How I AI: Zach Davis's 3 Workflows for Enterprise Engineering with AI

In this episode of How I AI, LaunchDarkly engineering director Zach Davis shows how he prepares a mature codebase for humans and coding agents, turns noisy test output into a prioritized cleanup queue, and coaches interviewers with a custom GPT.

The workflows reflect different stages of adoption. Zach demonstrated working examples and described practices his team was testing, but he did not claim that every tool or workflow had been rolled out across the entire engineering organization.

Workflow 1: Centralize documentation and agent rules

Zach’s starting principle is that improvements for humans often help language models too. He moved useful material from Confluence, Google Docs, and other repository locations into a central docs directory so engineers and agents could work from the same references.

How the rules system came together

1. Build human-readable docs with Devin: Zach began with Devin’s repository wiki, asked Devin to draft documentation and a structure, then reviewed and corrected the output.
2. Consolidate the guidance with Cursor: He asked Cursor to combine the reviewed docs and existing Cursor rules into a central `.agents/rules` directory.
3. Point tool-specific files to the center: Cursor’s own rule files reference the relevant material in `.agents/rules` instead of duplicating the full guidance.
4. Add a new tool with Augment: While trialing Augment, Zach pointed it at the Cursor rules and central agent rules and asked it to create Augment’s tool-specific file.

1. Refine rules where agents fail: Zach adds concise guidance for recurring mistakes, such as confusing LaunchDarkly feature-flag product operations with feature-flag code in the repository.

Example request for Augment

Create a unified rule file for our AI agents that encapsulates the rules from the cursor rules and the agent's rules files. This file should be easy for both humans and AI agents to understand and use.

The result is a shared source of guidance with thin adapters for each tool. Zach still reviews generated documentation and rules before treating them as repository standards.

Workflow 2: Turn test noise into a tech-debt queue

Zach used coding agents to break one neglected maintenance problem into work that a person or agent could pick up incrementally.

From log file to reviewed changes

1. Capture the problem: Run `yarn test` and pipe the console output to a log file. Zach asked Cursor how to construct the command.
2. Analyze it with Claude Code: Give the large log file to Claude Code and ask it to group warnings, identify the worst offenders, and estimate severity.
3. Create a task file: Store the prioritized checklist in `.agents/migrations` so humans and agents can share the queue.
4. Assign one bounded item: Give the next task to Cursor, Devin, a background agent, or a human owner rather than attempting the entire cleanup at once.

1. Review before merging: Zach had completed and merged an initial subset. The rest of the checklist remained a queue for future reviewed changes.

Example request for Claude Code

Analyze this test log file and create a prioritized task list for reducing test noise.  Categorize issues by type and severity.  Output in markdown checklist format.

The method exposed roughly 1,200 unwanted lines in a test run and surfaced recurring accessibility warnings. Its value is the shared, prioritized queue, not a claim that all of the debt had already been removed.

Workflow 3: Coach interviewers on scorecard quality

Zach built a custom GPT to evaluate the quality of interview scorecards against LaunchDarkly’s hiring rubric. It grades the write-up, not the candidate.

From rubric to coaching message

1. Define the standard: Supply the hiring rubric and clear evaluation criteria.
2. Show contrasting examples: Add strong and weak scorecards plus instructions for detailed feedback and a Slack-ready summary.
3. Paste a scorecard manually: Zach copied an interviewer’s scorecard into the custom GPT during the demo. There was no automated scorecard ingestion.
4. Review the response: The GPT rates the scorecard as excellent, good, fair, or poor, explains strengths and gaps, and drafts a short coaching message that Zach can edit and paste into Slack.

Example prompt for the custom GPT

Evaluate this interview scorecard based on the provided rubric.  Rate the scorecard as Excellent, Good, Fair, or Poor. Provide specific feedback on strengths and areas for improvement.  Also, generate a brief Slack message summarizing this feedback.

Zach found the tool useful for improving his own scorecards while testing it. The manager still decides whether the feedback is accurate and appropriate before sending it.

The operating pattern

Across the three workflows, Zach creates a shared artifact first: repository docs and rules, a prioritized migration checklist, or a hiring rubric. An agent then works against that artifact, and a person reviews the result. That pattern gives a large team room to experiment without treating every generated output as production-ready.