snitch-aws
You orchestrate. ~/.claude/skills/snitch-aws/snitch-aws.sh exposes thin tools — read-only data tools emit JSON; mutating tools (fix, panic) are explicit and idempotent. You synthesize: classify intent, prioritize findings, render prose. Run bash ~/.claude/skills/snitch-aws/snitch-aws.sh help for the surface.
Setup
- AWS credentials. Preferred: SSO profiles (
aws configure sso) or assumed roles. Refused when AWS_ACCESS_KEY_ID starts with AKIA* and an SSO profile exists in ~/.aws/config. Hard-refused: no creds.
aws CLI v2. jq required for parsing. doctor flags missing.- Optional: AWS-flavored MCP — export
AWSSEC_MCP_PRESENT=1 to prefer it for inventory reads.
Tool inventory
Read-only (JSON on stdout, errors as JSON on stderr):
| Subcommand |
Returns |
doctor |
env health (aws cli, jq, profile, sts caller, MCP) |
detect |
cwd signals: stacks, databases, storage, IaC, AI providers, hostnames, project_kind |
state <area> [slice] |
per-area state JSON; digest is the default |
analytics |
account-level signals: untagged spend, top services 30d, region distribution |
events |
last 100 IAM-significant CloudTrail events |
fit-matrix [stack] |
migration verdict + caveats (AWS-native landing options) |
stack-docs [stack] |
canonical doc URLs to WebFetch |
score [host...] |
SSL Labs / Mozilla Observatory / securityheaders / hsts-preload grades |
state subscopes (digest by default):
account, iam, s3, ec2, vpc, rds, dynamodb, lambda, cloudfront, route53, acm, cognito, secrets, cloudtrail, cloudwatch, wafv2, shield, config, inspector, macie, guardduty, securityhub, backup, kms, organizations, eks, ecs, eventbridge, sqs-sns, cost.
Each digest emits a summary (counts + key risk signals). Slice values vary per subscope — see per-area headers for the valid set.
Mutating (idempotent):
| Subcommand |
Behavior |
fix <area> [args] |
apply hardening for one area; safe to re-run. Areas: iam s3 ec2 vpc rds lambda cloudfront cloudtrail kms secrets wafv2 guardduty securityhub backup account all |
panic <action> [args] |
incident: revoke-key <id>, quarantine-role <arn>, block-ip <ip>, restore. Confirm with user before invoking. |
Utility: export, terraform, verify, refresh-docs, help.
Removed (print deprecation + exit 64): check, migrate, roadmap, report, diagnose, stacks. Compose primitives — see references/30-recipes.md.
CLI vs MCP vs skill — division of labor
The AWS CLI is the primary tool. Drive it via the skill's aws_run wrapper.
| Operation |
AWS service / API |
| Account / SCP / Org |
organizations, iam |
| Identity inventory |
iam list-users / list-roles / list-policies / list-access-keys, accessanalyzer list-findings |
| S3 reads |
s3api get-public-access-block / get-bucket-policy / get-bucket-encryption / get-bucket-versioning / get-bucket-logging |
| EC2 + VPC |
ec2 describe-instances / describe-volumes / describe-security-groups / describe-vpcs / describe-flow-logs |
| Data services |
rds, dynamodb, elasticache, dax |
| Edge + DNS |
cloudfront list-distributions, route53 list-hosted-zones, acm list-certificates, wafv2 list-web-acls |
| Detective |
cloudtrail describe-trails, config describe-configuration-recorders, guardduty list-detectors, securityhub get-enabled-standards, inspector2, macie2 |
| Secrets / keys |
secretsmanager list-secrets, kms list-keys / get-key-rotation-status |
| Containers |
eks list-clusters, ecs list-clusters |
| Lambda |
lambda list-functions / get-function-configuration / get-function-url-config |
| Cost |
ce get-cost-and-usage, budgets describe-budgets, ce get-anomalies |
Use an AWS MCP (if AWSSEC_MCP_PRESENT=1) for typed inventory reads (S3, DynamoDB, Lambda, RDS lists). Fall back to aws_run otherwise.
Use this skill for: digest views (counts + derived risk signals from many CLI calls), mutating actions (fix, panic — read-first, idempotent, recorded), offline tools (detect, fit-matrix, stack-docs, score).
If neither CLI nor MCP is available, the skill cannot operate. doctor is the pre-flight.
Orchestration
For any AWS request:
Classify intent yourself (audit, migrate, scale-plan, diagnose, incident).
Call the smallest set of tools that gives you what you need. Start with state account + state iam + state s3 digests in parallel; add state cloudtrail, state guardduty, state securityhub, state config for an audit. Run independent calls in parallel (single message, multiple Bash calls).
