Skip to content

robacarp/dotfiles / .claude/skills/tf-drift

tf-drift

Verified

Backport AWS drift into terraform code so that `terraform plan` is clean (or shows only an explicitly intended diff). Fans out one sub-agent per drifted resource, each working from `terraform plan -target=<addr>` and `terraform state show <addr>` to update the .tf source until that resource's diff disappears. Use when the user asks to "reconcile drift", "backport drift", "make plan clean", or "update code to match AWS".

Terraform drift backport

Use this when AWS state and terraform code have diverged and the user wants the code updated to reflect AWS reality (not the other way around — no terraform apply is involved).

Inputs the user typically provides

  • The terraform project directory (default: cwd).
  • An optional intended diff to preserve (e.g. "only the BusinessUnit tag should remain in the plan"). Pass this through to every sub-agent — they need it to know which diff lines are noise vs. signal.

If the user hasn't named an intended diff, ask once. Without it you cannot tell when a resource is "done."

Workflow

1. Bootstrap

terraform plan -no-color > /tmp/tf-drift-plan.out 2>&1

Plans on real infra are slow (often 60–120s). Run in the background and use Monitor (or a poll loop) to wait — never block the conversation on sleep.

If terraform init hasn't been run, ask the user to run it (or to confirm before you do — init can pull modules and take a while).

2. Enumerate drifted resources

From the plan output, extract the address of every resource whose action is update in-place, replace, create, destroy, or import and whose diff is not purely the intended diff.

grep -E "^  # .* will be (updated|created|destroyed|imported)|must be replaced" /tmp/tf-drift-plan.out

Skip resources that only show the intended diff (e.g. only + "BusinessUnit" = "amp" under tags_all). A quick heuristic: if -target plan for that resource produces only the intended diff, it doesn't need a sub-agent.

3. Fan out — one sub-agent per resource

For each drifted resource, dispatch a general-purpose Agent in parallel (multiple Agent calls in one assistant turn). Per-resource isolation keeps each sub-agent's context narrow and lets terraform plans run concurrently across resources.

The prompt for each sub-agent should be self-contained and include:

  1. Goal: "Update the terraform source for <resource address> so that terraform plan -target='<resource address>' shows only <intended diff, verbatim>. Do not run terraform apply."
  2. Working directory: absolute path to the terraform project.
  3. Resource address as a quoted string (module paths contain dots — use single quotes in shell).
  4. The two reference outputs the sub-agent should re-run itself:
    • terraform plan -target='<addr>' -no-color
    • terraform state show '<addr>' The sub-agent should not trust a snapshot you paste in — drift evolves and re-running gives ground truth. Tell it to run them itself.
  5. Source file pointer: grep for the resource block to point the sub-agent at the right .tf file:
    grep -rn "resource \"<type>\" \"<name>\"" *.tf
    
    For module-internal resources (module.X.aws_foo.bar), point them at the module input in this project — they cannot edit the module's own source from this project.
  6. Constraints:
    • Only edit .tf files in the project directory (not the module sources under .terraform/).
    • For resources that exist in AWS but not in state, add an import { to = ...; id = "..." } block.
    • For resources present in code but absent in state and AWS, delete them.
    • For pure attribute drift, edit code values to match terraform state show.
    • For module-internal drift that can't be fixed without changing the module: report it back, don't bash on it.
    • Verify by re-running terraform plan -target='<addr>' and confirming the output matches the intended diff. Re-iterate until clean or until the sub-agent declares the resource module-locked.
  7. Reporting: ask each sub-agent to return a one-line summary: <addr>: <clean | residual: <one-line reason>>.

4. Reassemble and verify

After all sub-agents return, run a full terraform plan once more. Compare the remaining diff against the intended diff:

  • If they match → done. Surface the per-resource summaries to the user.
  • If residuals remain, list them with the reasons collected from sub-agents. Do not silently retry — module-locked or AWS-side drift may need a human decision (bump module ref, manual aws ... fix, accept the residual, etc.).

Common drift patterns and how a sub-agent should handle them

  • Resource in AWS, not in state, in code: add import { to = <addr>; id = "<aws id>" }. Use the resource's natural import id (ARN for LB stuff, name for IAM, etc. — terraform docs per resource type).
  • Resource in AWS, not in state, not in code: write a new resource block matching terraform state show (after import), plus the import block.
  • Resource in state, not in AWS: rare — usually means someone manually deleted in AWS. Confirm with the user; do not blindly remove from state.
  • Resource in state and AWS, not in code: terraform plan says "will be destroyed". Add the resource to code (don't accept the destroy).
  • Attribute drift: edit the attribute in code to match terraform state show. Be careful with computed/(known after apply) fields — leave them alone.
  • tags_all differs by exactly the intended diff: leave it (this is the intended diff).
  • Module-internal drift (e.g. module.foo.aws_bar.baz): you can usually pass the right input via the module call (task_environment_variables, etc.) or bump the module ref to a version whose generated config matches AWS. If neither works, report and stop — don't fork the module from inside this project.
  • Imports counted in plan don't drop on re-plan: that's normal — terraform plan doesn't write state. The diff stays the same across plans until terraform apply -auto-approve or an explicit terraform apply runs the import. The user controls when to apply.

What this skill won't do

  • It will not run terraform apply. The output is code changes the user reviews and applies themselves.
  • It will not modify shared module source under ~/Repositories/.../terraform-modules unless the user explicitly asks.
  • It will not invent AWS resources that aren't in the plan output. If a sub-agent thinks it found drift outside the plan, it should stop and surface that.
writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deploymentStars: 319kUpdated Claude CodeCodexVerifiedInstall
xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy.Stars: 118kUpdated Claude CodeCodexVerifiedInstall
childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release.Stars: 117kUpdated Claude CodeCodexVerifiedInstall