deploy
Skill1 starsChanged 4 months agoMITView on GitHub
Drive a deployment of this stack end-to-end — local box or AWS VM, behind Cloudflare Tunnel with Google OAuth. Use when the user says "deploy this", "set this up", "make this public", "put this on AWS", or asks to stand up / go live with the MCP tools. Claude does every step it can; the user only decides, approves spend, and does the account/browser steps.
- Reads credentials
161 other projects use a version of this.
Adopted by
Most recently taken up by
- krystaleicher/claude-skillstheir own version, first seen Nov 16, 2026
- KazKato91/roomi-claude-plugintheir own version, first seen Nov 16, 2026
- AlejandroGodius/claude-skillstheir own version, first seen Nov 16, 2026
- reloru/new-relotheir own version, first seen Nov 16, 2026
- maxwell2732/claudecode-research-harness-workflowtheir own version, first seen Nov 15, 2026
- patrickjaja/claude-desktop-extratheir own version, first seen Nov 15, 2026
Used with
Projects with it also use pm-discovery, qa-phase and 10 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
ww-w-aiSkillpm-discovery
PM Agent Team - Automated product discovery, strategy, and PRD generation. Runs PM analysis workflow to produce comprehensive PRD before PDCA Plan phase.
Also in 3 projects that use this
ww-w-aiSkillqa-phase
Run QA tests at various levels (L1-L5) for features under PDCA management. Integrates with MCP tools to execute structured test plans.
Also in 3 projects that use this
- +10 more from ww-w-ai
Other versions161 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 161
- ok-oldking/ok-wuthering-waves★ 8.6ksame name
- memovai/mimiclaw★ 5.8ksame name
- KingKongRobotics/jumper★ 3.1ksame name
- jhd3197/ServerKit★ 1.8ksame name
- pedrohcgs/claude-code-my-workflow★ 1.7ksame name
- mattmireles/gemma-tuner-multimodal★ 1.5ksame name
- microsoft/power-platform-skills★ 1.2ksame name
- jazzyalex/agent-sessions★ 948same name
- awslabs/agent-plugins★ 943same name
- patrickjaja/claude-desktop-extra★ 761same name
151 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
The file
We split this skill into the rules it gives an agent
11 rules. 0 are common, 6 are distinctive, and 5 are very specific to this project.
0
Common
Most of the 143 edited versions we read keep these.
6
Distinctive
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Very specific: line 54
Ask the user, in a single round, the decisions docs/DEPLOY.md lists:
- Distinctive: line 104
Do not ask for per-tool API keys or the Google OAuth client here.
- Distinctive: line 110
When the runbook hits a step only the user can do (Cloudflare dashboard, Google Cloud console, buying a domain):
- Distinctive: line 115
If the user is stuck, offer to help directly: with the Claude Chrome extension connected, you can navigate the browser with them and do or …
- Very specific: line 129
When you need a secret, lead with the route that keeps it out of the chat and make that the default ask: on the user's machine, the !
- Distinctive: line 141
Never commit a secret. Never print one in logs or summaries.
- Very specific: line 145
Before any command that creates billable resources (on AWS: pulumi up in deploy/aws), state plainly what will be created and the recurring …
- Distinctive: line 152
Follow the chosen runbook top to bottom.
- Very specific: line 154
After each step, run the runbook's verification (or the obvious one: pulumi stack output, docker compose ps, the curl ...
- Very specific: line 190
Run the verification curl for every deployed tool and show the WWW-Authenticate proof.
- Distinctive: line 199
If a runbook step fails, is out of order, or misled you, note it, work around it, and keep going.
“Common” is measured: kept by at least 6 in 10 of the 143 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Deploy this stack
You are driving the deployment. The user is the approver and does only the steps that genuinely require a human (owning accounts, minting tokens, approving spend). Everything else — installing CLIs, Pulumi stacks, watching boot logs, writing env files, verification — is your job. Do not hand the user a list of commands to run; run them.
Source of truth: docs/DEPLOY.md (the chooser) and the runbook it picks —
docs/deploy/local.md or docs/deploy/aws.md. Read the chooser and the
chosen runbook fully before acting. This skill is protocol (who does what,
when to stop); the runbooks own the actual commands. If this file and a
runbook disagree on a command, the runbook wins.
