Add Claude Code skill and install command

[baelter]

Summary

• Adds cmd/skills/cloudamqp-cli/ — a Claude Code skill that teaches Claude how to use the cloudamqp CLI (commands, auth, async behavior, reference guides for scripting, upgrades, and VPC setup)
• Adds cmd/install.go — implements cloudamqp install skills which embeds the skill files in the binary and copies them to ~/.claude/skills/cloudamqp-cli/
• Updates README.md with a short Claude Code section

Mirrors the pattern from microsoft/playwright-cli.

Test plan

• cloudamqp install skills creates ~/.claude/skills/cloudamqp-cli/ with SKILL.md and references/
• Skill appears in Claude Code skill list after install
• go build succeeds with embedded skill files

[snichme]

There are references to non-existing flags in the skill.

Also is the skill needed? When I've asked Claude to use the CLI it has figured out the usage by appending --help to each sub command and then built up the knowledge.

[snichme]

--details and -o json isn't valid flags for the CLI

[baelter]

There are references to non-existing flags in the skill.

Will looking into it

Also is the skill needed? When I've asked Claude to use the CLI it has figured out the usage by appending --help to each sub command and then built up the knowledge.

It will be much better at knowing when and how to use the cli. You won't have to be as explicit in telling it to use the cli. E.g
"What regions are supported on CloudAMQP?" - without to skill I doubt it will figure out to use the cli?

[snichme]

"What regions are supported on CloudAMQP?" - without to skill I doubt it will figure out to use the cli?

Yes, if it makes the usage better of course.

I know this works now:

"Using the CloudAMQP CLI, create a new 3-node RabbitMQ cluster called 'production-bunny' in the US-East region with the bunny plan. Once the cluster is provisioned and ready, enable the MQTT plugin."

But you need to tell it to use the CLI :)

[baelter]

But you need to tell it to use the CLI

Exactly :)

[baelter]

Re the comment on scripting.md: --details and -o json are valid. --details is defined on instance list, and -o is a persistent global flag (--output) on the root command. cloudamqp instance list --help shows both.

[baelter]

Skill value: before/after comparison

Ran 3 prompts with and without the skill loaded (using --append-system-prompt to inject skill content for the WITH case; baseline used the model's training data only). The main finding: Claude's training data has a completely different and wrong interface for this CLI.

---

Prompt: "spin up a RabbitMQ instance on AWS in us-east-1"

Without skill — Claude fabricates a positional-argument interface:

cloudamqp instance create my-rabbitmq lemur amazon-web-services::us-east-1

It also invents plan names (tiger, bunny, rabbit) and gets the instance info syntax wrong.

With skill — correct flag syntax, and tells you to look up plan/region names first:

cloudamqp plans --backend=rabbitmq
cloudamqp regions --provider=amazon-web-services

cloudamqp instance create \
  --name=my-rabbit \
  --plan=<plan-name> \
  --region=amazon-web-services::us-east-1 \
  --wait

---

Prompt: "get the connection URL for instance 12345 without showing credentials"

Without skill — wrong flag (--format instead of -o), positional ID:

cloudamqp instance get 12345 --format=json | jq -r '.url | gsub(...)'

With skill — correct, plus knows about the --show-url flag for the inverse case:

cloudamqp instance get --id 12345 -o json | jq -r '.url'
# URL is masked by default; add --show-url if you want credentials

---

Prompt: "upgrade all instances tagged staging to latest RabbitMQ"

Both produced correct output here. The upgrade patterns are common enough that the model gets them right from training data alone.

---

The gap is widest where the CLI deviates from conventions Claude has seen before. The training data shows a positional-argument style CLI; the actual CLI uses named flags throughout. Without the skill, Claude confidently produces wrong syntax. With it, it uses the right interface on the first try.

[snichme]

Re the comment on scripting.md: --details and -o json are valid. --details is defined on instance list, and -o is a persistent global flag (--output) on the root command. cloudamqp instance list --help shows both.

My bad, I was checking against an older version of the CLI. Yes they exists and work.