onboarding（入门）

LaunchDarkly SDK Onboarding

Orchestrates LaunchDarkly setup in an existing codebase: on kickoff, show a roadmap in chat (see Kickoff: onboarding roadmap); Step 0 writes a living onboarding log so a new session or the user can resume; then explore the project, detect the agent, install flag-management skills, configure the LaunchDarkly MCP server early, install and initialize the SDK (sdk-install and nested detect/plan/apply), and create a first feature flag. Nested skills: mcp-configure, sdk-install, first-flag.

Prerequisites

• npx: Available on PATH when using npx skills add to install companion skills (see Step 3).
• LaunchDarkly account (deferred -- inferred, not asked upfront): Do NOT ask whether the user has a LaunchDarkly account at the start. Instead, let the workflow reveal account status naturally:

- Steps 0-3 (log, explore, detect, install skills) do not require an account. Run them first. - Step 4 (MCP): Present the MCP install link. If the user completes OAuth successfully, they have an account -- confirmed, no question needed. - Step 5 (SDK keys): If MCP is not configured (or the user declined it), account status becomes relevant at D7 when the user needs to provide keys. If they cannot provide keys, offer the signup link: https://app.launchdarkly.com/signup?source=agent. - This eliminates the upfront "Do you have an account?" question and lets the workflow itself surface whether the user needs to sign up.

• Keys and tokens (defer until needed): Collect these only when the path requires them.

- Step 4 -- MCP: Hosted MCP uses OAuth; no API token or SDK key needed to configure it. Local npx MCP (federal/EU, etc.): API access token per mcp-configure and MCP Config Templates. - Step 5 -- SDK: SDK keys / client-side ID / mobile key when wiring env in Apply code changes, after the integration plan is confirmed. ldcli / REST for discovery: use ldcli login or an access token when you first run those commands, not at hello. - Key type must match the integration: server-side SDK -> SDK key; browser/client-side SDK -> Client-side ID; mobile -> Mobile key. Env variable names and bundler rules: Apply code changes.

MCP (preferred): Complete Step 4 via mcp-configure/SKILL.md before SDK work when possible. If MCP is unavailable or the user opts out, use ldcli / REST fallbacks described in that skill (including MCP Config Templates for local npx fallback when hosted MCP does not apply) -- onboarding must still be completable.

Optional MCP tools (when configured):

• get-environments -- list environments for a project; the response includes SDK keys, client-side IDs, and mobile keys per environment. Use this as the single source for all key types -- do not make separate requests for individual keys.
• create-feature-flag -- create the boolean flag for Step 6: Create Your First Feature Flag.
• update-feature-flag -- toggle or patch flag configuration during Step 6; see Create first feature flag for ldcli/API fallbacks.

Other MCP tools you may use if present (not required): list-feature-flags, get-feature-flag, get-flag-status-across-environments.

Agent Behavior Directives

Progress Tracking

The roadmap (Steps 0-6 + follow-through) MUST be tracked using your agent's native task-tracking tool in addition to the onboarding log file.

• Cursor: Use TodoWrite to create a todo for each step before beginning work. Update status as each step completes.
• Claude Code: Use TaskCreate to create a task for each step (or TodoWrite if native tasks are unavailable).
• Other agents: If your agent provides a native task list or progress tracking tool, use it. If not, present a numbered checklist in chat and update it after each step.

Do NOT work through steps mentally or rely solely on the LAUNCHDARKLY_ONBOARDING.md log for in-session tracking.

Decision Points

When a step requires user input to determine branching, you MUST collect the answer by calling a tool — do NOT write the question as prose text in your response.

Call the tool directly. Use the first tool from this list that your environment provides:

1. AskQuestion — call it with a prompt and options array
2. TaskAsk or equivalent structured-input tool
3. (fallback) If the tool call fails or no such tool exists, THEN render the question as numbered options in text and wait.

Do NOT decide in advance whether you have the tool. Attempt the call. The tool call IS the question — do not also write the question as text.