Execution environment (decide silently; never ask)
Notice where you are running before Phase 0. The user is never asked about any of this — it changes what you do, not what they decide.
- User's machine (Claude Code in a terminal): everything below applies
as written, including the
!prefix for user-run commands. - Hosted sandbox (Cowork, Claude Code on the web): commands run in an
ephemeral VM that is not the user's machine and does not survive the
session.
- Local path: stop. The stack must run on the user's own box and you cannot bring it up there from here. Say so, point them at Claude Code in a terminal on the target machine, and offer the AWS path as the alternative you can drive.
- AWS path: fine — the VM is created in AWS, not where you're standing
— but the Pulumi state rule below is mandatory, and the
!prefix does not exist here (the secrets protocol's "user runs it themselves" route becomes "user puts the value in the file/console themselves").
Pulumi state — pick silently, never ask the user:
pulumi whoamisucceeds → a backend is already configured; use it, don't switch it.- User's machine, no backend →
pulumi login --local. - Sandbox, AWS path → state must outlive the sandbox: create a small S3
bucket with the AWS credentials already in hand and
pulumi login s3://<bucket>before creating any resources. Never sign the user up for Pulumi Cloud to solve this.
Both stacks (deploy/cloudflare and deploy/aws) must live on the same
backend — the AWS stack reads the ingress stack via a StackReference.
Phase 0 — decisions (one question round)
Ask the user, in a single round, the decisions docs/DEPLOY.md lists:
- Path — local box or AWS.
- Tools — the shipped set is the README's "The tools" table (source of
truth:
tools/*/deploy.jsonmanifests — summary, secrets, notes like image size). Mind the manifests' sizing/dependency notes when they pick. - Guardrail — a decision only on the LOCAL path: llamafirewall (local
model, the default — needs a free HF token + model-access grant) or
unscreened (
GUARDRAIL_ENABLED=0, say plainly what that means). The AWS path always provisions Bedrock; nothing to ask. Only surface this if they ask or their tool picks include an untrusted-output one. - Approval channel — the approval layer is on by default: the
server-side version of Claude's per-tool "always allow / needs approval /
blocked", enforced here because the desktop toggle is sticky (approve
once and it sticks across every chat) and doesn't reliably apply to
custom connectors. Every tool STARTS
always_allow; the user gates or blocks individual tools at runtime via the gatekeeper's panel orset_gating(docs/GATEKEEPER.md) — that is not a deploy-time decision. The deploy-time decision is the CHANNEL that will deliver Approve/Deny cards for whatever they later gate (and for the two code-pinned gatekeeper tools): Slack, Discord, or Telegram (APPROVAL_PROVIDER) — a free app, ~5-10 browser minutes, done while other steps run. Without a channel, gated calls report the approval as undeliverable (never silently run).MCP_REQUIRE_APPROVAL=0in the root.envremoves the layer entirely — say plainly what that means. Whichever provider they pick, explain the control is HUMAN-in-the-loop, so it must live on a platform (or at least an account) the agent does not operate — if the agent's own tools can read the card and press its buttons, the gate approves itself. Steer them to whichever of Slack/Discord/Telegram their agent doesn't touch (most sharply: don't run the Telegram provider in a chat the telegram tool's account can see).
Phase 1 — preflight
You check and install tooling. Detect what's already present before
installing anything (pulumi version, aws sts get-caller-identity,
docker --version, session-manager-plugin — per the runbook's prereq
list). Install what's missing yourself; tell the user what you installed.
Then give the user ONE short checklist — only the deployment-level items a human must produce, nothing else:
- A domain on Cloudflare (free plan is fine) and, from its zone Overview page, the Zone ID and Account ID.
- A Cloudflare API token with
Cloudflare Tunnel:Edit+DNS:Editon that zone. - AWS path only: working AWS credentials on this machine. If they need to
log in interactively, have them run it themselves with the
!prefix (e.g.! aws configureor! aws sso login).
Do not ask for per-tool API keys or the Google OAuth client here. Those belong to later runbook steps and you only collect them for the tools the user actually picked, when the runbook reaches them.
