Master prompts & AI boundary

Yes: the client receives the prompt-editor prompts

The master prompts are an explicit part of the MCP contract. The client does not need to know them before connecting, and you do not need to paste them into every user request.

On the mandatory first call to confbuild_start_design_session, the server builds a versioned promptBundle. It contains the effective Codex/OpenAI or Claude table prompt, the matching domain extension, and an MCP agent contract. The client reads this bundle and uses it as modeling knowledge for the current session.

Important: “The server provides a prompt” does not mean “the server calls an AI.” The bundle consists of locally read, reproducibly exported text resources. Codex, Claude, or another MCP client is the first component to interpret those texts with a model.

What the prompt bundle contains

Component Purpose
MCP agent contract Requires the client to plan, generate Sheets, call tools, and analyze screenshots.
Effective table prompt Exported directly from the prompt editor’s SystemPrompt logic for OpenAI/Codex or Claude.
Profile extension Adds rules and examples for building, machine, 3dprint, or generic.
Domain overlay Adds deterministic MCP rules for the chosen model class.
Optional specialist resources Component, assembly, analysis, and fix prompts for more complex work.
Hashes and manifest Make the source and version of every section verifiable.

Profiles are inferred deterministically from the request or selected explicitly by the client:

  • building: houses, rooms, halls, roofs, and architecture
  • machine: machines, conveyors, gearboxes, fixtures, and equipment
  • 3dprint: printable parts, FDM/FFF, SLA, SLS, and MJF
  • generic: products or models without a specialized classification

Client-specific prompt

client: codex receives the OpenAI/Codex-oriented table prompt. client: claude receives the Claude version. generic uses the general OpenAI-compatible base.

The prompt editor contains historical instructions such as “return JSON only” or “do not use tools.” In the MCP workflow, those rules apply only to the internally generated Sheet payload. The preceding MCP agent contract takes priority: the client does not return that JSON to the user; it sends it to confBuild through the edit tools.

Three detail levels

promptDetail Contents Best for
manifest Metadata, sections, and hashes without full bundle text Inspection, caching, and version checks
essential Agent contract, effective table prompt, and matching profile Default for almost every modeling task
full Also includes component, assembly, and review prompts Complex assemblies or missing specialist rules

essential keeps the context—and therefore client cost—small. The client can later read individual optional resources instead of loading everything preemptively.

Available prompt resources

confbuild_list_prompt_resources lists each URI, MIME type, character count, and SHA-256 hash. Important resources include:

confbuild://prompts/agent-contract
confbuild://prompts/prompt-editor/openai-table
confbuild://prompts/prompt-editor/claude-table
confbuild://prompts/prompt-editor/profile/building
confbuild://prompts/prompt-editor/profile/machine
confbuild://prompts/prompt-editor/profile/3dprint
confbuild://prompts/prompt-editor/agent-analysis
confbuild://prompts/prompt-editor/agent-apply-fixes
confbuild://prompts/component-mode
confbuild://prompts/assembly-mode
confbuild://schema/workbook

The master and extension sources are also available as local MCP resources for troubleshooting. During normal operation, the client should use the exported effective prompt instead of reinterpreting the TypeScript sources.

Where the prompts come from

Effective prompt artifacts are exported reproducibly from the same sources used by the confBuild prompt editor. This avoids a manually maintained MCP copy that can quickly become stale.

The export produces, among other artifacts:

  • OpenAI and Claude table prompts;
  • action-object prompts;
  • agent-mode analysis and apply-fixes prompts;
  • profile extensions for buildings, machines, and 3D printing;
  • routing and artifact manifests with hashes.

When runtime prompts change, refresh the export in the same change set:

npm run mcp:confbuild:prompts
npm run mcp:confbuild:test

The test fails when generated artifacts no longer match the prompt-editor sources.

Who performs each AI task

Task Responsible component
Understand user intent MCP client
Decompose the model into Sheets MCP client
Generate formulas and geometry data MCP client
Plan tool calls and iterations MCP client
Provide prompt texts and rules MCP server
Apply and validate Sheet patches MCP server
Save the confBuild project MCP server + Firebase under user permissions
Open the browser and capture PNGs MCP server
Interpret screenshots MCP client
Professional approval Responsible human

Cost reality

The MCP server creates no additional LLM request to confBuild or a model provider. Costs may still come from:

  • your Codex/Claude plan or API usage;
  • normal Firebase, network, or browser infrastructure;
  • unusually long client contexts and many visual iterations.

Use promptDetail: essential, filtered Sheet reads, localized patches, and focused screenshot iterations to reduce cost. Smaller prompts must not lead the client to read only part of a Sheet it is about to replace.

Next step

The browser and screenshot loop explains how the client performs visual AI work. Exact prompt tools and resources are summarized in the tool reference.

Back to Documentation