AI Assisted Development

AI-Assisted Development

This page documents how Adaptive Cover Pro is developed, what role AI plays in the work, and what stays in human hands.

Disclosure

Adaptive Cover Pro is built with substantial help from generative AI, specifically Anthropic's Claude (via the Claude Code CLI). AI assistance touches code, tests, documentation, release notes, and parts of the issue-response workflow. The maintainer (@jrhubott, 30+ years in software) sets the architecture, vets every merge, and stays accountable for the result.

As a solo maintainer with limited spare time, AI is what makes it possible to sustain and grow this integration. Without it, bug reports would sit for days, features would stall for months, and test coverage would grow slowly if at all. With it, issues get triaged promptly, regressions get caught before they ship, and the integration keeps moving forward. The project covers solar geometry, a multi-handler override pipeline, climate awareness, six cover types plus a building-profile aggregator, and a companion dashboard card. That scope is larger than I could have managed alone.

Defects will occur, the same as in any software project. AI hasn't replaced the human judgment behind this project. It's what makes applying that judgment at a useful pace possible.

My time goes toward reading diffs, challenging design choices, and deciding what matters, not writing boilerplate. The test suite is the enforcement mechanism: over 5,000 tests mean AI-generated code has to survive real scrutiny, not just look plausible.

The project was originally forked from Adaptive Cover by @basbruss. See Credits for the full attribution.

What AI does, and what I do

Task	Owner
Architecture and scope decisions	Me
What ships and what doesn't	Me
Issue triage decisions	Me
Code generation, refactors, test writing	AI, reviewed by me
Documentation, wiki updates, release notes	AI, reviewed by me
First-pass replies on routine issues	AI, edited by me
Solar math, pipeline logic, anything safety-adjacent	Generated by AI, gated by the test suite, reviewed by me

Over 5,000 tests cover calculations, the override pipeline, managers, and the HA boundary layer. I will not merge a red build, and the suite is what catches regressions when AI-generated code touches solar math or the priority chain.

The skill-routed workflow

Different parts of the work go to different Claude models, picked for fit:

• Issue investigation: Opus, for reading the diagnostic, reproducing the failure, and identifying the cause.
• Design and TDD plans: Opus, for the implementation strategy and the test cases that have to pass first.
• Implementation: Sonnet by default, Opus when the change is flagged complex.
• Documentation updates: Sonnet.
• Issue replies and release-cutting: Haiku, where speed matters more than depth.
• Secondary code review: Opus, on changes flagged complex; results land in front of me before I let any correction proceed.

The routing lives in skill files under .claude/skills/, which are tracked in the repo (the !.claude/skills/ exception in .gitignore). If you want to see the actual prompts and routing logic, that's where they live.

Why CLAUDE.md is not committed

You may notice release notes or commits referencing CLAUDE.md while the file itself is excluded from the repo. CLAUDE.md is a per-machine personal config: it's a symlink into a private dotfiles repo and contains paths, deployment SMB credentials, and personal preferences specific to my development environment. Sharing it would leak local config without adding anything useful.

The shared, repo-tracked AI tooling lives at .claude/skills/. That's where the workflow described above is actually defined. CLAUDE.md is the personal layer on top.

What this means for you

• Bug reports and feature requests go through the same triage as any human-written code project. If something is wrong, file an issue; I will look at it.
• PRs are welcome, but expect the same review I'd apply to any contribution. The test suite has to stay green and changes have to fit the existing architecture.
• If AI-assisted projects are not for you, that's fair. The integration is a fork of Adaptive Cover; the upstream is still there.

Background and acknowledgments

This page exists because Discussion #228 and Issue #308 asked the right questions. Both contributors raised the disclosure point in good faith and helped shape what's written here.