Project Postmortem (Behind the Scenes)

This post is part of an ongoing experiment using AI to help surface patterns hiding in plain sight. I’ve been exploring how to reflect on potential client inquiries and communication habits by reviewing emails sent and received over time.

The audit process revealed some blind spots that helped me improve systems so I thought why not apply the same technique to a difficult project from the past to see if I could learn anything.

I headed into the email archive and selected all email correspondence from a particularly stressful architectural project with hundreds of messages. I’ve anonymized the project details here (I’m calling it "X House") out of respect for the people involved. I printed them as a single PDF and then asked ChatGPT to serve as my ‘business communications consultant’ and help me look for “communication breakdowns, hidden stressors, and decision bottlenecks.”

The goal wasn’t to assign blame; only to learn from the patterns and fix my own blind spots.

Here’s what I learned:

The stress was a result of ambiguity.

The client was hesitant, often unsure, and slow to make decisions. Questions would pile up about cabinet layouts, siding types, energy systems, and trim details. The more they did so, the more the overwhelm set in and clear decisions lagged. When they did come, they often reopened earlier ones (questioning led to re-questioning).

In short:

  • Weeks of silence followed by last-minute asks.

  • Requests for redraws that reopen earlier decisions.

  • Client worry about making the “wrong” choice—without a framework to guide them.

  • Important decisions farmed out to consultants or mystery trusted ‘expert friends’... who didn’t attend meetings but weighed in behind-the-scenes.

There were long stretches of indecision followed by sudden urgency (usually on the weekend when the client had time to focus on the project). It was hard to plan around and taxing as I agreed to ride the same emotional roller coaster as my client.

To be fair, I allowed it. At the time there wasn’t a system in place to say: “this isn’t mine to fix.”

Actual chat excerpt

But there was something else. Reviewing this thread also surfaced how I communicate when under pressure. AI flagged moments where my tone became overly accommodating or vague, where I delayed saying “no,” or softened a boundary that should’ve been firm. Here were some of the communication patterns that created downstream issues (quotes are actual things I wrote, cringe!):

  • Delaying direct replies in hopes the client would self-resolve.
    “No rush on this.”

  • Burying key points inside overly long or overly polite responses.
    “I hope this makes sense, but happy to clarify further or walk through it together.”

  • Offering solutions too quickly, before clarifying whether the issue was mine to solve.
    “I can look into that and get some preliminary numbers if that helps.”

  • Avoiding reminders that a decision was overdue or already made.
    ”Just checking in on this again—totally understand if it’s still in process.”

There were dozens of improvements suggested, here are a few I’ve incorporated:

  • Write shorter emails with bolded action items.

  • Restate decisions when a client circles back. “Just to recap where we landed: we decided on [X] during our last review—let me know if something’s changed on your end.”

  • Respond to scope creep with a simple phrase: “Happy to explore this. Let me check how it impacts scope.”

how to run your own project postmortem audit:

  1. Pick the project. For this example, it’s the one that drained you most. You can also do the same for a great project and use that to refine your acquisition process + client avatar. You want more of those!

  2. Export all communications. Select all → Print → PDF. If you have subfolders: SD, DD, CD, CO, etc. export each separately to give ChatGPT more context. Perhaps everything goes smoothly during SD + DD but falls apart in CD and CO. You can extract those insights with this technique.

  3. Open ChatGPT (or another AI assistant). Use this prompt:

    “Act as a business communication strategist with special expertise in architectural design, construction and documentation. I’ll share all emails from a past project in a single PDF. Help me identify patterns of stress, confusion, or poor decision-making. Be specific and actionable.”

  4. Prompt for deeper insights. Use follow-up prompts like:

    • “Where do you notice tension or confusion building?”

    • “Which replies signal scope creep or boundary slippage?”

    • “What could I have said earlier or more clearly?”

    • “Which of these issues created the most downstream effects? Prioritize them.”

    • “How many times was this decision revisited?”

    • “Was the client ever given a clear, bounded choice?”

    • “Where did I continue designing without key decisions locked?”

  5. Improve your Process. Ask:

    • “What systems would’ve helped here?”

    • “What template or script might I reuse in the future?”

    • “Where did I unintentionally signal I was responsible for something I wasn’t?”

    • “Which messages created more work down the line, even if they seemed fine in the moment?”

    • “What else haven’t I asked that would be important to know?”

  6. Build better Systems. After gathering insights, prompt ChatGPT to help you design new tools:

    “Based on this thread, suggest specific tools, templates, or boundary-setting phrases I should develop. Focus on making communication clearer, reducing decision fatigue, and protecting my time.”

    Then ask: “Which of these patterns are likely to show up in future projects—and how can I systematize my response?”

