User story
As a user directing an AI agent to make a small, well-specified content change in a Webstudio project (e.g. inserting one authored component into an existing container), I expect the change to happen quickly, without a long chain of preparatory discovery calls each round-trip.
Summary
Server-provided tool instructions (meta.index guidance) direct the agent to perform multiple sequential discovery calls before any mutation, even when the user's request is already narrow and well-specified. For a single component-insertion task the agent ended up calling meta.index, meta.guide, components_search, meta.get-more-tools, inspect-design-context, and a resource read (project/expressions) before attempting the first mutation, and the user explicitly objected that this was unnecessary (they had not asked for a token/design-system audit) and that the overall process was taking unnecessarily long.
What the agent tried
- meta.index for orientation
- meta.guide with a detailed brief to confirm two specific technical questions
- components_search for timer/interval-related components
- meta.get-more-tools for insert-fragment/bind-props/create-variable schemas
- inspect-design-context, enumerating all pages, breakpoints, ~97 design tokens, assets, and variables
- resource read of project/expressions
- first insert-fragment dry run attempted, at which point the user interrupted
Expected behavior
For a bounded, clearly-scoped authoring task (insert one described component structure into a known container, wire a couple of variables/interactions), the number of mandatory discovery round-trips before the first mutation attempt should be minimal (ideally 1, at most 2), proportional to the task's actual ambiguity. Confirming open technical questions should not require a full project-wide design context dump when only a narrow yes/no answer about a mechanism was needed.
Actual result
The agent made 6+ discovery/orientation tool calls, several returning large payloads unrelated to the immediate question (e.g. a full list of ~97 existing design tokens and 18 pages) before reaching the first proposed mutation. This visibly elapsed real wall-clock time for the user and triggered a direct complaint about process speed, plus a correction that a specific discovery step (token enumeration) was not asked for.
Recovery attempts
- None attempted within session; user asked the agent to stop and file this report instead of continuing to optimize the discovery sequence ad hoc.
User impact
Users directing agentic edits on a live project perceive the tool as slow/over-cautious for simple tasks, and can lose trust when the agent appears to go off-script (broad discovery) relative to the specific, narrow thing they asked for. Especially costly for delegated/non-interactive sessions where each round trip has real latency and no visibility into intermediate reasoning.
Technical context
MCP server 'webstudio', reached via Claude Code's MCP tool integration. meta.index's bundled instructions recommend a proportional workflow but the practical effect in this session was several chained discovery calls, including a full inspect-design-context bundle, before the first mutation was attempted. No errors or crashes occurred; this is a workflow-efficiency/guidance issue rather than a functional bug.
Agent environment
- Client: claude-code-cli
- Provider: unknown
- Model: claude-sonnet-5
- Reasoning effort: low
- Trigger: user-requested
- Category: other
Technical runtime
- CLI: 0.291.0
- Node.js: 22.11.0
- Operating system: darwin 23 (arm64)
- Execution mode: mcp
- API contract: public-api:iqd6ok
Acceptance criteria
- Guidance/tool descriptions make it clearer when a narrow, targeted read is sufficient versus when a full inspect-design-context bundle is warranted
- For small, unambiguous authoring tasks, the recommended number of discovery calls before the first mutation attempt is reduced
- Documentation or tool descriptions help the agent avoid pulling broad project inventories (all tokens, all pages) when the user's question is narrowly scoped
User story
As a user directing an AI agent to make a small, well-specified content change in a Webstudio project (e.g. inserting one authored component into an existing container), I expect the change to happen quickly, without a long chain of preparatory discovery calls each round-trip.
Summary
Server-provided tool instructions (meta.index guidance) direct the agent to perform multiple sequential discovery calls before any mutation, even when the user's request is already narrow and well-specified. For a single component-insertion task the agent ended up calling meta.index, meta.guide, components_search, meta.get-more-tools, inspect-design-context, and a resource read (project/expressions) before attempting the first mutation, and the user explicitly objected that this was unnecessary (they had not asked for a token/design-system audit) and that the overall process was taking unnecessarily long.
What the agent tried
Expected behavior
For a bounded, clearly-scoped authoring task (insert one described component structure into a known container, wire a couple of variables/interactions), the number of mandatory discovery round-trips before the first mutation attempt should be minimal (ideally 1, at most 2), proportional to the task's actual ambiguity. Confirming open technical questions should not require a full project-wide design context dump when only a narrow yes/no answer about a mechanism was needed.
Actual result
The agent made 6+ discovery/orientation tool calls, several returning large payloads unrelated to the immediate question (e.g. a full list of ~97 existing design tokens and 18 pages) before reaching the first proposed mutation. This visibly elapsed real wall-clock time for the user and triggered a direct complaint about process speed, plus a correction that a specific discovery step (token enumeration) was not asked for.
Recovery attempts
User impact
Users directing agentic edits on a live project perceive the tool as slow/over-cautious for simple tasks, and can lose trust when the agent appears to go off-script (broad discovery) relative to the specific, narrow thing they asked for. Especially costly for delegated/non-interactive sessions where each round trip has real latency and no visibility into intermediate reasoning.
Technical context
MCP server 'webstudio', reached via Claude Code's MCP tool integration. meta.index's bundled instructions recommend a proportional workflow but the practical effect in this session was several chained discovery calls, including a full inspect-design-context bundle, before the first mutation was attempted. No errors or crashes occurred; this is a workflow-efficiency/guidance issue rather than a functional bug.
Agent environment
Technical runtime
Acceptance criteria