Lazy-load references that match findings. Read references/30-recipes.md for orchestration detail. Read references/<NN>-<area>.md when an area surfaces an issue. Read references/15-stack-best-practices/<stack>.md for stack-tailored guidance.
Synthesize the report. Group findings by area; mark OK / WARN / FAIL; surface paid-only items as [locked: <tier>+] with one-line value statements (from references/10-plan-tier-matrix.md).
For project file changes (fix project, fix gha), the tool emits proposed contents + unified diff to stdout:
=== FILE: <relative-path> ===
=== DIFF ===
<unified diff>
=== CONTENT ===
<full proposed file body>
=== END ===
Apply with Edit or Write after user confirms. The skill never writes inside the user's project. For Lambda secret values, the tool emits aws secretsmanager create-secret invocations — never type secret values yourself.
Recipes
Canonical recipes live in references/30-recipes.md. Read it for any specific recipe (audit, migrate, scaling, diagnose, attack response). Recipes assume the digest-by-default contract above.
Guardrails
- Refuses when
AWS_ACCESS_KEY_ID starts with AKIA and at least one SSO profile exists in ~/.aws/config. Redirect: AWS_PROFILE=<sso-profile> aws sso login.
- Hard-refuses when no AWS credentials are visible (env, profile, instance role).
doctor prints remediation.
fix is idempotent — no-op when state matches target. Each *_fix reads first, compares, mutates only on drift.panic records each action to .state/panic-<ts>.json; panic restore rolls back.- Honest verdicts: PHP/Rails/Django stacks usually get
partial or proxy-only — surface that first.
- The skill never lowers posture on
*_fix (never opens a Public Access Block, never decreases CloudTrail multi-region coverage).
1---2name: snitch-aws3description: AWS security + readiness skill. Thin tools for the agent to compose. Detects the user's project, audits account/IAM/S3/EC2/VPC/RDS/Lambda/CloudFront/Route53/CloudTrail/GuardDuty/SecurityHub/Config/KMS posture across services, applies idempotent hardening, and produces honest migration / scaling guidance. Triggers on audit my AWS account, harden AWS, AWS security audit, secure my AWS infrastructure, AWS WAF setup, AWS IAM audit, AWS scaling readiness, AWS incident response, AWS best practices, AWS hardening, should I move to AWS, migrate to AWS, AWS DNSSEC, AWS S3 public bucket, AWS CloudTrail audit, AWS GuardDuty.4---56# snitch-aws78You orchestrate. `~/.claude/skills/snitch-aws/snitch-aws.sh` exposes thin tools — read-only data tools emit JSON; mutating tools (`fix`, `panic`) are explicit and idempotent. **You synthesize:** classify intent, prioritize findings, render prose. Run `bash ~/.claude/skills/snitch-aws/snitch-aws.sh help` for the surface.910## Setup1112- AWS credentials. **Preferred:** SSO profiles (`aws configure sso`) or assumed roles. **Refused** when `AWS_ACCESS_KEY_ID` starts with `AKIA*` and an SSO profile exists in `~/.aws/config`. **Hard-refused:** no creds.13- `aws` CLI v2. `jq` required for parsing. `doctor` flags missing.14- Optional: AWS-flavored MCP — export `AWSSEC_MCP_PRESENT=1` to prefer it for inventory reads.1516## Tool inventory1718Read-only (JSON on stdout, errors as JSON on stderr):1920| Subcommand | Returns |21|---|---|22| `doctor` | env health (aws cli, jq, profile, sts caller, MCP) |23| `detect` | cwd signals: stacks, databases, storage, IaC, AI providers, hostnames, project_kind |24| `state <area> [slice]` | per-area state JSON; `digest` is the default |25| `analytics` | account-level signals: untagged spend, top services 30d, region distribution |26| `events` | last 100 IAM-significant CloudTrail events |27| `fit-matrix [stack]` | migration verdict + caveats (AWS-native landing options) |28| `stack-docs [stack]` | canonical doc URLs to `WebFetch` |29| `score [host...]` | SSL Labs / Mozilla Observatory / securityheaders / hsts-preload grades |3031**`state` subscopes** (digest by default):3233`account, iam, s3, ec2, vpc, rds, dynamodb, lambda, cloudfront, route53, acm, cognito, secrets, cloudtrail, cloudwatch, wafv2, shield, config, inspector, macie, guardduty, securityhub, backup, kms, organizations, eks, ecs, eventbridge, sqs-sns, cost`.3435Each digest emits a summary (counts + key risk signals). Slice values vary per subscope — see per-area headers for the `valid` set.3637Mutating (idempotent):3839| Subcommand | Behavior |40|---|---|41| `fix <area> [args]` | apply hardening for one area; safe to re-run. Areas: `iam s3 ec2 vpc rds lambda cloudfront cloudtrail kms secrets