    Use the feedback to draft:

    • A boundary-setting phrase bank

    • A “pending decisions” tracker shared with clients

    • A visual timeline that flags who decides what—and when

Taking it further

This kind of audit works for any aspect of your life that flows through your inbox. One of the most useful ways I’ve applied it lately is reviewing the last 30 days of sent emails. Export them as a PDF, then ask ChatGPT to look for patterns: Do I delay responses until a problem escalates? Do I soften direct feedback? Overuse certain phrases? Come across as vague—or domineering? Am I unintentionally assuming more risk than I should?

That last one can be particularly costly.

I hope this helps you spot habits you weren’t aware of, and once you do, you can begin re-wiring some of your defaults.

AI in my Architecture Practice - 2025 Update

Over the past few months I've been experimenting with AI and discovered something counter-intuitive: the most impactful AI tools I return to each day aren't design-related at all. It's the ones that eliminate the administrative burden stealing my creative time.

I tracked every hour of my work for a month and the results were sobering:

  • 35% administrative overhead

  • 25% meetings, site visits, coordination

  • 20% client management

  • 20% design and production

That doesn’t match how I see myself—or how most of us do. We imagine we’re in the business of designing. The reality is, we’re in the business of everything else!

While everyone's talking about AI for concept generation, I've found it's far more valuable as an administrative partner to help me reclaim some of that 80% (!!!) of non-design time.

Here are few quick use cases to inspire you:

1. Estimate Taxes + Optimize Paying Them

Tax time is closing in and this quarter, I used Claude 3.5 Sonnet to estimate my quarterly taxes for me, including the QBI deduction (that’s a bummer to calculate accurately, IYKYK!) I fed it my P&L and asked:

“Estimate my quarterly taxes, include QBI deduction, show your math, and document assumptions.”

It gave me a breakdown of AGI, deductions, and payments for Federal and State. Then I asked:

“What’s the best business credit card to use if I want to pay this tax bill and get the highest return, including signup bonuses? Factor in processing fees, annual fees, rewards and bonuses.”

It compared rewards, fees, and point value based on my actual spend. I picked one, applied and when I pay my taxes on April 15th I will have earned at least $1,100 in travel value plus some cash back. Now I have a 'tax project' in ChatGPT I can use moving forward.

BONUS HACK: When your new credit card arrives, don’t toss the Guide to Benefits. It’s a dense, jargon-filled list of every perk the card offers—trip delay insurance, extended warranties, purchase protection, and more. I used to ignore it too. Now, I snap a photo or upload the PDF to a dedicated ChatGPT project (mine’s called Credit Cards). I ask it to extract a summary: annual fees, renewal dates, and key benefits. Then I drop that into a Notion page for quick reference.

Even better, I’ve trained ChatGPT to create context-based reminders. For example, when I’m preparing to book travel, I just prompt:

“Act like my travel concierge. Based on the benefits from my [Card Name] guide, what should I remember to use, activate, or avoid when booking a trip?”

It replies with a checklist—flight protections, unused travel credits, which card to use for coverage, and claim instructions if something goes wrong. It’s actually saved me real money. I once left a new iPad on a plane and got it replaced because ChatGPT reminded me the card I used covered theft and loss.

You can do the same with memberships. AIA, for example, comes with perks that are easy to miss—rental car status, product + insurance discounts, even legal resources. Catalog them once, and your future self will thank you.

Other uses + ideas:

→ A. Forecast project cash flow month-by-month
Feed your live Google Sheet of invoices, payment milestones, and planned expenses into Claude or Gemini and prompt:

“Map expected cash flow by month for the next 6 months. Flag risk periods and suggest payment timing strategies.”
It’ll show you where things might get tight—before it happens.

→ B. Identify underperforming projects
Upload project-level budget actuals and ask:

“Which projects are tracking below projected profitability and why? Sort by fee structure and project type.”
This is insight you can act on—whether to adjust scope, schedule, or staffing.

→ C. Optimize recurring business expenses
Dump your last 6 months of business expenses into Claude and ask:

“Identify subscriptions, tools, or services that I haven’t used in 90+ days. Suggest where I can cut or consolidate.”
This is the digital equivalent of spring cleaning—and it saves real money.

2. Meeting Notes That Write Themselves

Client calls, site visits, coordination meetings—they pile up.

I use Granola.ai (you can also use Whisper transcription) to record all my meetings, then I upload the transcript to GPT-4o and prompt:

“Summarize this by (1) decisions made, (2) action items by stakeholder, and (3) unresolved issues.”

If the result is messy, I iterate:

“Now rewrite this as an email to each stakeholder in the form of a follow-up email—clear, concise, bullet points for each decision.”

