The Subscription Killer

Everyone is talking about AI. Almost nobody is telling you what to actually do about it. Most firms pick one of two paths: ignore it and hope it's a phase, or buy three more tools and call it a strategy.

There is a third path. The companies selling you tools have no incentive to mention it.

You can build your own. This is the subscription killer.

Tools tailored to how your firm actually works, connected to your data, solving the friction your team complains about every Friday afternoon. The past-project-finder vendor charges $1,800 a seat. Your version works on your file structure, your conventions, your actual projects. Running cost is a rounding error compared to the vendor invoices you pay now.

The real upside is bigger than the money. You finally get to point a tool at the things that consume time inside your firm: the 20-minute hunt for an old detail, the two hours building proposal precedents, the drawing schedule reconciliation. No SaaS company will ever optimize for those, because they don't know your firm exists.

The goal of this guide: by the end, pick one annoying weekly task. Spend a Saturday on it. End the weekend with a working tool. It doesn't have to be useful the first time. The first one is a test, and the test is for your head: you stop asking "which AI app should we buy?" and start asking "what do we want to build?"

Anthropic publishes a short framework called AI Fluency that names what separates firms quietly winning at AI from firms quietly getting burned. Two parts.

The Four Pillars describe what good AI work looks like: Effective (you got the outcome), Efficient (it didn't burn your week), Ethical (you'd defend the choices to a client), Safe (you accounted for failure modes, including the AI's own).

The 4Ds describe what you actually do.

Here is the part most teams miss. They get good at Description (writing prompts) and they enthusiastically Delegate (use Claude for everything), and then they ship work containing errors nobody caught. The teams that win lead with Discernment and Diligence. Everything else flows from there.

Read more at anthropic.com/ai-fluency. Pass it around your team.

Out of the box, Claude knows everything about everything and nothing about your firm. It will draft a fee proposal, just not your fee proposal. Write a spec, just not in your firm's voice. Analyze a drawing set, just not the way your team has decided to after twenty years of figuring out what actually matters.

The vendor demo never shows this. It shows generic output that looks impressive because you have no baseline. Take it back to your firm and the gap shows up: close but not yours, missing the thing your principal always asks for, citing the wrong standard, using a layout you stopped using three years ago.

You are not training the model. Anthropic trained the model. What you train is the system around it: standards, conventions, spec format, QA protocols, design principles, the way your principals like deliverables written. You do this through the mechanisms in Section 07 (project instructions, CLAUDE.md files, Skills, Plugins). Once well, then maintained.

That layer is the moat. A competitor can't copy it by signing up for the same Claude plan. It encodes how your firm actually works: the lessons, the standards, the "don't ever do this again" notes from past projects. Every project that flows through makes it smarter. The layer compounds.

Claude is three different surfaces stacked on one account, each good at a different shape of work. Most users live in Chat and wonder why they're underwhelmed.

Conversation → Chat. Job → Cowork. Tool → Code.

Models. The picker is a dial. Bigger model, smarter answer, more tokens burned. As of May 2026: Opus 4.7 for deep reasoning, Sonnet 4.6 as the everyday default, Haiku 4.5 for fast bulk tasks. Sonnet handles most professional work. When in doubt, take the latest Sonnet.

Tokens. The unit Claude reads, writes, and bills in. Roughly 5-6 characters of typed text per token, so a page of spec is ~500 tokens. Prompts, replies, and attachments all cost tokens. Run out and work stops mid-sentence.

Paid plans operate on a five-hour rolling window. Agentic surfaces (Cowork and Code) burn tokens far faster than Chat (often by an order of magnitude), because each agent step is its own model call. A few heavy Cowork sessions can flatten a Pro plan in an afternoon. If anyone in your firm uses Cowork or Code as a daily driver, give them Max. Pro is for exploring; Max is for working.

The context window. Claude's working memory for a single conversation. 200,000 tokens standard; a 1M-token beta on Sonnet for API users. Everything lives inside it: messages, replies, attachments, project instructions.

As it fills with stale messages and old files, Claude loses the thread, ignores earlier instructions, and produces strange output. The community calls this "context rot." It's the single biggest reason intermediate users get bad answers without knowing why. Fix: start fresh conversations when you switch topics, upload only what's relevant, and prefer Markdown over PDFs (a heavy PDF can eat 15,000 tokens).

Pricing. Pro $20/mo. Max $100-$200/mo. Team is per-seat with a five-seat minimum and adds SSO/admin controls. Confirm at claude.com/pricing.

One-time setup. Walk through Settings before your first real conversation. Free accounts: chats are used for training unless you opt out. Paid consumer accounts: training off by default. Make the choice consciously if you handle anything under NDA. Then write your personalization paragraph (it travels into every conversation):

The 4Ds from Section 02 show up here as practical skills. Deeper prompting guidance at docs.anthropic.com.

Description: How to Actually Prompt

Think step by step. Adding "think step by step" produces meaningfully better output on anything analytical. For code compliance review, structural assessments, or cost reasoning, structured chain-of-thought is non-negotiable.

