimal conceptual illustration on a very dark ink background (#111318) with a single restrained warm accent glow. High-end business-publication aesthetic, subtle depth, cinematic soft light. No text, no words, no letters, no logos, no UI labels. Prompt sent to Higgsfield · nano_banana_pro · 3:2 An AI coding assistant is only as reliable as the domain knowledge it can reach. Store that knowledge inside a base model’s weights and watch it rot. It drifts. It invents field names that never shipped. Google’s answer, released in the Ads API Developer Assistant v4.0.0 on August 25, 2026, grounds the assistant in the API’s real definitions instead of its memory (Google Ads Developer Blog, August 2026). It is a clean case study in ai coding assistant domain knowledge treated as engineering, not prompting. This guide walks through what Google changed and the grounding pattern worth copying. ## Why AI coding assistant domain knowledge should live in schemas, not memory A base model memorizes a snapshot. The Google Ads API does not hold still. It ships several named versions a year and enforces rules a model cannot infer: which fields are compatible in a single query, or how date segmentation constrains a SELECT. Ask a model to recall all of that. It will return GAQL that looks right and fails at runtime. The usual workaround is iterative guessing. The assistant writes a query and the API rejects it. It reads the error and tries again. Each loop burns tokens and a network round trip, and it trains no one, because the next query starts from the same stale memory. Grounding removes the loop. Point the assistant at the actual schema and the guess becomes a lookup. This is the same lever we described in how context shapes AI agent output: the context you feed the model, not the cleverness of the prompt, is what decides whether the output is usable. Google is explicit that this is the goal. The v4.0.0 assistant grounds more of its work in the API’s real definitions “rather than relying solely on an AI model’s existing knowledge” (Search Engine Land, August 2026). ## How schema grounding compares to RAG, MCP, context injection and fine-tuning Schema grounding is one option among several. Anyone searching how to give an agent domain knowledge has met the rest, so it helps to place it against them. Retrieval-augmented generation pulls matching documents into the prompt at query time. It fits prose: documentation, support tickets, code comments, the odd design doc. A schema is not prose. It is a typed contract, and similarity search can surface a field from the wrong API version as readily as the right one. Fine-tuning bakes the knowledge into the weights. That is the failure Google is routing around. Retrain on every version and the model holds another snapshot that rots. Context injection stuffs the whole schema into the system prompt. That holds until the schema outgrows the window, which the Google Ads schema does across many live versions. The Model Context Protocol is the near cousin. It standardizes how an agent reaches a live tool or data source, and a Protobuf inspector is exactly that. Grounding is the principle. MCP is one wire it can travel. v4.0.0 reads the definitions on local disk, the same bet by a shorter route. ## The four grounding moves in v4.0.0 The rebuild rewards dissection. Each choice generalizes. Read it as an engineer rather than a user, and you see four distinct bets on where knowledge should sit. - A globally installed plugin, not a workspace scaffold. Earlier versions dropped a standalone project structure into one repo. v4.0.0 installs once as a plugin, so its Google Ads rules and diagnostic commands follow the developer from repo to repo and IDE to IDE (Search Engine Land, August 2026). Knowledge stops being per-project copy. - Architectural rules embedded as files. The assistant runs on the Antigravity and Claude Code agent frameworks, and it leverages AGENTS.md and CLAUDE.md for persistent context and custom skills (Google Ads Developer Blog, August 2026). The rules live where the agent already looks. - Local Protobuf schema inspection. The assistant inspects Protobuf schemas on demand to discover resource fields and enum values for any active API version, without remote metadata overhead (Google Ads Developer Blog, August 2026). The source of truth is the definition, read locally. - Single-step GAQL validation. A /validate-gaql command dry-runs a query through the API’s validate_only parameter, and a local validator checks fields, resources, zero-impression metrics, and date clauses in one pass rather than a retry loop (Search Engine Land, August 2026). ## Step by step: how the assistant validates a query without guessing The GAQL validation flow is the sharpest example, so it is worth tracing in order. It swaps a probabilistic loop for a deterministic check. 1. The assistant reads the resource you are querying straight from the local Protobuf schema. That tells it the exact fields, nested structures, data types, and enum values that exist in the API version you target, so it never proposes a column that was renamed two versions ago (Search Engine Land, August 2026). 