The real shift happens when you use these meeting summaries to build a dedicated project in ChatGPT or Notebook LM (powered by Google's Gemini). I've started prompting:

"Compare this meeting's decisions to our last three meetings on this project. Identify any reversals or inconsistencies in client requests."

This catches potential scope creep early and provides documentation when change orders are submitted.

In Granola you can chat with any meeting too. I can't count how many times I've left a meeting only to forget the final decision we agreed upon for a minor project detail just a few hours earlier.

Other uses + ideas:

→ A. Auto-generate discipline-specific punch lists
After a consultant meeting or site visit, upload audio or transcript and prompt:

“Generate three punch lists: one for the GC, one for the electrical engineer, one for my internal team. Include photos if referenced.”
Use tools like SuperWhisper, Granola, or Whisper transcription + ChatGPT to pull this off.

→ B. Track design decision history
Feed in weekly notes and transcripts and ask:

“Track all decisions made about the kitchen design since project start. Include client rationale and status.”
It’s building a searchable memory—one you’ll be grateful for when changes come up.

→ C. Polish Design Narratives for Awards or Clients
LLMs excel at improving clarity and tone over multiple passes. Feed in your initial project description, then iterate:

  • Pass 1: Tighten the language

  • Pass 2: Highlight innovation without jargon

  • Pass 3: Match a specific style guide (e.g., “write this like a Dwell piece”)

The more you guide it, the more precise the result.

→ D. Throw your project meeting notes into Notebook LM
Take your project meeting minutes, site visit notes, and any related PDFs or emails and upload them to Notebook LM. Once the material’s in, prompt:

“Summarize these notes into a 6-minute podcast I can listen to on my way to the site. Prioritize unresolved issues, decisions made, and any changes since the last visit.”

Notebook LM will generate a natural-language conversational summary and you’ve got a quick, customized pre-visit briefing—no more scrambling to remember what happened last time.

3. scaling your expertise

I've started documenting any specialized workflows—the things only I know how to do—to make them transferable. For example, I recorded myself on Loom as I completed a zoning analysis for a new residential project. I took the transcript that it auto-generates and dropped it into Claude and asked:

"Transform this into a detailed SOP that a junior designer could follow. Identify steps where additional context is needed. Flag opportunities for automation and suggest ways to simplify or speed it up."

Once I had that draft, I asked GPT-4o to review it:

“Improve clarity, remove redundancies, and make sure a junior staff member could follow this without needing context.”

Now I have a reusable, self-contained + more efficient process I can easily share with a remote team member.Other uses + ideas:

→ A. Document your social media workflow
Record yourself outlining how you take a finished video or project photo and schedule it for Instagram or YouTube. Then ask:

“Turn this into a repeatable checklist with tool recommendations and time estimates. Make it usable for a VA.”
Now you’re delegating marketing without starting from scratch.

→ B. Formalize your proposal writing approach
Document how you create fee + scoping proposals. Feed into Claude and ask:

“Write an SOP for proposal development that captures performance goals, cost review, and documentation.”
You’ll spot steps that can be templatized or automated.

→ C. Create onboarding for new collaborators
Walk through your project folder structure, naming conventions, and communications process. Then:

“Create a ‘welcome doc’ for new freelancers or consultants that explains how to work with me effectively.”
You only write this once. It pays off every time you onboard someone new.

The most powerful addition I've recently implemented is a weekly "meta-review" where I ask the LLM to analyze its own outputs: "Review all AI-assisted tasks from this week. Identify patterns in what I'm asking for and suggest workflows that would anticipate these needs." This recursive learning loop has identified several repetitive tasks I wasn't even aware I was doing regularly.

Start small this week: pick the administrative task you find most draining and apply one of these approaches and experiment from there.

None of this is theoretical. These are small, compounding wins. Over time, they’re forming the backbone of something bigger—until I can think of a better name, I’ve been calling it, Project Intelligence. It’s a system that evolves alongside each project, remembers everything, and handles more of the overhead with every iteration. If you’re unfamiliar with the current AI tools, now’s a great time to begin experimenting.

Model Selection MATTERs:

Not all AI models perform the same. Each one has strengths depending on what you're trying to do. Here's how I choose:

GPT-4o (ChatGPT Pro)

  • Best for: General writing, email drafts, meeting summaries, everyday prompts

  • Why: Fast, reliable, well-rounded. Good at tone and formatting.

  • Use when: You want clarity fast, or you’re iterating on written content.

Claude 3.5 Sonnet / Opus (Anthropic)

  • Best for: Complex reasoning, SOPs, long context (e.g. project docs, transcripts)

  • Why: Structured, cautious, and good at step-by-step logic.

  • Use when: You need it to “think” clearly across multiple steps or large chunks of info.