Use XML tags. Claude is trained to recognize XML structure. Tags like <context>, <task>, <examples> produce more reliable results than dumping everything in prose. It's just labels on the parts of your request.

Show, don't tell. Examples beat instructions. Rather than describing how you want a fee proposal structured, paste in two finished proposals and say "match this style." Few-shot prompting is one of the highest-leverage techniques.

Put long context at the top. For documents over ~20,000 tokens, place them near the start of your prompt, above your instructions. Dramatically improves retrieval.

Give Claude a role. "You are a senior project architect reviewing this drawing set for code compliance" beats "review this drawing set." The role primes the response.

Iterate. Your first prompt is not the right prompt. Treat prompting as a development cycle: define good output, draft, see where it fails, fix, rerun. Pros iterate three or four times before trusting the output.

Discernment: How to Evaluate Output

The harder half. The more important one.

Quote grounding. For long documents, require Claude to extract direct quotes first, then perform the task. "Find every clause in this contract that addresses change order procedure. Quote each one verbatim with its section number, then summarize." Quotes are auditable; summaries become much harder to fabricate.

Citation and retraction. Have Claude cite sources for every factual claim. Check each cited source actually exists and contains the claim. If it doesn't, retract. Highest-leverage check against hallucination on knowledge work.

Best-of-N. For high-stakes outputs, run the same prompt two or three times. If answers diverge, the model is guessing. Consistency isn't proof of correctness, but inconsistency is proof of a problem.

Adversarial review. After Claude produces something, open a fresh conversation and ask Claude to tear it apart. Find errors, omissions, unsupported claims. Claude is often better at criticizing than creating.

Diligence: Human-in-the-Loop

Anthropic's research on agent autonomy shows the large majority of AI agent tool calls in mature production still have a human in the loop, with safeguards layered in: restricted permissions, approval requirements, review gates. That's the frontier norm. It should be your firm's norm.

The real risk isn't catching Claude inventing things once and learning to distrust it. The real risk is the workflow that almost always works. The team gets comfortable. Verification gets dropped. The failure mode shows up six months later in front of a client. Build the gates early, write them down, and don't let comfort erode them.

Eval guidance: docs.anthropic.com/en/docs/test-and-evaluate.

Four mechanisms do the heavy lifting. Learn these and the rest of Claude falls into place.

Projects and CLAUDE.md

Projects are how you scope context. In Chat, a Project is a cloud workspace with custom instructions and attached files. In Cowork and Code, a Project is a folder on your computer, and the equivalent of Project instructions is a Markdown file at the project root called CLAUDE.md.

If you remember one thing from this guide, remember CLAUDE.md. Plain text. Sits at the root of your project folder. Claude reads it automatically every time it works inside that project. It's the most underused mechanism in the toolchain and the primary place your firm's context gets encoded.

What goes in it: firm conventions, standards, spec format, project-naming rules, QA protocol, the way your principals like deliverables written, the people on the project, the patterns to follow, the mistakes Claude made last time. Anything a new hire would need to be briefed on. Update it after every project. (Use CLAUDE.local.md for personal overrides you don't want in a shared repo.)

Why Markdown over PDFs? PDFs carry layout instructions, font tables, embedded images, and structural metadata that an LLM has to wade through to find your content. Markdown strips all that out. For firm reference material reused across projects, convert once and stop fighting your tokens.

Skills

A Skill is a reusable Markdown bundle that captures a recurring task. If you find yourself typing the same instructions three times, that's a Skill. Examples: your firm's writing voice for technical reports, your design review protocol, your fee proposal assembly. Claude picks a Skill automatically on match, or you invoke it with /. Skills live in the cloud or locally; local ones don't roam between machines.

Connectors and MCP

Connectors bridge Claude to other software. Anthropic ships official connectors for the common SaaS layer (Gmail, Calendar, Drive, Notion, Asana, Canva). Behind them is MCP, the Model Context Protocol — the open standard for AI tools to talk to other software, originally from Anthropic. Connectors cover SaaS; MCP covers everything else, including the design software your firm uses. Section 08 covers the AEC MCP ecosystem.

Plugins

Plugins are how you ship a workflow. Where a Skill is one instruction set, a Plugin wraps several pieces (Skills, connectors, slash commands, MCP configurations) into a one-click install. Cowork-only, not in Chat. Use case: your firm has figured out a tricky workflow and you want every project team using the same approach without retraining.

A Note on Artifacts

When Claude produces structured output (document, slide deck, spreadsheet, small web app, diagram), it returns it as an Artifact rather than dumping it into chat. Supported: Word, PowerPoint, Excel, PDF, HTML, React, SVG, Markdown, Mermaid. In Chat, Artifacts are cloud-stored. In Cowork, local files in your project folder. Cowork's Live Artifacts feature (April 2026) gives you the cloud experience while keeping files local — useful when client work can't go to a cloud service.

