Connect an AI (DeepSeek, Claude…)

The AI reads the bill of quantities and answers per wall type. On the sample: 141.9 m² of EXT-01 larch siding and 98.4 m² of EXT-02 galvalume, plus 41 m of base flashing. Ask the same way for room areas, window counts, the 壁量 ratio or the estimate total.

The AI switches the frame system to 105: posts, 土台, beam widths and braces follow, including the one post the sample had at 120×120 — about a hundred members in one undo step. The structure check runs again by itself.

The AI finds the 2F sliding windows, searches the catalogue for the same size in the series you named, and applies the product: maker, code, size, U-value and frame colours update, the position stays. Tested on the sample with two 1,690 × 2,030 引違い窓 (LIXIL サーモスII-H → YKK AP APW 430 16520).

The AI moves the 2F east wall out by 910 mm and stretches the north and south walls, the 2F floor and the roof with it. Then the check tells what the change means for the structure — posts of the overhang that stand on nothing, a floor beam that is now too weak — each with its fix, and the AI asks before applying them. Tested on the sample.

The AI runs the structure checks and uses the fixes that respect your condition: bracing panels or braces, belt beams 胴つなぎ, stronger wood, rigid floors. Tested: with the bracing taken off five 1F walls (壁量 Y 0.10, NG) it added 構造用合板 in 16 bays — no posts — and every check passed. In harder cases (a large cantilever) some rows may stay NG; the AI then lists the options (a 105×210 post, glulam, a bigger post) and asks you. It never makes posts bigger without asking.

With an AI app that reads PDFs (Claude Desktop, Claude Code) connected through MCP: the AI reads the grid, dimensions, room names and window sizes, then lays out the floor — walls, rooms, doors, windows and the floor slab — in one step, and fixes what the check reports. Tested with a 1F plan sheet: 8 rooms, 8 doors and 8 windows, room areas equal to the PDF to 0.01 m². Check the result against the drawing; handwritten or scanned plans are harder.

With an AI app that can read your files (Claude Code, Claude Desktop with file access): the AI makes the project, sets the stories and the grid, the address (the region for snow, wind, energy and the permit follows from it), lot number, land area and GPS position, draws the site boundary, puts the survey heights on the terrain and reads the SWS ground survey table. Tested with a folder of text and CSV files.

The built-in AI assistant cannot read attached files or pictures yet. Until then use an AI app connected through MCP for the two examples above.

You pay the provider per token (about ¾ of an English word or one Japanese character). The assistant sends a summary of the project with each request, so one question that changes the model uses some 20,000–100,000 tokens: from a few yen (DeepSeek) to a few hundred yen (the largest models) per working session. Local models cost nothing. Prices change often — check the provider's page.

The assistant says what happened: the key was refused (save it again), no credit left (top up), the model name is unknown (check it in AI settings), too many requests (it waits and tries again), the model cannot use tools (pick another), or the saved key belongs to another server address (save it again).