Five Days of Working Alongside AI

This is a record of the work I handled together with AI from Monday to Friday this week. I mainly used Codex (Pro), with Claude (Pro) playing a supporting role.

In case anyone is curious about what it's actually like to work alongside AI, I wrote down a log of these five days. New technologies like Hermes and OpenClaw are everywhere these days, but for real-world work, this is about the level I can realistically apply.

First, I'd been wondering how much of an efficiency gain I'd actually get from working with AI — and, unsurprisingly, so far it seems pretty solid. To be clear once more: this wasn't a matter of just running the AI and taking whatever it produced. It was a process of running it, reviewing the output, running it again, and checking again — repeated over and over to get things ready for a live service. This wasn't something I ran just as a test. It's a record of actually using AI for real service work, from product features to promotion.

Daily Usage

Note: The total token count includes cached input tokens. The figure closest to actual new processing — based on non-cached input plus output tokens — comes to roughly 36.15M tokens.

Work Handled with Codex

Over the past five days, I ran operations, admin features, content management, analytics screens, and marketing prep across several solutions in parallel.

The scope went well beyond simple code fixes and included the following:

Solution A — Admin analytics screen improvements

• Cleaned up the visitor-by-country, referral, device, and visit-summary screens
• Fixed referral-classification errors
• Standardized the display criteria for the dashboard and visit-history screens

Solution B — Content management system improvements

• Improved the post-deletion feature
• Added bulk-select deletion, a custom confirmation dialog, and cleanup logic for related images/media
• Diagnosed and fixed the root cause of database query errors
• Tidied up admin UI fonts and screen layout

Solution C — Public content page improvements

• Fixed an issue where product info displayed differently from the admin data
• Resolved image-cropping problems
• Standardized image optimization and display criteria
• Improved the credibility and visual polish of the product pages

Solution D — Magazine/article screen improvements

• Improved the structure of the magazine detail view
• Adjusted the table-of-contents UI, header area, title sizing, and spacing
• Kept the article view similar to the magazine layout while giving it its own distinct character

Company/brand site cleanup

• Improved the footer layout
• Connected SNS links
• Fixed an issue where translation keys were exposed
• Removed unnecessary copy
• Cleaned up multilingual translation data
• Verified the actual screens after deployment

Marketing/SNS direction

• Reviewed the limits of repetitive image/copy posting
• Reorganized the SNS strategy per product
• Built a content strategy grounded in the actual product screens
• Tidied up SNS links, meta tags, typos, and profile info

Work Handled with Claude

A · B SNS promotion execution

• X: cleaned up account-browser mapping, wrote follow strategies and posts for A and B individually
• Facebook: set up both A and B pages, joined relevant groups and posted
• Instagram · Threads: posting and account cleanup

Recording / routine-building

• Consolidated all channel work into a single unified log (sns-promotion-log.md) — a status table, per-channel details, and a TODO list
• Wrote a new SNS routine checklist — fixing the order "respond to reactions first → post new content later" for every channel

Strategy / environment setup

• Laid out A's funnel-instrumentation direction (why measuring drop-off stages comes first)
• Set up scheduling for recurring tasks
• Configured Claude plugins/connectors to match each role

AI's Analysis

A workload of this size would typically take a week or two to handle alone by hand. Because it also included admin features, databases, image optimization, multilingual support, deployment, and live-service screen verification, the amount handled in five days is far from small.

That said, not every step went perfectly smoothly. There were cases where the AI misjudged a root cause, treated an image problem as a CSS-only issue, confused the deploy branch with the actual deployment state, or needed re-checking due to encoding/local-environment problems. As a result, a significant portion of the work went into verifying and correcting each item rather than trusting the AI's output as-is.

So the productivity assessment breaks down as follows:

• Workload: High
• Processing speed: Fast
• Verification cost: High
• Inefficiency: Some
• Overall productivity: High
• Perceived fatigue: Likely high

In conclusion, these five days weren't wasted time — they can be seen as a period where I used AI to push a large amount of work through quickly, while spending heavily on verification to protect operational quality.

Speed clearly improved thanks to AI, but for work that goes into a live service, it was still essential for a human to set the direction, check the results, and verify things before and after deployment. In other words, AI dramatically boosted the working speed, but what guaranteed the final quality was human review and judgment.