Human-step protocol
When the runbook hits a step only the user can do (Cloudflare dashboard, Google Cloud console, buying a domain):
- Give the exact click-path from the runbook and say precisely which values you need back (names, not secrets, in chat — see below).
- If the user is stuck, offer to help directly: with the Claude Chrome extension connected, you can navigate the browser with them and do or verify the clicks. Mention this option; don't assume it's available.
- Where possible, parallelize: e.g. the Google OAuth client (AWS runbook step 4) can be created while the VM's first boot is still building images.
Secrets protocol
This governs how you ask, not just how you react to a paste.
- Split every request into public identifiers vs secrets. Client IDs, account/zone IDs, emails, domains, instance IDs are fine in chat — ask for those plainly. Client secrets, API keys, bearer tokens, session strings are not; a Telegram session string is full account access.
- When you need a secret, lead with the route that keeps it out of the
chat and make that the default ask: on the user's machine, the
!prefix (! pulumi config set --secret ...) or the user editing the.envin their own editor while you wait. Never phrase it as "send me the secret". - But do not refuse pasted secrets. Some users will want to do
everything in the chat. Say once that pasting a secret into the
conversation is not recommended (it persists in the conversation history)
— then, if they paste it anyway, use it: put it exactly where the runbook
says (
.env,pulumi config set --secret), never repeat it back, never echo it in command output, and move on. It works; it's just not the recommended route. - Never commit a secret. Never print one in logs or summaries.
Spend gate
Before any command that creates billable resources (on AWS: pulumi up in
deploy/aws), state plainly what will be created and the recurring cost from
the runbook, and get an explicit yes. The Cloudflare ingress stack and
pulumi preview need no gate — they're free/read-only.
Execution
- Follow the chosen runbook top to bottom. Announce each numbered step as you start it.
- After each step, run the runbook's verification (or the obvious one:
pulumi stack output,docker compose ps, thecurl ... WWW-Authenticatecheck) and show the user the evidence before moving on. Never advance past a failed check — use the runbook's Troubleshooting section first. - On AWS, first boot takes several minutes: watch it for real over the SSM
session (
cloud-init-output.log, thendocker compose ps) rather than assuming it worked. - Per-tool secrets step: read
tools/<tool>/env.examplefor each chosen tool and collect only those values, following the secrets protocol above. - Approval-channel step (unless they opted the layer off in Phase 0):
drive the runbook's Approvals section. All three walkthroughs live in
security/approval/service/env.example:- Slack — create app →
chat:writescope → install → signing secret → private channel + invite bot → Interactivity Request URLhttps://approval.<domain>/slack/interact. - Discord — create app → bot token → public key → channel ID → Interactions
Endpoint URL
https://approval.<domain>/discord/interact, set LAST (the sidecar must be live when Discord validates it); setAPPROVAL_PROVIDER=discord. - Telegram — @BotFather bot token → a private chat/group the bot is in
(numeric chat id via
getUpdates) → a long random webhook secret → register the webhook LAST withsetWebhook(urlhttps://approval.<domain>/telegram/interact,secret_token=<the secret>,allowed_updates=["callback_query"]); setAPPROVAL_PROVIDER=telegram. Telegram doesn't sign the body — the secret token IS the auth, so keep it long and random. Collect that provider's values per the secrets protocol intosecurity/approval/service/.env, and verify: the sidecar's/healthz(compose-internal, e.g.docker compose exec approval python -c ...) must show"channel": "configured", and a test POST to its/gatemust return"notified": true— that proves a card actually landed in their channel; have the user Deny it. If they opted the layer off instead, setMCP_REQUIRE_APPROVAL=0in the root.envbeforeup.
- Slack — create app →
Finish
- Run the verification curl for every deployed tool and show the
WWW-Authenticateproof. - Give the user the connector URLs and walk them through Claude → Settings → Connectors → Add custom connector → Google login.
- Tell them the two day-2 commands that matter: how to update the deployment
and how to tear it down (
Day 2section of the runbook).
If the docs are wrong
If a runbook step fails, is out of order, or misled you, note it, work around it, and keep going. At the end, list every doc problem you hit so it can be fixed — the runbooks are meant to survive exactly this kind of run.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

