Building Custom GPTs and Claude Projects for AEC

A custom GPT or Claude Project is a pre-configured AI assistant with your firm's documents, standards, and instructions already loaded. Instead of pasting your code analysis template into ChatGPT every morning, you open a GPT that already knows you are an architect in Austin, Texas, working under IBC 2021 + City of Austin amendments, and responds in your firm's voice. For a working AEC professional, this is where AI productivity actually compounds. This lesson shows how to build the three most useful AEC GPTs/Projects.

What You'll Learn

• The difference between Custom GPTs (OpenAI) and Claude Projects (Anthropic)
• How to build a Specification Assistant, a Code Assistant, and an RFI Assistant
• What to upload vs. what to write as instructions
• How to keep these tools current over time

GPT vs. Project: Which to Use

Custom GPTs (ChatGPT) — Available on the Team and Enterprise plans. Each GPT has instructions, knowledge files (up to 20 files, 512 MB each), and optional actions. GPTs are shareable within your org. They respond in ChatGPT's standard interface. Best if your firm is already on ChatGPT Team.

Claude Projects (Anthropic) — Available on Claude Pro, Team, and Enterprise. Each Project holds custom instructions and uploaded files (with a generous context window — roughly 500,000 characters of project knowledge). Projects are conversations that inherit context. Best if your firm is on Claude Team or Enterprise, or if you need to work with very long documents (codes, specs, reports).

Most AEC firms use both: Claude Projects for long-document work, Custom GPTs for quick structured tasks.

Build #1: The Specification Assistant

Goal: a GPT/Project that helps you edit spec sections in your firm's style.

Instructions (write these as the system prompt):

You are a CSI-trained specification writer for {Firm Name}, a mid-sized architecture firm based in {city}. Your job is to draft and coordinate specs in CSI MasterFormat three-part format.

Always use the firm's standard boilerplate and formatting conventions (attached). Prefer performance language over prescriptive when both are acceptable. Never invent manufacturer model numbers — use "Approved manufacturers include [X], [Y], [Z] or approved equal." Always reference current ASTM, ANSI, NFPA, UL, ASHRAE, and ADA standards by number, not by language.

When the user asks for a spec section, respond with: (1) the three-part draft, (2) a list of assumptions I made, (3) a list of project-specific values the user needs to fill in. Cite every referenced standard.

Knowledge files to upload:

• Firm spec master table of contents
• 3-5 representative firm-style spec sections (so the AI matches your firm's voice)
• CSI MasterFormat 2020 section index
• Firm style guide for specs (if one exists)

Typical uses:

• "Draft Section 07 21 00 Thermal Insulation for a Class A office in Climate Zone 4A."
• "Compare this contractor-submitted spec section against our firm's current master 08 43 13."
• "I have this hand-written spec note. Format it properly."

Build #2: The Code and Jurisdiction Assistant

Goal: a GPT/Project tuned to a specific jurisdiction's code package.

Instructions:

You are a code consultant specializing in {City, State} amendments to IBC 2021, IEBC 2021, IMC 2021, IPC 2021, and IECC 2021. Your job is to answer plan-check and design questions based on the code documents loaded into this project.

Always cite the specific code section number. Always note whether the applicable provision is in the base IBC or a local amendment. If the user asks a question where the answer is not in the loaded documents, say so explicitly — do not guess. When interpretation is ambiguous, suggest the user contact the AHJ directly.

Knowledge files to upload:

• IBC/IEBC/IMC/IPC/IECC as PDFs (if your code subscription allows)
• Your jurisdiction's local amendments document
• The jurisdiction's published interpretations or FAQs
• Fire marshal guidance documents
• Historic preservation guidelines if applicable

Typical uses:

• "I have a 4-story Group R-2 building, Type V-A, in NYC. What is the allowed base area and height, and what are the increases for sprinklers?"
• "What is the egress width required for an assembly space of 312 occupants per NYC Building Code?"
• "Is this plumbing fixture count sufficient for a B-occupancy under NYC Plumbing Code?"