Throughout this skill and its nested skills you will see decision-point markers formatted like this:

**D1 -- BLOCKING:** <instruction to call your question tool>
- question: "<the question>"
- options:
  - "<option A>" -> <what happens>
  - "<option B>" -> <what happens>
- STOP. Do not continue until the user selects an option.

These are instructions for you to follow, not content to display. When you reach one: make the tool call (or render numbered options if no tool exists), then STOP and wait. Do NOT copy the marker text into your response.

User-Facing Communication

Every reply during onboarding must sound like a friendly, knowledgeable colleague walking someone through setup — not a workflow engine quoting internal instructions. Follow these rules in all user-facing output:

Required response structure. Every substantive onboarding reply must include:

1. What we just did — one or two sentences summarizing the completed action and its result.
2. What we're doing next — a plain-English preview of the next step.
3. What you need to do (only when the user has a manual action) — a concrete instruction, not a vague label like "Your turn." Include where to perform the action (e.g. "in Cursor's integrated terminal," "in the project folder," "in your browser," "in macOS Terminal").

Forbidden in user-facing output:

• Internal decision-point IDs (D1, D5, D7, etc.), step numbers as labels (e.g. "Step 5 -- detect"), or skill file names (e.g. "sdk-install/apply/SKILL.md").
• Quoting or paraphrasing raw skill instructions, directive headings, or markdown from these files.
• Workflow-engine language ("BLOCKING," "STOP," "call your structured question tool," "proceed to the next nested skill").

When telling the user to run a command, always say where to run it. Good examples:

• "Run this in the integrated terminal in your editor"
• "Run this from the project root in your terminal"
• "Open a terminal in the packages/api folder and run …"

Bad: "Run npm install" (without location context).

Tone: Friendly, conversational, and confident — like a knowledgeable colleague, not a manual. Use first person naturally (e.g. "I just detected that the flag was created, now I'm going to …"). Assume the reader is an engineer so don't over-explain basic concepts (what a package manager is, what an environment variable does), but do explain LaunchDarkly-specific concepts briefly on first mention (what a context is, what an SDK key is for, why there are different key types).

Step Execution Rules

Do NOT treat the user's initial request (e.g. "onboard me," "set up LaunchDarkly") as blanket permission for file writes, installs, or configuration changes. Each action that modifies the repo, installs packages, or writes secrets requires its own consent at the step where it occurs.

Blocking decision points (you MUST halt and wait for the user's response before continuing):

Non-blocking (you may proceed automatically): Steps 0-3 (log, explore, detect agent, install skills -- no user input needed), D6 plan preview (present and continue unless user objects), Step 5 detect (file reads only), compile check (Step 5 apply Step 4), follow-through file writes (LAUNCHDARKLY.md, editor rules).

Core Principles

1. Detect, don't guess: Inspect the repo for language, framework, and package manager.
2. Minimal changes: Add SDK code alongside existing code; don't restructure the project.
3. Match existing patterns: Follow env vars, config files, and initialization patterns already in use.
4. Validate end-to-end: Confirm the SDK is connected before treating the first flag as proof of success.
5. Paper trail: Keep the Step 0 onboarding log current so another agent or session can continue without re-deriving context.
6. Orient the user first: On a fresh onboarding request, show the Kickoff roadmap before substantive work so the user knows the full arc.
7. Defer credential questions: Do not ask about account status or keys upfront. Account status is inferred through MCP OAuth (Step 4) or surfaced at D7 (Step 5) when keys are needed. Ask for SDK keys / tokens only in Step 4-5 when that step's skill says they are required (Prerequisites).
8. Deep-link to the dashboard: When generating LaunchDarkly dashboard URLs and the project key and/or environment key are known (from MCP tools, user input, or the onboarding log), construct the most specific URL possible instead of linking to a generic page. Use these patterns: What you need to show URL pattern Project flags list https://app.launchdarkly.com/projects/{projectKey}/flags Specific flag https://app.launchdarkly.com/projects/{projectKey}/flags/{flagKey} Environment keys / SDK keys https://app.launchdarkly.com/projects/{projectKey}/settings/environments/{envKey}/keys Project environments list https://app.launchdarkly.com/projects/{projectKey}/settings/environments All projects https://app.launchdarkly.com/projects Only generate deep links when the required keys are known from tool responses or confirmed user input. If they are unknown, use the most specific generic path available and tell the user how to navigate from there (e.g. "Open your project in the LaunchDarkly dashboard, then go to Settings > Environments to find your SDK key").