wafv2 guardduty securityhub backup account all` |42| `panic <action> [args]` | incident: `revoke-key <id>`, `quarantine-role <arn>`, `block-ip <ip>`, `restore`. Confirm with user before invoking. |4344Utility: `export`, `terraform`, `verify`, `refresh-docs`, `help`.4546Removed (print deprecation + exit 64): `check`, `migrate`, `roadmap`, `report`, `diagnose`, `stacks`. Compose primitives — see `references/30-recipes.md`.4748## CLI vs MCP vs skill — division of labor4950The AWS CLI is the **primary** tool. Drive it via the skill's `aws_run` wrapper.5152| Operation | AWS service / API |53|---|---|54| Account / SCP / Org | `organizations`, `iam` |55| Identity inventory | `iam list-users / list-roles / list-policies / list-access-keys`, `accessanalyzer list-findings` |56| S3 reads | `s3api get-public-access-block / get-bucket-policy / get-bucket-encryption / get-bucket-versioning / get-bucket-logging` |57| EC2 + VPC | `ec2 describe-instances / describe-volumes / describe-security-groups / describe-vpcs / describe-flow-logs` |58| Data services | `rds`, `dynamodb`, `elasticache`, `dax` |59| Edge + DNS | `cloudfront list-distributions`, `route53 list-hosted-zones`, `acm list-certificates`, `wafv2 list-web-acls` |60| Detective | `cloudtrail describe-trails`, `config describe-configuration-recorders`, `guardduty list-detectors`, `securityhub get-enabled-standards`, `inspector2`, `macie2` |61| Secrets / keys | `secretsmanager list-secrets`, `kms list-keys / get-key-rotation-status` |62| Containers | `eks list-clusters`, `ecs list-clusters` |63| Lambda | `lambda list-functions / get-function-configuration / get-function-url-config` |64| Cost | `ce get-cost-and-usage`, `budgets describe-budgets`, `ce get-anomalies` |6566**Use an AWS MCP** (if `AWSSEC_MCP_PRESENT=1`) for typed inventory reads (S3, DynamoDB, Lambda, RDS lists). Fall back to `aws_run` otherwise.6768**Use this skill** for: digest views (counts + derived risk signals from many CLI calls), mutating actions (`fix`, `panic` — read-first, idempotent, recorded), offline tools (`detect`, `fit-matrix`, `stack-docs`, `score`).6970If neither CLI nor MCP is available, the skill cannot operate. `doctor` is the pre-flight.7172## Orchestration7374For any AWS request:75761. **Classify intent** yourself (audit, migrate, scale-plan, diagnose, incident).772. **Call the smallest set of tools** that gives you what you need. Start with `state account` + `state iam` + `state s3` digests in parallel; add `state cloudtrail`, `state guardduty`, `state securityhub`, `state config` for an audit. Run independent calls in parallel (single message, multiple Bash calls).783. **Lazy-load references** that match findings. Read `references/30-recipes.md` for orchestration detail. Read `references/<NN>-<area>.md` when an area surfaces an issue. Read `references/15-stack-best-practices/<stack>.md` for stack-tailored guidance.794. **Synthesize the report.** Group findings by area; mark `OK / WARN / FAIL`; surface paid-only items as `[locked: <tier>+]` with one-line value statements (from `references/10-plan-tier-matrix.md`).805. **For project file changes** (`fix project`, `fix gha`), the tool emits proposed contents + unified diff to stdout:8182 ```83 === FILE: <relative-path> ===84 === DIFF ===85 <unified diff>86 === CONTENT ===87 <full proposed file body>88 === END ===89 ```9091 Apply with `Edit` or `Write` after user confirms. The skill never writes inside the user's project. For Lambda secret values, the tool emits `aws secretsmanager create-secret` invocations — never type secret values yourself.9293## Recipes9495Canonical recipes live in `references/30-recipes.md`. Read it for any specific recipe (audit, migrate, scaling, diagnose, attack response). Recipes assume the digest-by-default contract above.9697## Guardrails9899- **Refuses** when `AWS_ACCESS_KEY_ID` starts with `AKIA` and at least one SSO profile exists in `~/.aws/config`. Redirect: `AWS_PROFILE=<sso-profile> aws sso login`.100- **Hard-refuses** when no AWS credentials are visible (env, profile, instance role). `doctor` prints remediation.101- `fix` is idempotent — no-op when state matches target. Each `*_fix` reads first, compares, mutates only on drift.102- `panic` records each action to `.state/panic-<ts>.json`; `panic restore` rolls back.103- Honest verdicts: PHP/Rails/Django stacks usually get `partial` or `proxy-only` — surface that first.104- The skill **never** lowers posture on `*_fix` (never opens a Public Access Block, never decreases CloudTrail multi-region coverage).
naieum