aws-ops

Asks about AWS itself, apart from any repo: "from where to delete access keys?", "what access keys do we have saved here?", someone's surprise S3 bill, and "give the exact prompt for what needs to be done" for an agent that holds the credentials. Reads run freely; anything that creates, changes or deletes waits for an explicit ask naming the resource.

Read first

• Your always-on rules, if you keep them. Secrets: keys and tokens by variable name only; a pasted credential is flagged for rotation, never used
• Account facts in the brain (brain recall bedrock, brain recall aws): auth method (bearer token vs SigV4 vs profile), which inference profile prefix works (us., eu., global.) and which is known-broken, the region per model family, and any data-retention setting a gateway must preserve. Record new ones with brain remember.
• Lessons for any account: confirm the auth method before any integration step; probe reasoning models with maxTokens >= 200 (a 500 is a fault, empty text with usage is token starvation); real Bedrock spend lives in AWS, not a vendor dashboard; Bedrock Mantle (OpenAI-compatible gateway) paths differ per region and model family, so verify live, and never route Claude through it if it drops a retention setting the account needs

Procedure

1. Identity first, always: aws sts get-caller-identity and aws configure list (key id masked, its source, the region). Say which account and identity the session is in; a work or client account is never touched from a personal identity.
2. Access keys, local: aws configure list-profiles; ls ~/.aws for which files exist; env | grep -oE '^AWS_[A-Z_]+' for variable names. Never open ~/.aws/credentials or an .env; the profile name is the answer.
3. Access keys, IAM: aws iam list-access-keys --user-name <user> and aws iam get-access-key-last-used --access-key-id <id>. Retire in two steps and only on ask: aws iam update-access-key --status Inactive --access-key-id <id> --user-name <user>, wait for the user to confirm nothing broke, then aws iam delete-access-key with the same arguments. Root keys are console only (account menu, Security credentials). Verify: list-access-keys no longer shows the id.
4. "If I give you a temporary access key can you find out?": yes, without pasting it. The user runs aws configure --profile temp in their own terminal and types the values there, says done, and every read below runs with --profile temp. The key never enters the transcript; the user deletes it after.
5. Unexpected S3 bill: aws ce get-cost-and-usage --time-period Start=<first of month>,End=<today> --granularity MONTHLY --metrics UnblendedCost --group-by Type=DIMENSION,Key=SERVICE, then filtered to S3 with --group-by Type=DIMENSION,Key=USAGE_TYPE (storage vs requests vs data transfer); aws s3api list-buckets; per bucket get-bucket-location and get-bucket-policy-status; CloudWatch AWS/S3 BucketSizeBytes daily; aws cloudtrail lookup-events --lookup-attributes AttributeKey=EventSource,AttributeValue=s3.amazonaws.com for who created what. On a "nothing created" account the usual causes, in order: a leaked key used by someone else (iam list-users, recent list-access-keys), a public bucket serving requests, CloudTrail data events writing to S3. Cost Explorer needs enabling once and lags about 24 hours; say so.
6. Least privilege: list the exact API calls the task makes, write policy.json with those actions and resource ARNs, run aws accessanalyzer validate-policy --policy-document file://policy.json --policy-type IDENTITY_POLICY, and stop. aws iam create-policy only on "do it".
7. Bedrock: aws bedrock list-inference-profiles --region <region> filtered on the model name, aws bedrock list-foundation-models --by-provider anthropic, aws service-quotas list-service-quotas --service-code bedrock. Liveness: aws bedrock-runtime converse --model-id <profile>.anthropic.<model> --messages '[{"role":"user","content":[{"text":"reply ALIVE"}]}]' --inference-config maxTokens=300 --region <region> with the account's auth (for example AWS_BEARER_TOKEN_BEDROCK) set in the shell. Spend: Cost Explorer filtered to Amazon Bedrock plus CloudWatch AWS/Bedrock InputTokenCount and OutputTokenCount by ModelId; cost = in / 1e6 * in_price + out / 1e6 * out_price with prices fetched live (the claude-api skill or the pricing page), never from memory.
8. Runbook prompt for an agent with the user's credentials: identity check first, region stated, read phase with exact commands, a stop-and-report gate before any create or delete, the output table below, and the Mantle endpoint table when asked to include Bedrock Mantle. Credentials are named (AWS_PROFILE, AWS_BEARER_TOKEN_BEDROCK, OPENAI_API_KEY), never pasted into the prompt.
9. Console-only steps (root keys, billing preferences, budgets) are printed as numbered clicks; the user does them and says done, then step 1 or 3 re-verifies.

Output

identity: <arn from sts, account id masked to last 4> | region <r> | profile <p>

| question | evidence (command) | result | next step (yours or mine) |
| keys in use | aws iam list-access-keys | 2 keys, one unused since <YYYY-MM-DD> | yours: confirm, then I run update-access-key Inactive |

console steps for you: <numbered, or "none">
unverified: <what could not be read and why>

Must not

• Never ask for, accept in chat, or echo an access key, secret, session token or bearer token; if one is pasted, say so, name what to rotate, and proceed through a profile the user configures themselves.
• Never cat ~/.aws/credentials, .env, or a desktop app config env block; bearer tokens surface there in plaintext.
• Never create, modify or delete an AWS resource, policy, key or budget without an explicit ask naming it; reads are always fine.
• Never use an inference profile or gateway the brain records as broken for this account.
• Never touch a work or client account from a personal identity, or store its details in the brain.
• Never quote model pricing or IDs from memory.
• Never chain commands with &&; one command per call.

Non-goals

Terraform in the user's IaC repos (normal coding there), Vercel, Neon and Render checks (prod-check), Claude Code settings and model switching (update-config, if available), Bedrock image and video generation (a Bedrock MCP server, if installed).