How Goose flipped my work ratio from 80/20 to 20/80

[clouatre]

I've been using Goose for infrastructure work and noticed something unexpected: my time split flipped from 80% implementing/20% deciding to 20% reviewing/80% judgment calls.

My setup

Environment:

GOOSE_PROVIDER: gcp_vertex_ai
GOOSE_MODEL: claude-sonnet-4-5@20250929

Tools Goose leverages:

• gh - Reviews PRs, issues, and repos before making decisions
• wrangler - Cloudflare Pages/Workers interaction
• aws - Route53, infrastructure discovery
• mermaid-cli - Architecture diagrams
• Standard POSIX tools (dig, curl, git, etc.)

Pattern: Goose discovers infrastructure state through CLI tools, proposes changes, I review and approve.

---

Two recent projects where this really showed up:

DNS Migration (2 hours, zero downtime)

Migrated production DNS (Route53 → Cloudflare) with 20+ records including email.

What Goose did:

• Discovered all Route53 records
• Pre-validated against Cloudflare nameservers before switching
• Generated validation report (100% match)
• Automated cutover

Results: 2h total, 88% faster deploys (5-8min → 38sec), zero downtime

Key insight: Pre-validation eliminated risk. I knew email would work before changing nameservers.

Read more: Migrating to Cloudflare Pages: One Prompt, Zero Manual Work

---

CI Modernization (20 minutes)

Added modern CI to a project with outdated tooling (slow mypy).

What worked: Goose learned my modern stack from another project (Ruff + uv + pytest-cov) and applied it automatically.

Results: 20min (vs 3-4h from scratch), 10-100x faster linting, 83% coverage

Key insight: Cross-project pattern transfer. I provided judgment (simplicity over complexity), Goose handled implementation.

Read more: AI-Assisted Development: Judgment Over Implementation

---

What I'm learning

Works well:

• Infrastructure tasks (complex, infrequent, high stakes)
• Pre-validation strategies (test before executing)
• Pattern reuse across projects
• Human gates at critical decisions

Doesn't work:

• Tasks where I lack domain knowledge to evaluate
• Time-sensitive fixes (no review time)
• Blind automation without oversight

The shift: Less time on "how to implement", more on "is this the right approach?" My value moved from execution to judgment.

---

Questions for you

1. Are you seeing this ratio shift too? More review, less typing?
2. What infrastructure recipes are you using? I'd love to see what patterns you've automated.
3. What concerns do you have about AI-assisted infra changes?

Happy to discuss approaches, share configs, or clarify details. All metrics verified via PR timestamps.

[clouatre]

A note on my evaluation process

I tested several AI CLI tools before settling on Goose (Claude Code, OpenAI Codex, OpenCode, Amazon Q CLI, Copilot CLI, Continue.dev CLI, Cursor CLI).

What made Goose stand out:

• Works with any provider (I use Claude Sonnet 4.5 via GCP Vertex AI)
• Strong tool calling with large models
• CLI-native workflow (leverages gh, aws, wrangler, etc.)
• Review/approve gates (not fully autonomous)

Important: Smaller models (7B-8B) don't work well with Goose due to tool calling format issues. The results I'm sharing require Claude Sonnet 4.5 or similar frontier models.