Terraform skills review: Anton Babenko's terraform-skill verdict, safety and alternatives

Anton Babenko's Terraform skill is a Markdown-only Agent Skill that makes agents plan before any apply or destroy. Our verdict is Recommended, 76/100, at Low confidence, with safety grade A. Its main costs are token weight and a narrow scope.

Get the skill: terraform-skill folder on GitHub · SKILL.md at the reviewed commit · skills.sh listing

What is Anton Babenko's Terraform skill?

Anton Babenko's Terraform skill is a Markdown-only Agent Skill for Terraform and OpenTofu. It routes each request through a failure-mode table and loads one of eight references on demand. Every answer follows a plan-first, rollback-aware contract.

Terraform is HashiCorp's infrastructure-as-code tool. People use it to describe cloud resources in code and apply the changes. OpenTofu is the open-source fork of Terraform, with the same language. The skill covers both.

An Agent Skill is a folder of instructions that a coding agent loads when a task matches. Our guide to what an Agent Skill is covers the format. This one holds a SKILL.md file and eight reference files. It contains no code at all.

Anton Babenko maintains it. He is the top contributor to terraform-aws-modules/terraform-aws-vpc, with 222 commits (GitHub API, 2026-10-20). His skill turns that experience into rules for an agent.

The workflow has four steps. The agent captures the runtime and version first. Next it classifies the request against a failure-mode table. Then it loads the one matching reference. Last, it answers in a fixed five-part shape: assumptions, risk, tradeoffs, validation commands and rollback notes.

Is the Terraform skill worth installing in Claude?

Yes, the Terraform skill is worth installing in Claude for IaC work: 76/100, Recommended, at Low confidence. Its plan-before-destroy rule came from a real Route53 incident. It costs about 5,105 tokens on activation.

Its strength is concrete safety rules learned from a real failure. A targeted destroy with auto-approve wiped Route53 records in issue #42. A plan-destroy-first rule shipped within three days in PR #43. Its weakness is heavy context. SKILL.md is about 5.1K tokens, and the references total about 44K by our estimate.

It suits platform and DevOps engineers who use Claude Code, Codex, Cursor or Kiro on IaC repos. The best fit is a team that writes or reviews modules, tests, CI pipelines and state operations. Those teams want the agent to state version floors and show a reviewed plan first. The poor fit is a Terragrunt or Terramate repo, or a quick one-line HCL question.

The worst failure for this job is an agent that applies or destroys infrastructure without a reviewed plan. The skill now guards that case in writing. Token weight and scope limits pull the score down.

Confidence is Low because nobody has measured the skill's output. Our page on how confidence levels work explains that rating. The evidence itself rates Moderate, covered in the community section. You can compare this skill with every skill ranked by verdict on one list.

How many tokens does the Terraform skill use?

The Terraform skill loads about 5,105 tokens on activation, and its eight references total about 44,000 by our estimate. One module PR review is estimated at 22,000 input and 3,000 output tokens. Our typical-call figure is 28,105 input tokens.

Each row below says whether the figure is a measured file size or our estimate.

Every response carries the five-part contract, even for a quick question. Deep tasks load references of 500 to 1,800 lines each. state-management.md alone runs to 1,834 lines. Our usage estimate assumes 10 uses a week.

TerraShark is a rival Terraform skill. Its README claims Babenko's skill costs over 7x its own core. That figure is vendor-reported and not independently reproduced.

What changed since our review?

Two issues opened after our review, and one affects a claimed install path. The skill itself has not changed. The verdict stays Recommended at 76.

GitHub showed this on 2026-10-21:

• No release since v1.17.1 on 2026-06-03.
• No commit since 2026-07-03, our reviewed commit.
• Issue #55 opened 2026-10-04 and is still open: the Copilot /plugin install command is not working.
• Issue #56 opened 2026-10-10 as a feature request for Azure DevOps pipeline examples.
• Stars rose from 2,391 to 2,420.

The repository had zero open issues at our review. Changes to verdicts go into our re-review log.

Is the Terraform skill safe to install?

Yes, the Terraform skill is safe to install, with safety grade A. It is nine Markdown files with no scripts, hooks, credentials or telemetry. On skills.sh, Snyk shows Warn while Gen Agent Trust Hub and Socket pass.

