Example prompts, 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 now builds a compact versioned promptBundle by default. It contains the agent contract, model loop, compact Sheet core, and domain overlay. The client then uses confbuild_get_prompt_sections to load the selected profile and only the semantic knowledge packs needed for its planned row types.

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

Copy-ready prompts: what you can ask confBuild to do

Ask for the outcome in normal language. The agent selects the MCP tools, validates every saved change, and normally renders the result for review. Better prompts name dimensions, materials, adjustable parameters, must-keep details, acceptance criteria, and desired files. Customer-adjustable numeric parameters default to sliders with sensible minimum, maximum, and step values; free number or text fields are reserved for cases without safe bounds or where arbitrary exact entry is required.

Generate a new parametric model

Create a parametric belt conveyor in confBuild. Make length, belt width, working
height, speed, and support spacing adjustable. Include the frame, legs, rollers,
drive, guards, materials, and a BOM. Validate the workbook, save it, render front,
right, top, and perspective views, inspect the images, and repair visible defects.
Return the editable project link.

Use the same pattern for a building, structural frame, furniture item, printable part, fixture, machine, skid, enclosure, or other configurable product:

Build a 20 m × 40 m steel portal-frame hall with adjustable bay spacing, eaves
height, roof pitch, cladding, doors, and windows. Add a cutaway review view and a
bill of materials. Keep all structural values explicitly marked as preliminary.

Design a printable wall bracket for FDM. Parameters: width 80 mm, height 120 mm,
wall thickness 4 mm, four countersunk holes, 0.4 mm nozzle, and no unsupported
overhang above 45 degrees. Render it, check printability, and export STL and 3MF.

Continue, inspect, or clone a project

Continue this project: https://app.confbuild.com/e/PROJECT_ID
Change only the frame width to 1,400 mm and add two cross-braces. Preserve every
other sheet, formula, script, and output ID. Validate, commit, render the changed
area, and compare it with the previous result.

Inspect https://app.confbuild.com/e/PROJECT_ID without changing it. Summarize its
parameters, sheets, materials, BOM readiness, unresolved references, validation
risks, and the parts that appear detached or colliding in a four-view render.

Clone this public confBuild template into my account, rename the copy “Customer A
Concept”, replace the dimensions with the values below, and leave the source
untouched: …

Review, repair, and recover

Find every row containing “drive_motor”, explain which sheets reference it, and
change only the motor power and mounting offset. Show the pending workbook state,
validate it, then save and render a close-up of that subassembly.

Render front, right, top, and perspective views of this project. Use the geometry
diagnostics to check collisions, detached parts, outliers, model bounds, and whether
the last patch changed the scene. Fix only confirmed defects and rerender.

List the rollback snapshots for this project and restore the snapshot from before
the last commit. Do not overwrite anything until you have shown me the selected
snapshot and the current revision.

Create CAD, manufacturing, and commercial outputs

Export this committed project as STEP, STL, GLB, and 3MF. Also deliver BOM JSON and
CSV. If an artifact is too large for the inline result, retrieve all chunks and
reassemble the original file.

Generate a vector drawing-set PDF with standard views and a sheet-metal flat-pattern
DXF. Check that every drawing page is populated and that hidden inactive sheet-metal
parts are excluded from the DXF.

Create a cost calculation as JSON, a customer quote as PDF, and purchasing lists as
JSON and CSV. Keep currencies separate, report price coverage, and stop the quote if
required prices or seller identity are missing. Do not invent prices or exchange
rates.

Compare saved variants

Compare these saved confBuild variants by total price in EUR, then by mass, generic
cradle-to-gate CO2e, X/Y/Z extent, and bounding volume: [URL A], [URL B], [URL C].
Show coverage and warnings for every row, keep failed variants visible, and never
treat missing values as zero.

Complete prompt-to-tool map

Users normally do not need to name tools. The following examples show which request should cause the client to use each of the 29 hosted MCP tools.

