What 223 Sessions Taught Me About Working with AI
I pulled 30 days of Claude Code session data — 223 sessions, 38,000+ prompts, 32 projects, 1.2 GB of conversation. The numbers tell a story about what AI-native development actually looks like when nobody's watching.
Nino Chavez
Product Architect at commerce.com
I’ve been building with Claude Code for months now. Shipping real products, writing client proposals, managing a photography gallery, coaching volleyball — all of it running through the same CLI tool on the same laptop. But I’d never actually measured what that looked like from the outside.
So I pulled the data.
Not from a dashboard or an analytics product. From the raw session files — the .jsonl logs that Claude Code writes to disk every time you start a conversation. I wrote a script, pointed it at the last 30 days, and let the numbers tell me what I’d actually been doing.
The Numbers
30 days. 223 sessions. 38,322 prompts. 1.2 GB of conversation data.
Across 32 distinct projects. Some days I ran 19 sessions. Some days, one session ran over 2,300 prompts.
Here’s the daily breakdown:
| Date | Sessions | Prompts |
|---|---|---|
| Feb 22 | 19 | 2,299 |
| Feb 24 | 11 | 2,341 |
| Feb 25 | 6 | 2,194 |
| Mar 1 | 16 | 2,404 |
| Mar 2 | 9 | 2,268 |
| Mar 5 | 13 | 2,156 |
| Mar 9 | 14 | 2,118 |
| Mar 12 | 4 | 2,180 |
| Mar 15 | 1 | 2,324 |
That last one — 2,324 prompts in a single session. One project, one conversation, one unbroken arc of work.
The average session was 172 prompts. But the distribution tells the real story.
How Sessions Actually Break Down
| Length | Prompts | Sessions | Share |
|---|---|---|---|
| Quick | 1-10 | 14 | 6% |
| Medium | 11-50 | 47 | 21% |
| Deep | 51-150 | 93 | 42% |
| Marathon | 151-500 | 54 | 24% |
| Ultra | 501+ | 14 | 6% |
The bulk of the work — 42% — happens in sessions between 51 and 150 prompts. Not quick fire-and-forget interactions. Not day-long marathons. Something in between: focused work arcs that last long enough to build real momentum.
But the marathons and ultras account for 30% of sessions and a disproportionate share of actual output. Those are the sessions where an entire feature gets built, a client proposal gets written and revised, or a new product goes from brainstorm to deployed prototype.
The most productive sessions aren’t the longest ones. They’re the ones where the context never breaks.
What the Work Actually Was
This is the part that surprised me. I expected “coding” to dominate. It didn’t. Not even close.
| Category | Sessions | Prompts | Share |
|---|---|---|---|
| Product Development (AI analytics) | 46 | 12,056 | 30% |
| Cross-cutting / meta-work | 55 | 8,039 | 21% |
| Product Development (sports tech) | 28 | 6,154 | 16% |
| Enterprise (employer) | 15 | 3,552 | 9% |
| Product Development (photography) | 28 | 3,330 | 9% |
| Client work | 11 | 1,879 | 5% |
| Content & thought leadership | 25 | 1,459 | 4% |
| Internal tooling | 7 | 1,113 | 3% |
| Brand & marketing | 7 | 613 | 2% |
| R&D / greenfield | 1 | 548 | 1% |
The biggest category — 30% of all prompts — was a single product: an AI analytics dashboard I’m building. That tracks. It’s the most complex project in the workspace and the one in the heaviest development phase.
But look at the second line. “Cross-cutting / meta-work” — 21% of all prompts — is the work that doesn’t belong to any single project. Brainstorming new ideas. Auditing the workspace. Organizing files across projects. Building shared tools.
This is the work that doesn’t show up in a sprint board or a git log. It’s connective tissue — the thinking that links one project’s lessons to another’s requirements.
The Session-per-Project Pattern
I don’t use Claude the way most people seem to describe using AI tools.
I don’t ask it quick questions. I don’t use it as a search engine with attitude. Instead, I spin up a conversation inside a specific project directory and ride it for hours. Sometimes days, resuming the same session across multiple sittings.
Each project has its own CLAUDE.md file — a persistent instruction set that tells Claude who it’s working with, what the project conventions are, what to avoid, and where to find things. When I start a session inside that project, Claude already knows the stack, the design system, the deployment target, the voice, the file structure.
That means I don’t spend the first 15 minutes re-explaining context. The session starts at full velocity.