Gemini Advanced (Google)

  • Best for: Financial modeling, spreadsheet logic, parsing PDFs from Drive

  • Why: Great with structured data and math-heavy prompts.

  • Use when: You’re projecting cash flow, comparing options, or dealing with numbers.

Perplexity Pro

  • Best for: Research, source-backed summaries, trend scans

  • Why: Cites everything. Good for learning something fast.

  • Use when: You’re researching materials, methods, or unfamiliar territory.

Grok (xAI / Twitter)

  • Best for: Real-time sentiment, trends, social-adjacent ideas

  • Why: Built on social data. Can be surprisingly creative.

  • Use when: You want to know what people are saying right now—or just explore ideas.

Bottom Line:

  • Use Claude for structure + logic

  • Use GPT-4o for clarity + versatility

  • Use Gemini for numbers + documents

  • Use Perplexity for citations + research

  • Ignore Grok unless you're very online

Using AI as Design Tool in My Architecture Practice

I’ve been practicing architecture for almost thirty years now and although my design process has evolved over that time, it’s one I’m wholly comfortable with. It’s efficient, it’s safe, and I know following it will yield good work. I’ve noticed too that it can produce similar-looking results as I return to familiar forms, plan layouts and details that have worked in the past. And this is a perfect use case for generative AI design tools - like Midjourney - to help break the frame and help me quickly explore options I may not have previously considered.

“Copying others is a good way to start a career. Copying yourself is a disappointing way to end one.” - Kevin Kelly

Using Midjourney, is strangely addictive. You enter a prompt and watch as four hyper-realistic images ‘hatch’ before you, resolving like an old Polaroid photo as the algorithm compiles the pixels. It feels a lot like gambling. Each prompt promises a new chance to get lucky. The results are captivating images of seemingly perfect buildings and it’s easy to mistake them for finished work which they’re not. Accepting these first results would deny the true power of Midjourney as a tool for ideation. Remixing, upscaling and blending images recaptures the autonomy of the designer to shape the outcome and preserves the capabilities of Midjourney to quickly deliver many iterations of the same idea.

Watch the video to see how I’ve been experimenting with it in my practice.

Heading to MidJourney without an idea or a concept guiding your efforts means you’ll likely spend a lot of time creating seductive images of buildings that have no place in a design process. Images without substance. For me, the design process starts with a sketch and an idea. I head to Midjourney to solve a problem I had before heading to Midjourney which is quite different than using it to cycle through an averaging of the imagery that the algorithm has ingested.

Midjourney is a text-to-image generative AI app that uses Discord as the prompt interface and creates a four panel image spread based on a text prompt. Entering something simple like: “iron lace by Tadao Ando ––ar 16:9” produces this image:

A short prompt like this allows Midjourney the freedom to be inventive. It recognizes Ando is an architect and produces a building with his characteristic elemental form language. Importantly, the algorithm has ingested many photographs of his work which tend to be dramatic, sweeping one or two point perspective views with water features and concrete and you’re receiving an average slice of the hybrid of those and the ‘iron lace’ images it’s been trained on. The ––ar 16:9 is an aspect ratio modifier that delivers a wide angle view.

To wrest more control from the AI, you may be inclined to create longer, more precise prompts which, in my experience, can produce mixed results. The recipe below balances precision with room for Midjourney to be creative.

Prompt Recipe:

Once you get beyond the clunky Discord interface used to create the images. The key to unlocking Midjourney as a creative assistant lies in controlling the output through specific prompt language. Words closer to the beginning of the prompt will have more weight in the final result and you’ll want a basic understanding of the parameters before getting too far, Midjourney keeps an updated parameter list here. Now, on to the prompt recipe:

Trying to recreate someone else’s visual style in Midjourney is less interesting to me than the process of experimenting with the tool to produce the kind of image that’s useful to my design process. For example, I’m more interested in creating building forms that don’t look anything like their antecedents. I often force Midjourney to produce buildings without common attributes, like windows or doors, by using the modifier “no” (––no windows ––no doors). I find the results more interesting and useful to me.

Aberrations

What I love about using generative AI for ideation is that it introduces error and randomness into an intentional, too-highly controlled process. It’s similar to analog processes like model-making or sketching, where an error presents an unexpected opportunity. New ideas often come from noticing flaws, where something is strange, misplaced, or missing altogether. Exploring these aberrations can push the design forward. The circled corner of the abstract barn image I created suggested I might consider eroding a heavier base and tucking the work program below the living program vertically separating it rather than the initial idea to separate it horizontally.

Finding a secret sauce of your own making will take some experimentation, just remember that AI is a mirror for your own creativity. Overly constrained thinking will yield similar results. Plying it with interesting prompts it’s become an indispensable office intern who has helped to reinvigorate my design process with fresh ideas and new perspectives.