The skill folder holds no scripts, hooks, MCP config or network indicators. It needs no credentials. Examples use placeholders, such as the AWS documentation key AKIAIOSFODNN7EXAMPLE in security-compliance.md. A skill linter flagged that key as a secret, and skillsaw issue #147 called the hit a false positive.

Four safety findings sit in or near the files:

• The repo root ships an mcp.json for Kiro Powers. It runs hashicorp/terraform-mcp-server:latest through Docker, unpinned.
• A reference file holds a curl-pipe-sh install snippet for trivy. The skill never runs it; users do.
• code-intelligence-lsp.md warns against auto-running terraform-ls in untrusted directories.
• The skill gives no general prompt-injection guidance.

The Terraform MCP server is a separate HashiCorp tool that agents call over MCP. Our explainer on how skills differ from MCP servers covers that split. The unpinned image matters only to Kiro users who turn on the Power.

The response contract is the strongest safety feature. It forbids a production apply without a reviewed plan. It forbids a destroy without plan -destroy and explicit confirmation. It bans -auto-approve on destroy.

The audits conflict. skills.sh shows Snyk Warn, with Gen Agent Trust Hub and Socket at Pass. Tessl's registry showed a "Critical" label on 2026-10-20, with a 94% quality score. Neither Snyk nor Tessl itemises the cause, so we name none. Our guide to vetting agent skills for safety shows how to read these labels.

What does the community report about the Terraform skill?

The community reports about the Terraform skill lean positive, with two strong items. Issue #42 reported a targeted destroy that deleted all Route53 records. PR #43 added a destroy rule within three days.

These are the 14 dated reports we weighed: 5 positive, 4 negative and 5 neutral. Two are strong, and the evidence as a whole rates Moderate.

What are the Terraform skill's known failures?

The Terraform skill's known failures were fixed by v1.17.1, but three limits remain. Its rules are prose with no hook enforcing them. Most AWS provider pins show ~> 5.0, and Terragrunt is unsupported.

The worst failure for this job is an agent that applies or destroys infrastructure without a reviewed plan. A second form is HCL that silently forces destroy and recreate, or corrupts state in production. The skill guards both cases in prose.

Five limits stand out:

1. Prose rules: no hook enforces plan-before-apply, so the model must choose to follow it.
2. Pin drift: most examples pin hashicorp/aws ~> 5.0, while quick-reference.md uses ~> 6.0.
3. Wrapper gap: the maintainer declined Terragrunt and Terramate support in issue #52.
4. Answer overhead: the five-part contract applies to quick one-line questions too.
5. Generic scope: issue #12 shows it is not AWS- or provider-specific.

Prose rules depend on the model reading and obeying them. Pin drift sends mixed signals about which AWS provider major version to use. The wrapper gap leaves Terragrunt users with plain Terraform patterns that do not match their repos.

Two earlier defects are fixed. The Route53 cascade in issue #42 led to PR #43. Factual errors found in April were fixed in PR #21.

How does the Terraform skill compare with HashiCorp's skills and TerraShark?

Anton Babenko's Terraform skill is broader than HashiCorp's narrower skills and focuses on safe changes. HashiCorp's Code Generation skill focuses on HCL code generation. TerraShark claims about 600 tokens but has no verdict from us.

Six Terraform skills compare here. HashiCorp publishes three of them in hashicorp/agent-skills.

The HashiCorp skills split the work by job. Code Generation writes HCL. Module Generation refactors modules and builds stacks. Provider Development helps people write Terraform providers. Babenko's skill covers the whole module lifecycle in one package.

TerraShark's site calls it "the #1 Terraform Skill" and claims it is 7x more token-efficient. Those are vendor claims. We have not evaluated TerraShark.

Kiro ships its own Terraform skills, and one dev.to article builds a custom skill for Kiro. The SKILL.md has 375 exact copies on GitHub code search, including forks. AWS's agent toolkit is building its own Terraform skill in aws/agent-toolkit-for-aws issue #115. Babenko asked them to reuse his.

HashiCorp's skills each teach the agent to produce code for one task. Babenko's skill teaches the agent to change live infrastructure carefully. Its contract asks for risk and rollback notes on every answer.

TerraShark is built to be small. Its site describes a small core, five failure modes, a six-step workflow and 18 reference modules. Babenko's skill keeps a bigger SKILL.md and eight long references. A team that cares most about tokens has a reason to test TerraShark. A team that cares most about destroy guards has more public evidence for Babenko's skill.

