guide

name	guide
description	Index of the Prisme.ai plugin skills plus essential Prisme.ai environment context (environments, workspaces, tools, event schema). Use this when starting work on anything Prisme.ai-related and unsure which skill applies, or to discover what is available.
allowed-tools	Read, AskUserQuestion

Prisme.ai skills catalog

This plugin ships a set of skills for Prisme.ai workspace development. They are grouped by domain below — pick the one that matches your task. Skills are namespaced by the plugin: invoke them as /prisme-ai:<name> (e.g. /prisme-ai:app-mcp-implement).

---

App + MCP connectors

The connector workflow: scaffold a new SaaS connector, test it, consolidate tests in a consumer workspace, document it, and keep the whole fleet in sync over time.

Skill	When to use
/prisme-ai:app-mcp-implement	Scaffold a brand-new App + MCP workspace for a third-party SaaS (REST or GraphQL). Produces index.yml, security.yml, .import.yml, helpers, Custom Code, MCP Core, tool/method automations, and pushes to the target env.
/prisme-ai:app-mcp-test	E2E smoke-test the tools exposed by an existing App + MCP workspace. Asks for credentials, lists every MCP tool, executes them one by one, fixes errors until the suite passes. Supports static tokens, Basic, OAuth2 client-credentials, and OAuth2 authorization-code (PKCE).
/prisme-ai:app-mcp-build-consumer	Build or audit the matching *-consumer workspace that consolidates the proven smoke-test coverage into a durable DSUL test suite. Typically chained after /prisme-ai:app-mcp-test.
/prisme-ai:app-mcp-document	Generate the public MDX documentation page for an existing connector in the prismeai/docs repo (Mintlify). Mirrors the layout of existing pages like gryzzly.mdx, data-galaxy.mdx.
/prisme-ai:app-mcp-fleet-sync	Find every connector workspace (by the app-mcp label), diff each against the current /prisme-ai:app-mcp-implement templates + rule checklist, and apply approved fixes. Run when a template trap-fix lands and the deployed fleet drifts.

---

Agents and evaluation

Skill	When to use
/prisme-ai:agent-builder	Scaffold or update an Agent Factory-backed Prisme.ai agent workspace with DSUL templates, deployment automations, tests, and agent instructions.
/prisme-ai:agent-workspace	Create a bootstrap workspace that provisions an Agent Factory agent, attaches Storage-backed sources, seeds Agent Evaluations cases, runs evaluation, and validates the setup through the Agents, Storage, and Evaluation product apps.
/prisme-ai:agent-implement-a2ui	Add A2UI (Agent-to-UI) surfaces to a Prisme.ai MCP workspace so that LLM agents can render interactive UI (cards, forms, tables, action-cards, confirmations…) through MCP tool calls. The host UI reads the tool's __surface payload and renders it from the prisme://blocks/v1 catalog.

---

Embedded React apps

Skill	When to use
/prisme-ai:workspace-page-implement	Edit (or bootstrap) the React app embedded in a Prisme.ai workspace following the starter-spa pattern. Modifies src/ (React) and/or automations/ (DSUL), builds the CJS bundle, and pushes to the workspace.

---

Investigation and validation

Skill	When to use
/prisme-ai:prisme-assistant	Investigate Prisme automations, apps, events, or search Prisme documentation. The expert for event-feed debugging (correlationId, activity feed, execution logs).
/prisme-ai:debug-events	Debug Prisme.ai runtime events, activity feeds, and correlationId traces with focused event-search patterns.
/prisme-ai:dsul-rules	Check current DSUL authoring/deployment rules and edge cases before editing automations or workspace configuration.
/prisme-ai:ticket-validator	After implementing a feature, verify every ticket requirement was met. Run before code review.