2. The local validator checks the query’s shape in a single step (Search Engine Land, August 2026). Are the selected fields mutually compatible? Does the date segmentation obey the rules the metrics impose? 3. /validate-gaql sends the query to the API with validate_only set, so the server confirms it would run without executing it or spending on live data (Search Engine Land, August 2026). Three checks, one pass. The developer sees a verdict before a real query ever runs. Compare that to the old shape, where the model guessed, ran, failed, and guessed again, and you can see the cost that grounding quietly deletes. The retry loop wastes time. Worse, it lets confidently wrong code slip into a codebase, because a query that happens to parse can still return the wrong numbers. ## Claude Code plugin architecture and why “global” is the real change The claude code plugin architecture decision looks like packaging. It is actually the load-bearing one. A plugin scoped to a single workspace has to be reinstalled and re-taught for every project, which means the domain knowledge decays the moment a developer opens a new repo. A global plugin carries its rules and validators across every project and editor at once. That portability came at a real cost, and Google did not hide it. v4.0.0 breaks compatibility with every earlier installation and forces existing users to reinstall from scratch (ppc.land, August 2026). A team that treated the grounding as a durable investment accepted a one-time migration to get it. That trade is the tell. You do not break every install unless the thing you are moving, the domain knowledge, is worth carrying everywhere. ## The pattern any agent platform should copy Strip the Google Ads specifics and a general rule remains. Put the volatile domain facts in a source the agent reads, and keep them out of the source the agent memorizes. Grounding ai agents in schemas is the sharpest version of that rule, because a schema is versioned and already the source of truth your systems run on. The pattern has three parts, and they reinforce each other. First, read the real definition instead of recalling one. Validate once, deterministically, rather than looping through failures. The last part is portability, so the capability follows the work instead of being rebuilt for every repo. Each part is cheap alone. Together they turn an assistant that merely sounds authoritative into one that is checkable, and checkable is the only property that matters when the output is code you will ship. This is the same argument we make about the layer beneath any agent. In AI agent readiness we argued that most agent failures are data-foundation failures, not model failures: an agent reproduces the mess it is given, faster and with more confidence. Google’s release is that argument shipped as a product. The schema is the foundation. The validator is the quality gate that sits on top of it. The plugin is how both travel. When you evaluate an agent platform, this is the question worth asking. Does it ground its answers in your real definitions, or does it ask you to trust a model’s memory of them? ## What to take from it Domain grounding is not a feature you bolt on at the end. Ask where the domain knowledge lives. Then ask whether the agent can verify a claim before it hands you code. Our other engineering write-ups trace how we apply the same discipline. Tags [#ai-coding-assistant](/tags/ai-coding-assistant)[#schema-grounding](/tags/schema-grounding)[#gaql-validation](/tags/gaql-validation)[#claude-code-plugin](/tags/claude-code-plugin)[#ai-agents](/tags/ai-agents) Share Written by [Carlos Arias](/authors/carlos-arias) Builder of AstroAgent, an AI-run website platform. On this page - Why AI coding assistant domain knowledge should live in schemas, not memory - How schema grounding compares to RAG, MCP, context injection and fine-tuning - The four grounding moves in v4.0.0 - Step by step: how the assistant validates a query without guessing - Claude Code plugin architecture and why “global” is the real change - The pattern any agent platform should copy - What to take from it ## Continue reading [Guides](/categories/guides) · September 23, 2026 [### 5 Ways to Automate SEO With AI Agents (and Wire Them to Your Data)](/blog/guides/automate-repetitive-seo-with-ai-agents) Automate SEO with AI agents across five concrete jobs, from a daily briefing to hreflang sitemaps, plus how to connect agents to your data safely. Carlos Arias · 6 min [Guides](/categories/guides) · August 19, 2026 [### AI Agent Readiness: The Data Foundation Most Teams Skip](/blog/guides/ai-agent-readiness-foundation) AI agent readiness is a data problem, not a model problem. A checklist for the knowledge base, connected data, and rules an agent actually needs. Carlos Arias · 6 min [Guides](/categories/guides) · August 8, 2026 [### AI Agent Context Configuration: How It Shapes Output Quality](/blog/guides/context-shapes-ai-agent-output) AI agent context configuration — not the task prompt — is the primary lever for output quality. Mapped to AstroAgent's real artifacts. Carlos Arias · 7 min ## Stay in the loop. One email when it’s worth it — new posts and updates, no spam. Thanks — check your inbox to confirm. Free. Unsubscribe in one click.