Does the Terraform skill work in Codex, Cursor and Copilot?

No run is documented for the Terraform skill outside Claude Code, where it is verified. The README gives install paths for Codex, Cursor, Copilot and six more hosts. Issue #55 reports the Copilot plugin install failing.

"Claimed" means the README documents the path, but we found no public report of a run. The paths below come from the README on 2026-10-21.

The README lists ~/.copilot/skills/ as a second Copilot path. Issue #55 covers only the /plugin install route. Our list of Claude Code skills worth installing covers the agent where this skill is verified.

How do you install the Terraform skill?

Install the Terraform skill with npx skills add https://github.com/antonbabenko/terraform-skill. In Claude Code, add the antonbabenko/agent-plugins marketplace, then install terraform-skill@antonbabenko. You need the terraform or tofu CLI for validation, but no credentials.

Prepare these tools first:

• Install the terraform or tofu CLI, which the validation commands need.
• Add terraform-ls for code intelligence, as an option.
• Add tflint, trivy or checkov for linting and security scans, as options.

Then follow these steps:

1. Run npx skills add https://github.com/antonbabenko/terraform-skill for any Agent Skills host.
2. Or, in Claude Code, run /plugin marketplace add antonbabenko/agent-plugins.
3. Run /plugin install terraform-skill@antonbabenko to finish the Claude Code route.
4. Ask a Terraform question and check the answer shows assumptions, risk, tradeoffs, validation commands and rollback notes.

That last step is our suggested check. A five-part answer shows the skill loaded.

Two warnings apply. agenticskills.io offers a curl route that copies SKILL.md only, so the eight references go missing. The README warns against adding antonbabenko/terraform-skill as a separate Codex marketplace, because the names clash. Setup friction scores 85. Our guide on how to add skills in Claude Code covers the general steps.

How do you use the Terraform skill?

You use the Terraform skill by asking the agent to review a module, refactor resources or change state. It asks for versions, classifies the failure mode and loads one reference. Answers include risk, validation commands and rollback notes.

Three tasks fit it best:

1. Review or refactor modules with count or for_each identity and moved or import blocks. Module authors do this weekly.
2. Guard apply, targeted destroy and backend migration. Engineers on production stacks do this weekly.
3. Set up native terraform test, mocks or Terratest. Module maintainers do this monthly.

A moved block tells Terraform that a resource changed address, so it is not destroyed and recreated. An import block brings an existing resource under Terraform's control. Both appear in the skill's refactor guidance.

Two uses are claimed only by the publisher. One is CI pipelines on GitHub Actions, GitLab or Atlantis with OIDC. The other is backends, auth and secrets across AWS, Azure and GCP. A third use, removing bootstrap provisioners, came from an outside contributor in PR #31.

Every answer follows the same contract:

• Assumptions about runtime and version.
• Risk of the change.
• Tradeoffs between options.
• Validation commands to run.
• Rollback notes.

No published with-and-without benchmark exists. Test it yourself with two prompts in a sandbox repo:

• Ask the agent to rename a count-indexed resource list.
• Ask it to destroy one targeted resource.
• Check that it uses moved blocks.
• Check that it shows plan -destroy before acting.

Common questions

For Claude skills in general, read Claude skills explained.

Does the Terraform skill support OpenTofu?

Yes, it covers Terraform and OpenTofu. The same rules apply to both CLIs.

Does the Terraform skill support Terragrunt?

No, the maintainer declined wrapper tools in issue #52. Terragrunt repos get raw Terraform patterns.

Is the Terraform skill an official HashiCorp skill?

No, Anton Babenko maintains it. HashiCorp publishes separate skills in hashicorp/agent-skills.

Is the Terraform skill free?

Yes, its LICENSE and SKILL.md say Apache-2.0. GitHub's API shows NOASSERTION, and agenticskills.io shows Apache-2.0.

Where is the Terraform skill on GitHub?

It sits at skills/terraform-skill in antonbabenko/terraform-skill. Our page on what sits inside a SKILL.md file explains the main file.

Does the Terraform skill work in Kiro?

Kiro support is claimed: the README documents a Kiro Power with an unpinned Terraform MCP image. No run is documented.