Kickoff: onboarding roadmap

When the user invokes this onboarding flow (for example by asking you to follow this skill, run LaunchDarkly onboarding, or set up feature flags in the project), treat it as a fresh kickoff unless you are clearly resuming (see Resuming below).

Kickoff sequence (new run — before any numbered step)

Perform these in order in the same assistant turn, then proceed directly into Steps 0-3:

1. Task list: Call your native task tool (Progress Tracking) and create one task per step for Steps 0 through 6 (seven tasks minimum — one each for Steps 0, 1, 2, 3, 4, 5, and 6, even though Steps 0-3 are grouped as a single row below). Do this before rendering the roadmap so progress tracking is in place.
2. Roadmap: Give the user a brief, friendly preview of what you are about to do. Keep it conversational -- a short paragraph or a compact list is fine. Do not render a large table by default (the table below is your internal reference). The user should understand the arc (explore the project, set up tooling, install the SDK, create a first flag) without seeing step numbers or internal labels.
3. Begin Steps 0-3 immediately. These steps do not require a LaunchDarkly account or any user action. Run them in the background and surface only the results: what you found (language, framework, agent) and what you installed (companion skills). Do not narrate each step as a separate heading -- summarize them together when presenting findings to the user. Account status is inferred later (see Prerequisites).

• Resuming: If LAUNCHDARKLY_ONBOARDING.md already exists, read it first per Steps 0-3. Show a shorter "where we are" summary. Refresh the task list to match the log; continue from the log's Next step.

After presenting the roadmap preview, proceed directly into Steps 0-3 (they require no user input or account). Then continue with Step 4.

Workflow

Follow Steps 0-6 in order unless an Edge case says otherwise. When Step 6 (first flag) completes successfully, continue with Default follow-through.

Steps 0-3: Setup (run silently -- do not narrate each step)

These four steps run automatically without user input. Perform them all, then present a single summary of what you found and what you set up. Do NOT show individual step headings, log-creation messages, or install output to the user.

Step 0: Onboarding log. Create or refresh LAUNCHDARKLY_ONBOARDING.md silently.

1. Look for an existing log at the repo root: LAUNCHDARKLY_ONBOARDING.md. If the project keeps docs under docs/, prefer docs/LAUNCHDARKLY_ONBOARDING.md when that folder already exists and the root file is absent.
2. Create or update the log file directly without asking for permission.
3. If resuming: read the log first, align with the stated next step, and only redo work the log marks incomplete or invalid.
4. What to write (update after each numbered step finishes or when something important changes):

- Checklist: Steps 0-6 with status (not started / in progress / done / skipped + brief reason). - Context: coding agent id (once known), language/framework summary, monorepo target path if any, LaunchDarkly project key and environment key when known (never paste secrets or full SDK keys -- say "stored in env" or "user provided offline"). - MCP: configured yes/no, hosted vs fallback, link to config path if relevant. - Commands run: e.g. npx skills add... (no secrets). - Blockers / errors: what failed and what was tried. - Next step: single explicit step number and name (e.g. "Step 5: Install and Initialize the SDK").

1. After errors: append or edit the log with what broke and where you are resuming.

This file is a working log during onboarding. After success, it is deleted and replaced with LAUNCHDARKLY.md (Onboarding Summary).

Step 1: Explore the project. Understand what you are integrating.

1. Identify language and framework. Check dependency files: package.json, go.mod, requirements.txt / pyproject.toml / Pipfile, pom.xml / build.gradle, Gemfile, *.csproj / *.sln, Cargo.toml, etc.
2. Check for existing LaunchDarkly usage. Search for launchdarkly, ldclient, ld-client, LDClient, @launchdarkly.

- If already present: note SDK version and patterns; you may shorten or skip Step 5 per edge cases. - If not present: plan full SDK setup.

