Agent skill | Deck Twin

Deck Twin ships with an agent skill: a SKILL.md and reference files. With it, a coding agent can:

find and install the right component for a request ("add a seawater lift pump");
assemble a whole plant from any mix of documents: P&IDs, PFDs, datasheets, equipment and line lists, plot plans, descriptions and notes. It extracts the equipment, maps it to components, sizes it, lays it out, wires the ports, checks the result and renders it.

The steps are explicit and every name is looked up through the CLI, so a small, inexpensive model can do the assembly. A larger model can prepare the equipment list.

Install the skill

npx deck-twin skill

This writes .claude/skills/deck-twin/: SKILL.md plus the references/ files. Claude Code loads skills from there automatically.

Other agents (Cursor, Codex, Gemini CLI, custom agents): run npx deck-twin skill --dir <folder> and point the agent at SKILL.md, or paste its contents into the system prompt.
Read it online: /skill/SKILL.md.

What the skill contains

File	Purpose
`SKILL.md`	The commands, `validate`, the two workflows and the rules
`references/component-map.md`	Maps P&ID symbols and equipment names to registry names
`references/sizing.md`	Rules of thumb for pressure class, line size, vessel, pump and driver sizes when the drawing is silent
`references/layout.md`	Deck levels, flow order, spacing, partners, pipe racks, cranes
`references/documents.md`	How to read each kind of document: drawings, lists, datasheets, plot plans and notes
`references/plant-json.md`	The schema, with the rules the checker and renderer enforce
`references/pid-symbols.md`	Generated: every component's P&ID symbol, tag letters, ISA codes and datasheet → props mapping
`scripts/tile.py`	Cuts large P&ID scans into legible tiles
`references/examples.md`	Six worked plants from basic to advanced: well pad, control room, wellhead platform, water injection module, separation train, crude export terminal

From P&IDs

The skill reads P&IDs supplied as images or PDFs.

Large sheets: they are tiled first (scripts/tile.py).
Symbols: each one is matched against the P&ID symbol legend. Process symbols follow ISO 10628-2 / PIP PIC001 conventions, instruments ISA 5.1, electrical IEC 60617, and F&G NFPA 170 / ISO 7010. The drawing's own legend always wins.
Tags: tag letters confirm each match.
Line numbers: they set line size, service, class and insulation.
Datasheets: fields map straight onto props. For example, an API 610 datasheet's driver rated power sets kw.

Every component page shows its symbol, tag letters and datasheet mapping, under P&ID.

Machine-readable endpoints

URL	Contents
`/r/index.json`	Every component with category, description, props (type, range, options), defaults, ports at defaults, bounds at defaults, standards and its P&ID entry (symbol, tags, ISA letters, datasheet → props)
`/symbols/legend.svg`	All P&ID symbols on one sheet, for side-by-side matching
`/r/<name>.json`	One item in the shadcn registry-item format, with file contents
`/r/<name>.md`	The component's page as Markdown: description, standards, install command, usage, props, ports, P&ID entry and source. Each component page has Copy as Markdown and View as Markdown buttons for it
`/llms.txt`	Short catalogue for LLM context
`/r/examples/*.plant.json`	The example plants

Prompting a cost-effective model

Give it the skill, the equipment table and one instruction:

Follow Workflow B in SKILL.md. Use only names from deck-twin list and ports from deck-twin info. Run deck-twin validate plant.json after every change until it reports no errors, then run deck-twin assemble plant.json.

The check loop does most of the work. Every error names the item and tells the model what to change.