Here’s what 30 days of that looks like across projects:
- QuantifAI (AI analytics platform): 47 sessions, from red team security audits to landing page copy to domain name brainstorming to user invitation flows
- Rally HQ (tournament platform): 19 sessions, from Swiss format implementation to organizer UX audit to beta user onboarding documentation
- Photography gallery: 27 sessions, including a full migration of 20,000+ photos from SmugMug to Cloudflare Images
- Client sites: 11 sessions across four clients — site redesigns, MVP launches, DNS migrations, SEO preservation
- Signal Dispatch (this blog): 17 sessions for writing, editing, and publishing
None of those are “just coding.” Every project blends strategy, design, writing, debugging, and deployment. The AI handles all of it in the same conversation.
What I’m Actually Typing
I looked at the most common ways sessions start:
| Opener | Count |
|---|---|
| ”Implement the following plan:“ | 47 |
| (Resumed/interrupted session) | 49 |
| Brainstorm / consider an idea | 15 |
| ”Pull latest” | 11 |
| Fix / address an issue | 4 |
| New project kickoff | 3 |
| Audit / review request | 3 |
Almost a quarter of sessions start with “Implement the following plan:” — meaning I’ve already done the thinking and I’m handing off execution. Another 22% are resumed sessions — picking up where a previous conversation left off.
That ratio feels important. Most of my interaction with AI isn’t ideation or exploration. It’s delegation with context. I show up with a plan, hand it to the agent, and monitor execution.
The brainstorming sessions (15) are real too — product ideas, competitive analysis, “consider this conversation from Gemini and tell me what you think.” But they’re the minority. The dominant mode is directed execution.
The Cross-Pollination Effect
One pattern I didn’t expect: I frequently paste output from other AI tools — Gemini deep research reports, Slack conversations, screenshots, even other Claude sessions — directly into a Claude Code conversation as input.
This happens constantly:
- Pasting a Gemini exchange about e-commerce protocols into a blog drafting session
- Dropping a Slack message from a colleague into a project planning session
- Screenshotting a deployed site and asking Claude to audit the layout
- Copying a client’s feedback email into a design revision session
Claude Code isn’t a silo. It’s the execution layer for ideas that originate everywhere. The thinking might start in a Gemini conversation, a Slack thread, a competitor’s website, or a phone call. But it converges in a Claude Code session where it gets turned into artifacts — code, docs, proposals, deployments.
The Infrastructure That Makes It Work
The raw numbers don’t explain why this works. The infrastructure does.
CLAUDE.md files — every project has one. It’s the persistent context layer. Conventions, stack details, deployment targets, off-limits files, preferred approaches. Claude reads this on session start and operates within those boundaries without being told.
Memory system — a file-based memory that persists across sessions. Who the user is, past decisions, project state, feedback corrections. This is how Claude knows that I prefer dark mode, that 630 Volleyball uses Supabase, that I don’t want emoji in my output.
Custom skills and hooks — automated behaviors triggered by specific actions. When I edit a .tsx file, a React best-practices checker runs. When I start a dev server, a browser verification agent launches. When I write a blog post, the voice guide loads automatically.
Plans and tasks — structured thinking for non-trivial work. Before a complex implementation, Claude drafts a plan. I approve, adjust, or reject it. Then it executes against that plan, tracking progress through discrete tasks.
This isn’t a chat interface with autocomplete. It’s a development environment where the AI has persistent state, operational constraints, and enough context to make judgment calls.
What the Data Doesn’t Show
The numbers miss some things.
They don’t capture the sessions where I started, hit a wall, and abandoned the approach. The 49 “resumed/interrupted” sessions represent just the ones I came back to. There are more I didn’t.
They don’t show the quality distribution. Some of those 38,000 prompts were “commit and push.” Others were “audit this entire site for cognitive overhead and rate how self-serve it is.” The prompts aren’t equal.
And they definitely don’t show the human judgment layer — the decisions I made between sessions about which project to work on next, when to stop iterating and ship, whether to accept Claude’s approach or redirect. That orchestration is invisible in the data but it’s the whole game.
What I’d Tell Someone Starting
The question I keep getting — from colleagues, from LinkedIn connections, from friends who see me ship at this pace — is some version of: how do you actually work with this thing?
The honest answer is: I stopped treating it like a tool and started treating it like a team member who needs onboarding. The CLAUDE.md file is the onboarding doc. The memory system is the tribal knowledge. The skills and hooks are the standard operating procedures. The project-scoped conversations are the stand-ups.
None of that is complicated. It’s just… deliberate. The same thing that makes any team effective: clear context, consistent conventions, and enough trust to delegate without micromanaging.
The 223 sessions and 38,000 prompts aren’t the impressive part. The impressive part is that most of them didn’t require me to explain the same thing twice.