Until recently, "talking to your Revit model" or "scripting Rhino in plain English" was a research-paper fantasy. As of 2026, it's a Tuesday afternoon. The AEC MCP ecosystem is young — most servers appeared in late 2025 or early 2026 — but real, and a working setup is feasible today.

Rhino and Grasshopper. Community MCP servers (grasshopper-mcp, rhino-mcp) open a two-way connection to a running Rhino instance: inspect geometry, manage layers, query and manipulate objects, drive Grasshopper components in natural language.

Revit. Autodesk shipped an official AEC Data Model MCP server, and Revit 2027 has built-in MCP support. Claude can read element categories, parameters, geometry, and spatial data, and drive operations like tagging elements or generating views and schedules.

Dynamo. Autodesk has signaled MCP support on the roadmap, with Claude becoming an executable step inside a Dynamo workflow rather than just an external assistant.

Civil 3D and AutoCAD. Both have working community MCP implementations. AutoCAD's gives Claude direct AutoLISP execution. Civil 3D servers expose project data for automated create/modify/delete.

Figma. Official MCP server with design context (components, variables, layout, FigJam) and the ability to write to the canvas or generate code from frames. Useful for firms that use Figma for diagrams, briefs, or research presentations.

A note on stability. Official servers from large vendors mature steadily; community servers move faster but break more often. For firm-wide deployments, prefer official. For prototypes, community is fine.

There is also a separate browser path. Claude for Chrome is an extension that lets Claude operate inside a tab the way a human would: clicking, filling forms, extracting data. It is slower and less reliable than a real MCP connection, but it works on permitting portals, supplier websites, manufacturer catalogs, and other AEC-relevant sites that have no API.

Concrete examples first. Tools AEC firms have built or could realistically build in a weekend. None replace core software; they sit on top of it.

The Subscription Audit

Pull every tool the firm pays for. For each one, ask: load-bearing software (Revit, Rhino, Bluebeam, AutoCAD, ERP), or thin layer over a recurring workflow (fee proposal generator, one-feature QA tool, spec-template service)?

The thin layers are the candidates. Most firms pay $500–$2,000 per seat per year for tools that wrap a workflow they could now own internally. Multiply by seat count. Two or three subscriptions retired pays for the in-house build many times over in year one. And you stop being dependent on a vendor's roadmap, pricing changes, data export policies, or sunset announcements.

Where Claude Code Fits

Claude Code writes, modifies, and reviews software. You describe what you want in plain language; it plans, writes, tests, fixes, and keeps going. Terminal, desktop app, or claude.ai/code.

You still need a baseline: what a code repo is, how Git works, what your tool runs on. Inside firms that have started building, the leads aren't full-time developers — they're technology directors, design tech specialists, computational designers, or operations leads who learned the basics. Treat Claude Code as a thought partner, not a code generator.

The Weekend Playbook

Most of these steps are about capturing context, not writing code. The tool is the easy part.

Your first tool will take longer than expected. Your fifth will be fast.

One scope warning. Don't start with the most ambitious workflow. Start small, bounded, low-risk. Build trust in the process before pointing it at anything sealed, billable, or contractually sensitive.

If You Are Starting Today

Install the Claude desktop app from claude.com/download. The browser version misses Cowork and most of the AEC integration story. Subscribe to Pro for personal exploration; move to Max or Team if your firm is committing to the toolchain.

Spend an hour on one-time setup: privacy settings, personalization paragraph, your first Project with a CLAUDE.md, one Skill that captures your writing voice.

Designate a context owner. One person responsible for maintaining the firm's shared CLAUDE.md, Skills, and Plugins. Not a side-of-desk task. Without an owner, the layer rots; with one, it gets sharper every quarter.

Pick one AEC integration based on the software your firm uses most. Rhino-heavy: install a Rhino MCP server. Revit-heavy: set up the AEC Data Model MCP. Don't wire everything up at once.

Capture patterns. When you find yourself doing the same prompt three times, save it as a Skill or Plugin. That's the firm context layer growing.

Anthropic Academy

Three courses at anthropic.com/learn for AEC technology leaders:

• AI Fluency: Framework and Foundations. The 4Es and 4Ds in depth. Run through as a leadership team.
• Claude API Development Guide. Prompting, tool use, function calling, agentic systems. For the people who will be building.
• Model Context Protocol (MCP) Fundamentals. Building MCP servers and clients. For whoever maintains your custom integrations.

Closing

The cost of building firm-specific software has collapsed. You own your data and your workflows. Vendors don't have either. The layer on top of your core tools is yours to build, if you want it to be. That's the whole pitch.

Pick one boring workflow. Spend a Saturday on it. Worst case, you lose an afternoon. Best case, you've started the muscle that lets your firm own the tooling layer for the next decade.

The tools that matter to your firm are the ones you start building this weekend.

Future AEC Hub guides go deeper on integration walkthroughs, case studies, Claude Code for AEC, and templates. A live course is in development for fall 2026; join the waitlist.