Run any Skill in Manus with one click

install-promptworld

Stars0

Forks0

UpdatedJune 8, 2026 at 20:58

WHAT: the deploy WIZARD for PromptWorld — YOU (the agent) build the docker image and boot a container running the whole system (the Engineer-CEO that has every skill + a native subagent per component type, the 7 PromptGym specialist agents, the assistant-ui chat + Monaco file explorer + an embedded terminal), find the Claude credentials on THIS machine, copy them into the container at boot, and then just hand the user a URL. WHEN: when the user wants to install / run / launch / boot / spin up / deploy PromptWorld, or says 'set up promptworld', 'start promptworld', 'boot the promptworld container' (any of).

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Run Skill in Manus

Source

sancovp

sancovp/promptworld

View GitHub Repository View Creator Repositories

Download

Run Skill in Manus

Related occupationsSOC

Based on SOC occupation classification

Software DevelopersComputer and Mathematical Occupations·SOC 15-1252

SKILL.md

readonly

name

install-promptworld

description

install-promptworld — the agent's deploy how-to

YOU (the agent reading this) do the ENTIRE deploy. The user fills in NOTHING and runs NOTHING — when you finish, you hand them ONE URL to open. There are no placeholders, no bind mounts, no -e flags, no human-typed docker commands. You build the image, find the Claude creds on THIS machine, and boot via create → docker cp creds → start (which works even when docker is the host daemon via a mounted socket, because it needs no -v bind mount of a local path).

What gets deployed

A single promptworld:latest container, on port 3858, serving:

the Engineer-CEO (claude -p) — has every skill + a native subagent per component type (skill / mcp / harness / operating_system / prompt / team / workflow);
the 7 PromptGym specialist AIOS agents (each scoped to its own promptgym/<type>/ directory);
the assistant-ui chat SPA + the Monaco file explorer over /workspace + an embedded xterm terminal (/ws/terminal → a PTY shell as ceo);
doc-mirror installed for every agent. Journal is file-only (DOCMIRROR_CARTON=0) — no DB.

The image ships UNAUTHENTICATED (no creds baked). Auth arrives at boot, two ways — you handle #1, the user can do #2 if you have no creds:

You cp the creds in (the normal path): you find the Claude OAuth creds on this machine and docker cp them into the container.
The user logs in live in the embedded terminal (claude auth login) — the fallback when this machine has no creds file (e.g. macOS, where creds live in Keychain).

Do this (you, the agent)

Locate the deploy dir: <plugin-or-repo>/application/promptworld/deploy/. Confirm deploy/build.sh + deploy/Dockerfile exist.
Build the image (stages promptworld + the doc-mirror plugin + cave + sdna, then docker build; a few minutes the first time). NO creds go into the image:
```
bash <...>/application/promptworld/deploy/build.sh build
```
Find the Claude creds on THIS machine. The OAuth credentials file is at ${HOME}/.claude/.credentials.json (on this dev box that resolves to /home/GOD/.claude/.credentials.json). Check it exists:
```
ls -l "${HOME}/.claude/.credentials.json"
```
- If it exists → you'll cp it in (step 4 does this automatically).
- If it does NOT exist (e.g. macOS Keychain) → the container still boots; tell the user to sign in via the embedded terminal (claude auth login).
Boot it — create → docker cp creds → start (this is exactly what build.sh boot does: it docker creates with -p 3858:3858 and NO bind mount, docker cps ${HOME}/.claude/.credentials.json to the container's /creds/.credentials.json, then docker starts; the entrypoint copies /creds into the ceo user's config + chowns it, so it authenticates for any uid):
```
bash <...>/application/promptworld/deploy/build.sh boot
```
Verify through the real surface (don't trust "it started"):
```
bash <...>/application/promptworld/deploy/build.sh e2e
```
It must show a real CEO /api/chat reply AND the skill-specialist subagent writing a file inside promptgym/skill/. (e2e uses docker exec, so it works regardless of host networking.)
Hand the user the URL. Tell them: open http://localhost:3858 — chat with the Engineer-CEO, watch it dispatch specialists, edit files in the Monaco panel. If step 3 found no creds, add: "open the Terminal panel at the bottom and run claude auth login to sign in on your own subscription."

Done when

build.sh e2e shows a real CEO /api/chat reply AND the specialist-subagent artifact in-container. The build/boot are idempotent (safe to re-run). On failure, read docker logs promptworld-v1, fix, re-run.

More from this repository

same repository

compile-a-world

sancovp/promptworld

WHAT: PromptWorld's meta-compiler skill — the architecture patterns + worked examples for building a *World (a sim app = a team-leader + a specialist set + a domain + a driver) such as JobWorld, GameWorld, HealthWorld, or PromptWorld itself, plus how to reuse one pattern, combine them into a new World/app, and compose+publish it over the nomicon. WHEN: when the user (or the Archwizard) says 'compile/build/make a World', 'make a new jobworld/gameworld/healthworld', 'what World patterns are there / how is a World built', 'turn this into an app / a sim', 'meta-compiler', 'world compiler', or wants to scaffold a CAVE-fork sim app or fold one into the nomicon (any of).

