Senden-Kaigi specialized copywriting assistance AI

Concept

In the realms of logic and calculation, AI has been surpassing humans one field at a time. Yet “crafting words that move people” — the domain of creative writing — has long been considered territory that belongs to humans alone.

Is that really true? We wanted to find out for ourselves. We chose the Sendenkaigi Award as our proving ground. Entries are published, briefs are explicit, and evaluation criteria are clear. A third-party jury could tell us whether AI-assisted copywriting could reach a standard that holds up in the real world.

That said, we didn’t go in the direction of having AI write everything. When we lined up large volumes of output, we hit a wall: “AI can produce 80-point copy in bulk, but it can’t produce 90-point copy.” That led us to redesign the approach. Generate at scale with AI; let humans curate from that output. Every line we submitted was chosen by human hand.

How It Works

1. Training the AI on a “60-Category Taxonomy of Copy”

We have someone in-house who once won the Sendenkaigi Award Grand Prix. Over years of studying past award-winning copy and distilling their thinking, they built a “60-Category Taxonomy of Award-Winning Copy” — a map of angles, organized around concepts like “discovery of essential value,” “discovery of emotion,” and “discovery of expression.” We trained the AI on this taxonomy.

As a result, the AI no longer imitates surface-level language. Instead, it operates from a higher-order question: “Which angle should I take against this brief?“

2. Exhaustive Generation via the Gemini Batch API

Against 31 briefs, we combined over 60 techniques and ran them through the Gemini 2.5 Pro Batch API to generate copy at scale — up to 1,900 lines per brief.

Batch jobs were managed through a CLI that supports pause and resume; all results were saved as JSONL. Generation and internal scoring happened simultaneously.

3. Training a Scoring AI from Scratch

The heart of the project. Rather than relying on prompts, we built an actual machine learning model in-house.

• We created a judgment UI that displays two candidates side by side (A/B)
• A copywriter with award experience made intuitive calls — “this one’s better” — one after another
• 400 A/B judgments became the training data for the scoring model
• On the validation set: ROC-AUC 0.81, binary accuracy approx. 74%

A real comparison example:

A: "Try deep-frying anything your kid refuses to eat."   → Score 3.57
B: "It's nice when it doesn't feel greasy."             → Score 0.60
AI judgment: A wins with 95.1% probability

This allowed us to cut 1,900 lines down to 300.

4. A Web App Built for Human Selection

The 300 lines the scoring AI passed were then read one by one by human reviewers, who finalized the 100 submissions. We built a dedicated internal web app in Convex + React.

• Per-reviewer evaluation queues (unreviewed copy prioritized)
• Real-time tallies of likes / passes / total evaluations
• Final selection of 100 from 300, resolved entirely within the database

The scoring AI is useful for “eliminating the obvious failures,” but by its nature it cannot “identify the 90-point line.” The final call of taste stays with the human.

5. A Final Polish with Gemini

For the 100 selected lines, we applied a last pass with Gemini — nudging it with prompts like “make this punchier” or “try this angle” — to sharpen the final wording. Rewrites by AI; final approval by human.

1. 01 INPUT Load briefs and angles 31 briefs and the "60-Category Taxonomy of Award-Winning Copy" — systematized by a Grand Prix winner — are fed in as input.
2. 02 GENERATE Exhaustive generation with Gemini 2.5 Pro Batch API generates copy at scale — up to 1,900 lines per brief. 1,900lines / brief
3. 03 SCORE Cut-off by the scoring AI A machine learning model trained on 400 A/B judgments (ROC-AUC 0.81) eliminates the obvious failures. 1,900→300lines
4. 04 PICK ━ HUMAN Hand-picked by humans Using the evaluation UI built in Convex + React, reviewers read all 300 lines one by one and confirm the 100 submissions. 300→100lines
5. 05 POLISH A final pass with Gemini Nudging with prompts like "make this punchier" or "try this angle," we refine the wording of the 100 submissions. Rewrites by AI; final approval by human.

Results & Impact

• 38 first-round passes / pass rate 1.44% (overall average 0.98% — approximately 1.5× the industry average)
• 1 second-round pass (surviving among 500 out of approximately 560,000 total entries)
• Scoring AI: ROC-AUC 0.81 / binary accuracy approx. 74%

Objective

• Operate AI not as a “vending machine for answers” but as a partner that draws out judgment criteria
• Turn the loop itself — mass generation → scoring AI → human selection — into an in-house asset
• Roll out the same pipeline to naming, concept development, and proposal writing

Insights

We started with an experiment of “leaving everything to AI.” The speed — hundreds of lines in an hour — was overwhelming. But reading them side by side, we kept hitting the same wall: “AI can produce 80-point copy in bulk, but it can’t produce 90-point copy.”

Late in the project, a line slipped out internally — “We couldn’t think of any way to dramatically raise performance beyond manual prompt tuning.” This was not a technical defeat. It was a conclusion: the final ten points of copy live only in human sensibility.

What we wanted to leave behind through this project was not the submissions, nor the pass rate — it was the writer who sharpened their judgment by running alongside AI. Getting to “AI is fast” is something anyone can reach now. When you go after the 90-point line beyond that, the final call belongs to human instinct. Generate with AI, select with humans. The more you run that loop, the sharper the selector’s eye becomes.