---

Quick reference

Situation	Skill
Build a new third-party SaaS connector (App + MCP)	/prisme-ai:app-mcp-implement
Test the tools of an existing connector with real credentials	/prisme-ai:app-mcp-test
Consolidate test coverage into a *-consumer workspace	/prisme-ai:app-mcp-build-consumer
Write or update the public docs of a connector	/prisme-ai:app-mcp-document
Propagate a template fix across all connectors	/prisme-ai:app-mcp-fleet-sync
Scaffold an Agent Factory DSUL workspace	/prisme-ai:agent-builder
Provision + evaluate an Agent Factory agent	/prisme-ai:agent-workspace
Render interactive UI from agent tool calls	/prisme-ai:agent-implement-a2ui
Edit the React frontend of a workspace	/prisme-ai:workspace-page-implement
Debug events / trace an execution	/prisme-ai:prisme-assistant
Configure an MCP environment or token	/prisme-ai:setup
Check DSUL rules before editing automations	/prisme-ai:dsul-rules
Check an implementation against its ticket	/prisme-ai:ticket-validator
I don't know where to start	/prisme-ai:guide (this)

---

Prisme.ai environment context

Authentication

The MCP authenticates with user-created API tokens. When a tool call fails with "No credentials for environment …" (or a 401), do not ask the user to paste their token into the chat — that would send it to the LLM provider.

Instead, relay the CLI command from the error message and ask the user to run it in their own terminal. Preserve it as one shell command; do not insert line breaks inside quoted paths.

node "<plugin>/build/index.js" set-token <environment> --config-dir "<config-dir>"

The error message already contains the exact path and config dir. The command prompts for the token with hidden input, then asks for the Prisme API URL (e.g. https://api.sandbox.prisme.ai/v2). If the user is unsure, tell them to open the Prisme instance in a browser and copy the API base URL from the Network tab. It validates the token and saves it. After it succeeds, just retry the request — the server picks up the new token automatically (no restart). To create the token, the studio page is <studio-origin>/settings/tokens (e.g. https://sandbox.prisme.ai/settings/tokens).

Only if the user explicitly insists on pasting the token in the conversation should you fall back to the set_token tool — and first warn them the token will be transmitted to the LLM provider as part of the chat.

Environment configuration

Environment	Aliases	Default workspace
sandbox	sb	ai-knowledge
prod	production	-

Default behavior: search on the ai-knowledge workspace in the sandbox environment unless specified otherwise.

Workspace parameter rules

Use the correct parameter based on user input:

User provides	Use parameter	Example
Raw ID (e.g., e-sdfwe)	workspaceId	workspaceId: "e-sdfwe"
Named alias (e.g., ai-knowledge)	workspaceName	workspaceName: "ai-knowledge"

Known workspace aliases: ai-knowledge (AIK), ai-store (AIS). Any other workspace has no workspaceName to provide to tools. Never pass a raw ID as workspaceName — it will fail.

Local workspace priority

IMPORTANT: The local codebase may contain Prisme.ai workspace folders. Always search locally first using Read, Glob, Grep before calling remote APIs.

1. Each workspace folder contains automations, pages, and config as YAML
2. Read local YAML files directly — they represent the current working state
3. Check local folders before search_workspaces — if the workspace exists locally, its index.yml contains the id field with the workspace ID; use it with the workspaceId parameter
4. Only call remote APIs for remote-only data (events, execution) or when the workspace is not found locally

Tool usage guide

Tool	Use when
search_events	User mentions: correlationId, activity feed, events, execution history, logs, tracing, debugging, errors/failures
get_automation	Need to inspect a specific automation's YAML
list_automations	Need to see all automations in a workspace
get_prisme_documentation	Need Prisme.ai syntax, patterns, or feature reference
validate_automation	Validate automation YAML for syntax/semantic errors
get_app / list_apps	Working with apps from the marketplace
pull_workspace / push_workspace	Syncing workspace files locally
execute_automation	Testing an automation with a payload
set_token	Register or rotate the API token for an environment

push_workspace must never upload pages/<name>/ subfolders

When a workspace embeds a React app under pages/<name>/ (starter-spa / nested-app convention), that subfolder is a full frontend project (src/, node_modules/, dist/, build tooling), not DSUL. Never push_workspace it:

• node_modules can make the importer fail and leave the workspace locked mid-import.
• The remote keeps app sources at canonical paths that differ from the local nested layout; pushing the local pages/<name>/ can overwrite that structure and make the Studio page disappear.
• push_workspace does not update runtime config.value, so it cannot deploy the bundle pointer.

