Dogfooded my own analytics tool for 31 days and here's what 2,277 Claude Code prompts looked like
I've been building toqn for about a month. It's a behaviour analytics tool for developers who use coding agents like Claude Code, Cursor, and Codex — specifically, it tries to answer one question that none of the existing tools answered for me: which of your AI prompts actually turn into committed code, and which ones don't?
I'm on the Claude Max plan, so I wasn't getting metered bills, but I could feel the token volume growing and had no idea whether any of it was productive. So I did the obvious thing: I tracked myself building toqn with toqn.
31 days. 2,277 prompts. 6.3B tokens. ~$6,142 equivalent spend (unmetered on Max). 130 commits. 6 PRs.
Five things I didn't know about my own coding habits until I measured them:
1. Only 5.3% of my prompts shipped code.
121 of 2,277 prompts produced a commit or PR. The rest were research, planning, debugging, or the wrong tool for the job. I wouldn't call it "waste" — exploration is legitimate work — but before toqn I would have guessed 20-30%. Seeing the real number recalibrated my sense of what an agent session actually looks like.
2. 99.9% of my spend went to a single model.
I built toqn partly because I wanted to compare Claude Code vs Codex for my own work. And but mostly use Claude Code for now. I just love Opus 4.6
3. Friday ships 2.7× more than Sunday, for 1/10th the cost.
Friday afternoons: 32 commits, $245 equivalent spend. Sunday evenings: 12 commits, $2,659. Sunday me is doing exploration and architecture spikes that burn tokens and ship nothing. Friday me is executing. I don't want to stop the Sunday work — that's where the ideas come from but it's useful to know which mode I'm in.
4. Process skills shipped. Brainstorming didn't.
Of the skills I used (superpowers plugin: brainstorming, writing-plans, TDD, subagent-driven-development, etc.), the top shippers were the process ones: subagent-driven-development 77.8% ship rate, TDD 60%, writing-plans 38.5%. My most-used skill (brainstorming) shipped 11% of the time. Brainstorming isn't supposed to ship, but I couldn't have quantified that before.
5. One $2,521 day, zero commits.
A single day in March where I burned 2.7B tokens prototyping an agent architecture I later abandoned. Without toqn, I would have called it "a busy day". With it, I can see it clearly — and I'd make different decisions next time.
---
What I'm taking away
The measurement itself changed how I work. I catch myself mid-session asking "am I shipping or just exploring?" — and either answer is fine, but knowing the answer is new. I didn't expect the tool to change my behaviour. I expected it to just give me a dashboard.
What's next
The dataset is still n=1. I want to see whether the Friday/Sunday pattern is universal, whether other people also default 99.9% to a single model, and which skills actually ship code for devs who aren't me. If you want to see your own numbers: toqn.dev. Free, works with Claude Code, Cursor, Codex, and Copilot via a hook, profiles are private to the owner, no paywall.
What I'd love feedback on
- Which of these five findings would you most want to see in a tool for your own workflow?
- Am I missing an obvious angle? (Cost per feature shipped? PR review time? Which agents unblocked you the fastest?)
- If you're already a multi-model user, what would you want the comparison view to look like?
For the multi-model comparison view, I’d separate post-hoc analytics from “what am I burning right now?” A lightweight live layer for current token/cost pace and reset timing changes behavior mid-session, not just after the bill. That’s the macOS menu-bar layer we built TokenBar for: https://tokenbar.site/
Really smart idea! I'm always wondering, when I move from one coding agent to another, if the new one is really better, or if it's just an impression/I'm giving it easier tasks. I feel like this could solve this issue.
yes try one for a while and the other for a while and the difference will be revealed clearly
This is fascinating data — the 5.3% ship rate really puts things in perspective. I've been using Claude Code heavily too and my gut says my ratio is similar. The Friday vs Sunday pattern resonates hard. I notice I do my best "execution mode" work early in the week and drift into exploration by Thursday/Friday.
One angle you might find interesting: tracking which reusable skills and prompts consistently lead to shipped code vs one-off exploration. I've been collecting community-shared Claude Code skills and the pattern I see is that structured, task-specific skills (like commit reviewers or migration validators) have way higher ship rates than open-ended ones. Would love to see toqn break down ship rates by skill type.
Thanks!
toqn has that already! You can see how often a SKILL is invoked and then toqn calculates SKILL vs commit correlation periodically. You can get this as an insight periodically, though there is no real time calculation and maybe specific chart for this.
Do you think it would be better to see this as a chart, table or something like that?
Took a look at your website: wanted to note 2 things, hopefully they are helpful!
1) Your design is awesome. Honestly one of the cooler landing pages I've seen, love how the green follows the mouse around the screen! You did a great job branding, I would be curious to learn how you did that if you're open to it.
2) As someone that is a full-on vibe coder, the language that you use on your landing page is a little hard for me to understand. Reading your post above, it was written in more "plain" language, which was much easier for me to understand exactly what you do.
Hope that is helpful, seems like a great idea and there will definitely be demand with so many new apps being released every day! Good luck.
Thanks a lot for feedback. Really appreciate it!
1) This I guess comes by experience, and really caring about it. Obviously all of the code is generated but design language, colours, branding and wording is all me refining so many times. Actually this gives me idea of breaking down BE vs FE work, how many of prompts went into FE and how many BE and mix of both. Hmm... I am you sure if you point your agent to toqn.dev it will reveal most of the client side code. As for background it is https://en.wikipedia.org/wiki/Conway%27s_Game_of_Life
2) Hmm... this is a good feedback. I am not sure how can I do better, any specific suggestions?
Very helpful, thank you! I've personally spent a lot of time improving the design within the actual app, and need to create an improved landing page next.
Maybe I'm an outlier, but I just felt like the wording in your post made it easier to understand what toqn does than the website did. Worth considering that there will probably be a lot of vibe coders like me in your target market, who may not have the same technical knowledge as you. Any way you can "dumb-down" the language I think is helpful.
That's just my two cents, hope it helps!
thank you!!!