Working Surfacelink · 11 Nov 2026
03
A player piano running a punched paper roll that a figure seen from behind has just finished cutting by hand at a bench beside it, one hand still resting on the rack that chooses which roll plays next, the whole scene drawn with wide empty margins on every side.
LITALICO Engineers (Zenn)

How test coverage was raised from 23% to 84% with AI

Eguchi (eggc) · 10 November 2026

How-to · artifacts · agents · Agent context files

Read the original
Takeaway

A LITALICO engineer raised test coverage from 23% to 84% by writing a personal testing routine into AI agent skills, while keeping the trade-off decisions for people.

Summary

Eguchi, an engineer on LITALICO's Nursing Net development team, published an account in Japanese of an effort to raise automated test coverage over a year and a half. LITALICO is a Japanese company that runs software and services in education, disability welfare and elder care. Hand-written tests and a new continuous-integration setup lifted coverage to 42%, and AI agents writing tests from shared skills took it to 84%.

Key points
  • Eguchi wrote the steps for writing a test by hand, plus traps specific to the product, as prompts. These became skills that both GitHub Copilot and Claude Code read.
  • Early prompts led the AI agent to overuse mocks to make tests pass, so the skills now limit mocks and forbid direct tests of private methods.
  • Given one file at a time from a list, the agent wrote 50 to 100 test files per pull request; giving it ten files at once lowered quality.
  • Some generated tests were worthless, such as tests that confirm a bug's server error, so Eguchi searches the output for such patterns and checks a sample of files by hand.
Implication

A team with low test coverage could start by writing its most careful tester's routine and trade-offs into a shared skill. Manual testing can stay in place while the agent fills gaps. This is an inference and is not the author's claim, and the account is one engineer's experience on one product.

Plays
Engineering

Write the steps an experienced engineer follows to write a test, and the team's testing trade-offs, into a skill the AI agent reads before writing tests.

Engineering

Search AI-written tests for known worthless patterns, such as tests that expect a server error, and review a sample of files by hand.

Source issue

11 November 2026: as AI agents write most of the code, teams are writing their engineers' judgement into shared skills, review rules and testing policies that the agents read every time.