2026-07-100

cave-teams

sancovp/promptworld

WHAT: Spin up a programmatic agent TEAM on the fly from a spec — mixed Claude-Code (claude -p, subscription) leaders + cheap MiniMax workers — with a LIVE WEB GALLERY that shows each agent in its own window as it works. cave-teams is a STANDALONE library (claude -p + MiniMax + files; no running CAVE required); it optionally streams its event seam to a gallery, a CAVE Channel, or Discord. WHEN: when you need to launch a sub-team to do parallel/heterogeneous work and watch it live — equip when spawning a team, unequip when done. Triggers: spawn a team, run a team, cave-teams, agent squadron, watch the agents, team gallery.

2026-07-100

ingest-into-tome-app

sancovp/promptworld

WHAT: the PromptWorld app-rung — a single flow that ingests an existing Claude-Code system directory (a dir of AIOSes / skills / rules / CLAUDE.mds), converts it into its own {name}-nomicon (the author's SkillTome — the framework-of-frameworks holder/search/router), and agentifies that tome into a standalone sidecar PromptWorld app (a CEO that carries the tome as its world-model, with the ingested AIOS dirs as its specialist agents). It WIRES the already-existing pieces (ingest = copy into the workspace; convert = the skill2framework→fold-into-tome pipeline; agentify = PromptWorld scoped to the tome) into one drop-dir→get-app flow. WHEN: when the user says 'ingest this claude-code dir', 'turn this system into an app / a webapp', 'make a {name}-nomicon from this', 'convert this dir into its own app', 'app-rung', or wants a Claude-Code directory packaged as a runnable sidecar PromptWorld with an AIOS (any of).

2026-07-100

wield-the-tome

sancovp/promptworld

WHAT: how a PromptWorld wizard WIELDS the SkillTome — consult the author's tome (which framework fits this task?), route through a framework's volume row, run the skill2framework chapter chain on something you built, and FOLD the finished framework into a tome (a code op, skilltree fold). WHEN: when a task smells like an existing framework might cover it; when a build is finished and should become a chapter; when the user mentions the tome, the nomicon, folding, a volume, or 'which of my frameworks fits this' (any of).

2026-07-100

name

install-promptworld

description

install-promptworld — the agent's deploy how-to

What gets deployed

A single promptworld:latest container, on port 3858, serving:

the Engineer-CEO (claude -p) — has every skill + a native subagent per component type (skill / mcp / harness / operating_system / prompt / team / workflow);
the 7 PromptGym specialist AIOS agents (each scoped to its own promptgym/<type>/ directory);
the assistant-ui chat SPA + the Monaco file explorer over /workspace + an embedded xterm terminal (/ws/terminal → a PTY shell as ceo);
doc-mirror installed for every agent. Journal is file-only (DOCMIRROR_CARTON=0) — no DB.

The image ships UNAUTHENTICATED (no creds baked). Auth arrives at boot, two ways — you handle #1, the user can do #2 if you have no creds:

You cp the creds in (the normal path): you find the Claude OAuth creds on this machine and docker cp them into the container.
The user logs in live in the embedded terminal (claude auth login) — the fallback when this machine has no creds file (e.g. macOS, where creds live in Keychain).

Do this (you, the agent)

Locate the deploy dir: <plugin-or-repo>/application/promptworld/deploy/. Confirm deploy/build.sh + deploy/Dockerfile exist.
Build the image (stages promptworld + the doc-mirror plugin + cave + sdna, then docker build; a few minutes the first time). NO creds go into the image:
```
bash <...>/application/promptworld/deploy/build.sh build
```
Find the Claude creds on THIS machine. The OAuth credentials file is at ${HOME}/.claude/.credentials.json (on this dev box that resolves to /home/GOD/.claude/.credentials.json). Check it exists:
```
ls -l "${HOME}/.claude/.credentials.json"
```
- If it exists → you'll cp it in (step 4 does this automatically).
- If it does NOT exist (e.g. macOS Keychain) → the container still boots; tell the user to sign in via the embedded terminal (claude auth login).
Boot it — create → docker cp creds → start (this is exactly what build.sh boot does: it docker creates with -p 3858:3858 and NO bind mount, docker cps ${HOME}/.claude/.credentials.json to the container's /creds/.credentials.json, then docker starts; the entrypoint copies /creds into the ceo user's config + chowns it, so it authenticates for any uid):
```
bash <...>/application/promptworld/deploy/build.sh boot
```
Verify through the real surface (don't trust "it started"):
```
bash <...>/application/promptworld/deploy/build.sh e2e
```
It must show a real CEO /api/chat reply AND the skill-specialist subagent writing a file inside promptgym/skill/. (e2e uses docker exec, so it works regardless of host networking.)
Hand the user the URL. Tell them: open http://localhost:3858 — chat with the Engineer-CEO, watch it dispatch specialists, edit files in the Monaco panel. If step 3 found no creds, add: "open the Terminal panel at the bottom and run claude auth login to sign in on your own subscription."