source-05-tool-schema-x-leads

Here’s what a targeted search across X turned up for people discussing the three problem areas you’re investigating. Results are from recent posts (mostly mid-2025 to August 2026) and focus on those whose content actually matches the research questions rather than generic AI chatter.

1. Enum-value / parameter naming, schema semantics, and descriptions

These users have posted directly about how names, enums, descriptions, and schema design affect tool/parameter selection:

User	Handle	Relevance
Adaline	@tryadaline	Explicitly lists “Misleading parameter names” as a core failure mode. States that parameter names carry independent selection signal separate from descriptions. Also covers ambiguous overlap and missing constraints.
Mr Alonso	@buildwithalonso	“El schema es UX para un LLM. Si el nombre no dice QUÉ y la description no dice CUÁNDO… el agent alucina la llamada.” Directly on naming + description effects.
mikaeru	@mikaeru676523	Notes that JSON Schema description/enum act as the real prompt; schema only prevents format errors, not plausible-but-wrong arguments.
Build with Abdul	@buildswithabdul	“the tool schema is the whole game… a vague description and the model just never calls the tool, or calls it wrong.”
nexus	@nexusx2026	“most agent failures aren’t hallucinations. they’re choosing the wrong tool because the description was 30 words of marketing copy instead of 3 examples of when to use it.”
Robat Das Orvi	@orvi_onethread	Points out that schema-valid JSON with a hallucinated enum value still fails at the handler — a constrained-decoding gap.
PostHog	@posthog	Concrete case: inconsistent parameter keys across similar tools (insightId vs id vs short_id) caused agent confusion; fixed by accepting aliases.
manish (AgentWorks)	@manish_iitg	Agent inventing tool names that no longer exist; fixed by tighter naming conventions and clearer error signals.
Dan Mercede	@danmercede	Built schemafit — CLI that lints JSON Schema / tool definitions against each major LLM provider’s constraint surface and suggests repairs.

2. Decision-boundary metadata (“when to use / avoid”, counterexamples, guidance)

Fewer pure hits on the exact terms useWhen/avoidWhen, but several users discuss the decision-boundary problem:

User	Handle	Relevance
Adaline	@tryadaline	Explicitly calls out “Missing constraints: Without a rule for when not to use a tool…” as a top failure mode.
Alex Yarosh	@AleksejAros	“split by decision boundary” when tool calls can’t be disabled/tested independently.
nexus	@nexusx2026	Advocates concrete examples of when to use a tool instead of vague marketing descriptions.
GreatFrontEnd	@greatfrontend	Practical rewrite notes: turning hard “never” rules and worked examples inside tool descriptions into sharper parameters/enums.

3. Constraint models, validation, and invalid combinations

User	Handle	Relevance
Dan Mercede	@danmercede	Most directly on-topic: tooling to validate JSON Schema / structured-output / tool definitions against provider constraint surfaces, with repair suggestions.
Tanat Tonguthaisri	@gastronomy	Shared work on autoformalization of agent instructions into policy-as-code.
z3n	@zench4n	Emphasizes treating tool outputs as untrusted and enforcing strict Pydantic / JSON Schema validation before acting.
OpenAI Developers	@OpenAIDevs	Official posts on Structured Outputs limits (including enum value counts), which is the practical constraint surface many agents hit.
Kimi Developers	@KimiDevs	Discusses “Tool Definition Bloat” and the resulting wrong-tool / invalid-argument problems when many schemas are present.

Notable additional voices

• Taelin (@VictorTaelin) — high-engagement post on the broader difficulty of injecting domain knowledge (relevant to retrieval of guidance/metadata).
• Praveen Singh (@PraveenKS30) — practical discussion of tool-list context bloat and deferred loading patterns.

Summary of coverage

• Strongest cluster: people talking about tool/parameter naming + description quality and the resulting selection errors (Adaline, Alonso, Abdul, nexus, PostHog, manish).
• Constraint/validation side: Dan Mercede stands out as someone actively building tooling in this space.
• Design-system-specific language (variant, component useWhen/avoidWhen, etc.) is rarer on X; most discussion is framed in general tool-calling / agent terms. The mechanisms people complain about map closely to your three research questions, even if they don’t use design-system vocabulary.

Would you like me to:

1. Pull full threads or more posts from any of these users,
2. Run a narrower search (e.g., only posts mentioning “enum” + “tool” + “wrong”), or
3. Search for people discussing design-system + LLM agents more specifically?