Any MCP push/sync must exclude every pages/*/ subdirectory. To deploy a workspace React app / SPA bundle, use /prisme-ai:workspace-page-implement; it builds the CJS bundle, uploads it, and patches config.value correctly.

After creating or editing YAML/DSUL automations, always run validate_automation to check for syntax and semantic errors. It is authoritative for DSUL validation — trust it over existing workspace patterns (which may contain legacy mistakes). If validation conflicts with documentation, list issues and report to the human before fixing.

Event search patterns

Common Elasticsearch DSL patterns for search_events:

{"bool": {"filter": [{"term": {"source.correlationId": "uuid-here"}}]}}

{"bool": {"filter": [{"term": {"type": "runtime.automations.executed"}}]}}

{"bool": {"filter": [{"term": {"source.automationSlug": "automation-name"}}]}}

{"bool": {"filter": [{"term": {"type": "error"}}]}}

Sorting: always use @timestamp (not timestamp) for time-based sorting.

Event schema quick reference

Field	Description
type	Event category (e.g., runtime.automations.executed, error)
source.correlationId	Groups related events from same operation
source.automationSlug	Automation name
source.workspaceId	Workspace identifier
source.userId	User who triggered event
payload	Event-specific data
@timestamp	Event timestamp (use for sorting)

Common event types:

• runtime.automations.executed - Automation completed
• runtime.fetch.failed - HTTP request failed
• error - Generic error
• workspaces.automations.updated - Automation changed

Products quick reference

Product	Nickname	Purpose
AI Knowledge	AIK	RAG, knowledge bases, LLM middleware
AI Store	AIS	Chat interface, agent marketplace
Builder	-	Custom workflows, automations, pages
SecureChat	-	Enterprise conversational interface
Collection	-	Tabular data with AI querying
Governance	-	Platform admin, RBAC
Insights	-	Conversation analytics

AI Knowledge query flow

Entry point is genericQuery, which orchestrates:

1. Config phase: rate limits, model specs, attachments
2. Prompt building: system prompt, history, context
3. Tool loop: callLLMWithTools handles tool calling
4. Post-processing: suggestions, source filtering

Key automations: query, genericQuery, buildPrompt, LLMCompletion, callLLMWithTools, executeTool

AI Store architecture

• Each "agent" in AI Store = AI Knowledge project
• Uses Knowledge Client app to interface with AIK
• Conversations stored via Conversations Service App
• SSE streaming for real-time responses

Agent URLs

URLs to access agents in the Prisme.ai interface follow this pattern:

https://{environment}.prisme.ai/fr/product/ai-knowledge/documents?id={projectId}

Environment	Base URL
sandbox	https://sandbox.prisme.ai
prod	https://prisme.ai

Finding the project ID: the id parameter is found in the workspace's imports/Knowledge Client.yml file.

Usage: when working on a workspace with a Knowledge Client, provide the full clickable URL to the user so they can easily test the agent in their browser.

Isolate workspace dev in a worktree

When the user starts a development task on a workspace (any request to implement, edit, or build a feature/fix on a workspace — i.e. before writing or changing any DSUL/automation/page), spontaneously propose, as the first thing, to do the implementation in a dedicated git worktree on a new branch.

• Make this a short one-line offer, e.g. "Je crée un worktree + branche workspace/<slug>-<short-desc> pour isoler ce dev ?" — derive the branch name from the workspace slug and a short kebab-case description of the task.
• If the user accepts, create the worktree (use the worktree isolation mechanism), branch from sandbox, and do all edits there.
• If the user declines or says to proceed in place, work in the current directory without asking again for that task.
• Skip the offer for pure read-only work (diagnostics, event tracing, questions) — it only applies when actual file changes are coming.

Development steps to edit a workspace

• Check if there is a local version of the workspace in ./
• Read the documentation of the relevant products and automations before anything
• Edit the automations locally
• Once completed, run validate_automation to check for common mistakes
• Execute a sub-task to review the changed code and give a list of eventual issues rated MAJOR | NEED_HUMAN, formatted with emoji in a bulleted list:
  • 🔴 Solve the MAJOR issues
  • 🟠 Ask the human for the NEED_HUMAN issues

If you encounter any issue (e.g. the Prisme.ai API returns an error), use the report_issue_or_feedback tool so we can enhance the tooling.