
AWS Startup Advisor
PluginAWS 官方为初创公司出的插件:解答 AWS Activate 云额度的申请条件和 60 多项创业优惠,按融资阶段、团队规模和资金能撑多久给出架构与成本建议,也能规划从 GCP、Azure、Heroku 迁移到 AWS。需安装 Claude Code;查额度与优惠不需要 AWS 账号,迁移和搭建项目要配合代码仓库,成本估算仅供参考。
The Migration DSL — plugin documentation
This folder documents the phase DSL that the migration skills are
built on: a declarative frontmatter grammar an LLM interprets at runtime to drive
a multi-phase migration, with a static validator (mise run lint:frontmatter) that
checks the structure before anything runs.
If you are about to author a new migration skill (or extend an existing one) and you want your agent to understand the grammar, point it at this directory. The docs are written to be read top-to-bottom by a human and loaded wholesale by an agent.
Why this exists
A migration ("move a Heroku app to AWS", "move a GCP project to AWS") is a long, multi-step process: discover what exists, clarify intent, design the target, estimate cost, generate artifacts, collect feedback. Two failure modes dominate when you express that as a plain prose skill:
- Silent drift between phases. A phase claims to produce an artifact the next phase needs, but the wiring is only in prose — nothing catches a rename, a typo, or a hollow "produces" list until a run fails halfway through.
- Orchestration prose the model half-follows. Hardcoded phase-order tables, "set status to in_progress" recipes, and gate-rule restatements pile up in every phase file, drift out of sync with each other, and add load cost without adding behavior.
The DSL answers both by splitting each phase file into two layers:
- Structure → frontmatter (checkable). The phase's identity, ordering, inputs,
outputs, gates, and data dependencies live in a closed-vocabulary YAML block. A
zero-dependency validator parses every phase/fragment/assembler file and fails the
build on a broken reference, an unknown key, a dangling phase edge, a hollow
_produces, an ambiguous artifact creator, and more. - Judgment → prose (interpreted). The actual work — how to parse Terraform, how to size a database, how to word a question — stays in markdown the LLM reads and executes. The grammar deliberately does not try to make judgment checkable.
The dividing line is the whole design: structure is checkable; judgment is the LLM's. See 01-concepts.md for the full rationale.
What it is (and is NOT)
- It is a declarative grammar the LLM interprets. The markdown is the program; the LLM is the interpreter. There is no engine and no server — nothing executes the phase files except a language model reading them.
- It is statically validated.
mise run lint:frontmatterruns a typed parser + structural checks over a skill's phase files. Green means the structure is sound (closed vocab, resolved references, a consistent phase chain, single-creator artifacts). It does not mean the prose is correct — opaque conditions (_when) and judgment assertions (_assert) are bound but never evaluated by CI. - It is not the MCP engine. A separate effort explores deterministic runtime tools (provision, dump/restore, verify). That is execution; this DSL is planning. Keep the two apart: the DSL plans a migration; it does not run one.
Getting started
- Read 01-concepts.md — the mental model (10 min). Without it the grammar looks arbitrary.
- Skim 02-grammar-reference.md — every frontmatter key, its shape, and a real example. This is the lookup you return to.
- Follow 03-authoring-guide.md — build a phase (its
orchestrator + a fragment + the assembler), wire it into the backbone, and get to
a green
lint:frontmatter. - Keep 04-validator-checks.md open while authoring — when a check fails, it tells you what the check enforces and how to fix it.
- Read 05-exec-agent-dispatch.md for how a phase can run
its work in an isolated sub-agent (the
_execexecution mode) — the mechanism, the generic worker, and the platform caveats.
The canonical sources (what these docs answer to)
These docs are a teaching layer. The authorities they describe — and that win on any disagreement — are all in the plugin:
| Source | Role |
|---|---|
skills/shared/dsl/INTERPRETER.md | the runtime semantics — what the LLM does with each key (the spec) |
tools/frontmatter-validator/types.ts | the shape of every frontmatter key (the typed model the parser binds to) |
tools/frontmatter-validator/parse.ts | the closed vocabulary (which keys/verbs/actions are legal) |
tools/frontmatter-validator/check.ts | the structural checks the build enforces |
skills/heroku-to-aws/ | the living example — a complete 6-phase skill authored to this grammar |
If a doc here and one of those sources disagree, the source is right — please open a fix to the doc.
The living example
skills/heroku-to-aws/ is a full migration skill built on this DSL: six phases
(discover → clarify → design → estimate → generate, plus an off-backbone feedback
checkpoint). Every construct in these docs is used there. When an explanation is
abstract, go read the corresponding phase file — the examples in these docs are
pulled from it verbatim.
相關資源

Agent Plugins for AWS
PluginAWS Labs 官方插件市场:含 deploy-on-aws(架构推荐、成本估算与 IaC 部署)、aws-serverless、aws-amplify、databases-on-aws、sagemaker-ai 等插件。AWS 建议新项目优先用 Agent Toolkit for AWS。
awslabs/agent-plugins924
Agent Skills for AWS
SkillAWS 官方技能合集(Agent Toolkit for AWS):覆盖 IAM、CDK/CloudFormation、Serverless、容器、可观测性、成本管理与 Bedrock,另有按服务划分的专项技能。执行 AWS 操作需本地配置凭据。
aws/agent-toolkit-for-aws2.9k
Agent Toolkit for AWS
PluginAWS 官方正式发布的 agent 工具包与插件市场:提供 aws-core、aws-agents、aws-data-analytics 等插件,打包 AWS MCP Server 与经过评测的技能,覆盖 IaC、Serverless、容器与账单。调用 AWS API 需配置凭证。
aws/agent-toolkit-for-aws2.9k
Pulumi
CLI用 TypeScript、Python、Go 等通用编程语言定义云基础设施,pulumi up 即可预览并部署到 AWS、Azure、GCP、Kubernetes 等 300 多个提供方。Pulumi 官方维护,开源。
pulumi/pulumi25.8k
AWS CLI
CLIAWS 官方统一命令行,用一套命令管理 EC2、S3、IAM 等各项 AWS 服务,支持 --query(JMESPath)过滤与 JSON 输出,便于 agent 查询和变更云资源。使用前需用 aws configure 配置访问凭证。
aws/aws-cli17.3k
render-deploy
Skill分析代码库后生成 render.yaml Blueprint,或经 Render MCP 直接创建服务,把应用部署到 Render 云平台。Render 官方维护,OpenAI 官方精选技能也收录了它;需要 Render 账号,代码须推送到 GitHub/GitLab/Bitbucket。
render-oss/skills88