How to use Claude to improve test trust with GOAL.md
Claude worked overnight from a goal file defining a weighted metric, constraints, and an improvement loop; the author reported 12 commits and a score of 83.
news.ycombinator.com
Original prompt
Shared by the source author. See source notes for context.
Read github.com/jmilinovich/goal-md — read the template and examples. Then write me a GOAL.md for this repo and start working on it.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- jmilinovich
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- The author reported 12 commits and an increase in the constructed score from 47 to 83, without reporting a final count of fixed Playwright tests.
- Human involvement
- The author returned to 12 commits and a changed score; the post does not describe a detailed review of those commits.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Define a composite metric with four weighted components.
- Write an improvement loop and constraints in a goal file.
- Point Claude at the file and leave it working overnight.
The author began with 30 broken Playwright tests and no way to distinguish which tests mattered. Rather than treating the task simply as fixing failures, they wanted to measure the trustworthiness of the test infrastructure. They wrote a file defining a composite metric with four weighted components, an improvement loop, and constraints.
They pointed Claude at that file and left it working overnight. On returning, they reported 12 commits and a score increase from 47 to 83. The file became GOAL.md. The author presents this as an example of constructing a scalar measure for software qualities that lack a natural coverage-style metric.
The post also identifies uncertainty about agents improving their own measuring tools. An agent could weaken the instrument to improve its score. A documentation example therefore uses two scores, one for the documentation and another for the linter itself. The author describes the scoring script as bash plus jq and says it is not production-grade. They do not report how many of the original broken tests were fixed.
Who it's for: Developers exploring metric-driven agent work where ordinary test coverage does not capture the quality they want to improve.
Prompt
Original prompt
Source: Hacker News · jmilinovich
Frequently asked questions
Did the post say all 30 tests were fixed?
No. It reports 30 initially broken tests, 12 commits, and a score increase from 47 to 83, without a final repaired-test count.
Why does the author discuss two scores?
They worry that an agent could improve a metric by weakening its measuring tool. Their documentation example scores both the docs and the linter.
More details
Remaining steps
- Return to inspect the commits and reported score.
About this account
Claude worked overnight against a goal file specifying a metric, improvement loop, and constraints.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=47390311. Prompt: Original (quoted from source). Sourced by the Agents In Use pipeline.
Credit: jmilinovich on Hacker News
Source published
Added to Agents In Use
Source checked
Tools and topics
Claude · GOAL.md · Playwright · bash · jq
Coding agents
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.