A jurisdiction-specific code GPT is by far the highest-value AI asset an AEC firm can build. It takes a Friday afternoon to assemble and it saves hundreds of hours a year.

Build #3: The RFI and Submittal Assistant

Goal: a GPT/Project that drafts CA correspondence in your firm's voice.

Instructions:

You are a senior construction administration architect for {Firm Name}. Your job is to draft professional, firm-voiced responses to RFIs, submittals, and field issues.

Tone: firm but collaborative; never defensive; never apologetic. Always cite the specific drawing sheet, detail, or spec section that answers the question. When a request falls outside the contract, state so clearly and offer an Additional Services Request path. Never commit the owner to extra work. Never acknowledge fault. When information is ambiguous, direct the contractor to consult the specific contract document and follow up.

Knowledge files to upload:

• 10-20 examples of high-quality RFI responses from completed projects (anonymized)
• The firm's preferred sign-off and signature block
• Any firm CA playbook or standards manual
• The current project's contract form identification (AIA A201, ConsensusDocs, etc.)

Typical uses:

• "Draft a response to this RFI: [paste]"
• "Review this shop drawing submittal against the structural drawings and suggest action."
• "Draft an ASI resolving the conflict between spec 09 29 00 and drawing A-301."

What to Upload vs. What to Put in Instructions

A practical rule: facts go in files, behavior goes in instructions.

• Your firm's spec boilerplate? File.
• Your firm's voice and tone rules? Instructions.
• The IBC PDF? File.
• The rule "always cite the code section number"? Instructions.
• A past RFI response as a style exemplar? File.
• The rule "never commit the owner to extra work"? Instructions.

Keeping These Assistants Current

The biggest risk with a GPT or Project is that it drifts from reality. Practical maintenance:

• Monthly: check for new code editions, adopted amendments, or firm master spec revisions.
• Quarterly: review the instructions — has firm policy changed?
• After every project: add 2-3 new representative examples to the knowledge files.
• Annually: refresh all knowledge files and run a sanity-check set of 20 prompts through the GPT to verify it still produces good output.

Assign one person in the firm as the "owner" of each assistant, otherwise they rot.

Sharing Within the Firm

For shareable assistants:

• Custom GPTs: share within ChatGPT Team workspace
• Claude Projects: share within Claude Team/Enterprise workspace
• Document the assistant's purpose and intended use in a short internal wiki page
• Include a "do not use for" list — e.g., do not use the spec assistant for Division 00 front-end documents

Where Not to Build an Assistant

Do not build a GPT/Project for:

• Stamped calculations. Your software is the tool of record.
• Safety-critical life safety review. A human does this.
• Legal contract interpretation. A lawyer does this.
• Final engineering design. Your engineering process remains the authority.

A GPT/Project is a productivity tool. It is not a shortcut around the licensed professional's responsibility.

Example: A One-Afternoon Build

A realistic afternoon:

• 1:00 pm Decide on the scope of the assistant (say, "Texas Commercial Code Assistant").
• 1:15 pm Draft the instructions (200-300 words).
• 1:45 pm Upload IBC 2021, Texas amendments, and your city's amendments.
• 2:15 pm Add 5 example Q&A pairs to train voice.
• 2:45 pm Run 15 sanity-check prompts; note where it fails.
• 3:15 pm Refine instructions based on failures.
• 4:00 pm Share with the team.

Total: 3 hours. Payoff: hundreds of hours across the year.

Key Takeaways

• Custom GPTs and Claude Projects pre-load your firm context so every prompt starts smarter
• The three highest-ROI AEC assistants: Spec Assistant, Jurisdiction Code Assistant, RFI Assistant
• Facts belong in uploaded files; behavior belongs in instructions
• Assign an owner and a monthly/quarterly maintenance cadence
• Never build an assistant that ends up replacing a licensed professional's judgment