MCP tool	Example user prompt
`confbuild_start_design_session`	“Plan and build a configurable workbench with adjustable width, height, drawers, and material.”
`confbuild_get_prompt_sections`	“Use the machine profile and load only the geometry, assembly, and catalog rules needed for this design.”
`confbuild_get_prompt_bundle`	“Before editing, show which versioned confBuild rule bundle applies to a structural-hall model.”
`confbuild_list_prompt_resources`	“List the available confBuild prompt resources with their hashes so I can audit what the agent uses.”
`confbuild_auth_status`	“Check whether my confBuild MCP connection is authenticated and tell me which account context is active.”
`confbuild_resolve_project_reference`	“Resolve this project/configuration URL and tell me whether it is owned, public, editable, and which saved configuration it targets.”
`confbuild_create_project`	“Create a new private parametric conveyor project in my confBuild account.”
`confbuild_clone_project`	“Clone this public template into my account and leave the source unchanged.”
`confbuild_read_project`	“Read this project without changing it; return only Main Part and Frame plus the project script.”
`confbuild_begin_edit`	“Open this owned project for a revision-protected edit and preserve the complete current workbook.”
`confbuild_apply_sheet_patch`	“Change cells C2 and C5 in Main Part and add one brace row without replacing unrelated sheets.”
`confbuild_validate_edit`	“Validate the pending edit for formula errors, bad references, duplicate IDs, invalid row types, and engine traps.”
`confbuild_read_edit_workbook`	“Show me the current unsaved Frame sheet from the active edit session.”
`confbuild_commit_edit`	“Save the validated edit atomically, create a rollback snapshot, and rename the project to ‘Conveyor 1400’.”
`confbuild_list_project_snapshots`	“List the newest rollback snapshots for this project with their dates and revisions.”
`confbuild_restore_project_snapshot`	“Restore the selected pre-commit snapshot as a new revision, while preserving the current state as another snapshot.”
`confbuild_discard_edit`	“Discard my uncommitted edit session; do not change the saved project.”
`confbuild_browser_capabilities`	“Tell me whether rendering will use an open browser tab or isolated server-side Chromium and which capture features are available.”
`confbuild_render_project`	“Start a background render of front, right, top, and default views and give me the job ID.”
`confbuild_get_render_result`	“Wait for that render, return the images and geometry diagnostics, and compare it with the previous iteration.”
`confbuild_render_and_wait`	“Render this committed project now, wait up to 90 seconds, and return four review views with diagnostics.”
`confbuild_export_project`	“Export STEP plus BOM CSV,” “create the drawing PDF and flat-pattern DXF,” or “generate cost, quote, and purchasing outputs.”
`confbuild_get_export_result`	“Wait for the export job and return the file inline when it is small enough.”
`confbuild_get_export_chunk`	“Download every chunk of the large completed STEP export and reassemble it without exposing private storage paths.”
`confbuild_compare_variants`	“Compare these three saved variants by EUR price, mass, CO2e, or scene bounds and rank only adequately covered values.”
`confbuild_get_variant_comparison_result`	“Wait for the comparison and show rankings, deltas, coverage, warnings, and per-variant failures.”
`confbuild_find_rows`	“Find every row that contains ‘motor’ in the type and identifier columns without reading every sheet in full.”
`confbuild_list_edit_sessions`	“List my open edit sessions and recover the one for this project after the previous client disconnected.”
`confbuild_finish_design_session`	“Finish the design session only after the final commit and render; return the project link and remaining limitations.”

Exact parameters, export formats, coverage rules, and safety limits are documented in the MCP tool reference.

What the prompt bundle contains

Component	Purpose
MCP agent contract	Requires the client to plan, generate Sheets, call tools, and analyze screenshots.
Compact generation core	Provides general Sheet, formula, geometry, and assembly rules without the complete table prompt.
Profile extension	Loaded as the required first deferred section for `building`, `machine`, `3dprint`, `structure`, or `furniture`; `generic` uses the core.
Domain overlay	Adds deterministic MCP rules for the chosen model class.
Semantic knowledge packs	`foundation`, `geometry`, `systems`, `assembly`, `catalog`, and `completion`, reproducibly exported from the same `SystemPrompt` logic.
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
structure: halls, steel construction, portal frames, trusses, crane runways
furniture: furniture, cabinets, shelves, tables, workbenches
generic: products or models without a specialized classification

Client-specific prompt

client: codex loads OpenAI/Codex-oriented knowledge packs as needed. client: claude receives the Claude variants. generic uses the general OpenAI-compatible base. Complete table prompts remain available as resources and in the compatibility modes.

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.

Four detail levels

`promptDetail`	Contents	Best for
`manifest`	Metadata, sections, and hashes without full bundle text	Inspection, caching, and version checks
`progressive`	Compact base plus a plan for targeted section loading	Default for direct MCP clients
`essential`	Agent contract, complete table prompt, and matching profile	Compatibility for clients without progressive loading
`full`	Also includes component, assembly, and review prompts	Complex assemblies or missing specialist rules

For a machine request, the first payload drops from roughly 259,000 to roughly 45,000 characters. The required machine profile adds about 47,000 characters; large catalog and specialist-system rules stay unloaded until needed. Explicit essential and full calls remain backward compatible.

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/openai/pack/geometry
confbuild://prompts/prompt-editor/openai/pack/systems
confbuild://prompts/prompt-editor/openai/pack/catalog
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 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;
six semantic prompt packs for OpenAI/Codex and Claude;
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 in the OAuth UID path
Capture PNGs	Open confBuild project tab
Relay private render jobs and PNG results	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 the default promptDetail: progressive, targeted section loading, 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