Project: a branded Claude Code plugin that turns reference images of houses into a design, drawings and a schedule of accommodation. No financials.
- Decision: the product takes pictures and sketches of houses the user likes and produces a design in that style. The money, appraisal and go/no-go elements are removed.
- Reasoning: Sanjiv corrected the direction. The real use case is taste-led design (matches his opening message about having elevations of a house he likes), not commercial appraisal.
- Alternatives considered: feasibility and go/no-go engine (rejected by user); full lifecycle product (rejected as too large).
- Decision: ship as a branded Claude Code plugin with one orchestrating command, run inside Claude Code.
- Reasoning: user chose "skill pack only" for lowest build and fastest path, accepting that the team uses Claude Code.
- Alternatives: branded web app (deferred, possible later for resale); configured Claude agent or Project (rejected, looser output).
- Decision: keep outputs 1 (scheme) and 2 (schedule of accommodation) from the original five, drop 3 (appraisal), 4 (planning check) and 5 (go/no-go). Add visual drawings (elevations, plans, 3D) as a core output.
- Reasoning: user said "remove the money bit, but 1 and 2", and the whole point is producing the design they can see.
- Decision: produce styled exterior elevation images (the look) and simple floor-plan line drawings, plus a basic 3D model.
- Reasoning: "copy the houses they like" implies the output should resemble the references (favours image generation), while a usable design also needs plans (favours line drawings). Both serve the brief.
- Status: Open Question OQ2, default chosen so the build is not blocked.
- Decision: Claude vision to read references; existing AEC skills for analysis and layout; Higgsfield generate_image MCP for styled images; a small SVG/DXF script for plans (DXF gives a real CAD file); Blender MCP or existing SketchUp script for 3D; xlsx skill for the schedule.
- Reasoning: uses what is already installed and tested. DXF via ezdxf answers the user's CAD question with a genuine CAD output at schematic quality.
- Alternatives: ecosystem skills (eachlabs floor-plan-generation, interior-design-visualization) noted as future options.
- Decision: ~/Documents/Claude/Code/house-design-studio (code location per user rules; never the Desktop).
- Reasoning: it is code. Business entity not yet confirmed, see OQ1; can be moved or re-homed later.
- Decision: requirements.txt uses minimum pins (ezdxf>=1.3, openpyxl>=3.1, pytest>=8.0) rather than exact pins.
- Reasoning: the machine runs Python 3.14.4, which is newer than the spec's exact pins. Minimum pins let pip pick 3.14-compatible wheels. Installed clean: ezdxf 1.4.4, openpyxl 3.1.5, pytest 9.0.3 (all pure-Python wheels, no build step).
- Alternatives: exact pins (rejected, risk of missing 3.14 wheels and a Cython build failure for ezdxf).
- Decision: the floor-plan SVG widens its canvas to fit the heading and uses a 13px title, instead of a fixed plan-width canvas with a 16px title.
- Reasoning: at the eyeball step the heading overflowed the right edge on narrow plans. Visual quality is the point of this product, so the heading must stay on-canvas. Tests are unaffected (they assert room rectangles and labels, not the title metrics).
- Decision: add a full M&E module (electrical + plumbing + heating/ventilation) that produces schematic services layouts (SVG + DXF) per unit and level, plus M&E schedules, driven by the building-services skill. Built after the design MVP, drawn on top of the existing floor-plan engine.
- Reasoning: client asked for "all the electrical plans, everything"; Sanjiv chose options 1 and 2 (electrical now + full M&E now). Electrical and other services layouts are drawn on the floor plans, so the existing plan engine is the base layer.
- Depth: schematic, design-and-spec stage only. Not certified construction or Part P drawings; an M&E engineer or electrician develops and certifies them. Stated in the pack's limits note.
- Architecture: one shared services framework (hds/services.py) classifies rooms and returns placed M&E items tagged by discipline; one parameterised SVG renderer, one DXF renderer, and one schedule module filter by discipline. Not three separate copies.
- Alternatives: electrical only (superseded by "1 and 2"); defer M&E to a much later phase (rejected, client wants it now).
- OQ1: Product and brand name, and which business it sits under. Needs Sanjiv.
- OQ2: Drawing style, styled images vs line drawings vs both. Default: both.
- OQ3: One bespoke house vs several of one house type on a site. Default: support both, ask at intake.
- OQ4: Approve use of the Higgsfield image MCP (costs credits). Default: yes.
"Copy" is implemented as "in the style of". The tool takes inspiration and adapts; it does not reproduce a specific protected house design exactly. Flag to revisit if the client wants closer replication.
- Completed: brainstormed the concept through the brainstorming skill; agreed a design-from-references house tool delivered as a Claude Code skill pack; wrote this spec and notes.
- In progress: spec written, awaiting Sanjiv's review.
- Blocked: none. Four open questions have working defaults so a plan can proceed.
- Next: Sanjiv reviews the spec, then move to the implementation plan (writing-plans skill).