1. Identify environment type: server-side app, client SPA, mobile, edge, etc. -- this drives SDK choice.

Deep detection details: Detect repository stack (nested under sdk-install).

Step 2: Detect the agent environment. Infer silently -- do not ask the user.

1. Check for indicators (in priority order — stop at the first strong match):

- Cursor: .cursor/, .cursorrules, or CURSOR_ env vars - Claude Code: ~/.claude/, CLAUDE.md, or CLAUDE_ env vars - Windsurf: .windsurfrules - GitHub Copilot: .github/copilot/ - Codex: ~/.codex/, AGENTS.md

1. If multiple indicators are present, pick the one whose runtime you are currently executing inside. If none match, default to the agent whose tool surface you observe at runtime.
2. Remember the agent id for Step 3 (e.g. cursor, claude-code).

Step 3: Install companion skills. Install flag-management skills from the public repo so later steps can delegate when appropriate.

npx skills add launchdarkly/ai-tooling --skill launchdarkly-flag-create launchdarkly-flag-discovery launchdarkly-flag-targeting launchdarkly-flag-cleanup -y --agent <detected-agent>

Replace <detected-agent> with the value from Step 2. Confirm success; skip skills already installed.

Bundled vs public: Orchestration and setup for this flow live in this folder -- parent SKILL.md, nested mcp-configure, sdk-install (detect / plan / apply), first-flag, and references/ (SDK recipes, snippets, summary, editor rules, etc.). The command above installs flag-management skills from the public launchdarkly/ai-tooling repo only.

After Steps 0-3 complete: Present a single summary to the user covering what you found (language, framework, environment type, whether LD is already integrated, detected agent). Then proceed to Step 4.

Step 4: Configure the MCP Server

Hand off to mcp-configure/SKILL.md for setup (hosted MCP, quick install, manual JSON, agent authorization).

MCP setup requires the user to act outside the agent (clicking a quick-install link, completing OAuth, enabling the server in editor settings). After presenting the instructions, tell the user to restart or refresh the agent so MCP tools become available. Then continue to Step 5 — do not block here with a confirmation question.

Auto-verify on next turn: When the conversation resumes (after the user restarts or sends a message), probe for MCP by calling a lightweight MCP tool such as list-feature-flags with the known project key. If the tool responds normally, MCP is live — note it in the onboarding log and use MCP tools for later steps. If the call fails or no MCP tools are visible, fall back silently to ldcli/API for Steps 5-6 and note the fallback in the onboarding log. Do not ask the user whether MCP is working — find out by trying it.

Do not duplicate MCP procedures in this file. Do not block Step 5 indefinitely on MCP.

Step 5: Install and Initialize the SDK

If the project already has LaunchDarkly installed and initialized (see detect decision tree), skip to Step 6: Create Your First Feature Flag.

Otherwise hand off to LaunchDarkly SDK Install (onboarding), which runs nested skills in order: Detect repository stack -> Generate integration plan -> Apply code changes, using SDK recipes and SDK snippets. If the user asked for both server and client (e.g. API + SPA, Next.js server + browser), follow Dual SDK integrations through plan and apply so both SDKs are really installed and initialized.

Blocking decision points inside Step 5 (see nested skills): D5 (SDK scope), D7 (secret consent), D8 (dependency changes). Do NOT batch tool calls across these boundaries. D6 (plan preview) is non-blocking -- present the plan and continue unless the user objects.

Step 6: Create Your First Feature Flag

Create and evaluate a boolean flag; toggle and observe end-to-end.

1. Follow Create first feature flag.
2. If the launchdarkly-flag-create skill (installed in Step 3) is available, you may use it for create/evaluation wiring only while still completing the verify/toggle checklist in Create first feature flag. Onboarding must remain completable without it.

Install or refresh flag skills via:

npx skills add launchdarkly/ai-tooling --skill launchdarkly-flag-create -y --agent <detected-agent>

See D9 in first-flag for the blocking stop on auth errors.

Default follow-through (not numbered steps)

Do these when finishing onboarding -- same session when possible. They are documentation and handoff tasks, not repeats of Steps 0-6. Do not skip this section -- it is the primary deliverable the user keeps after onboarding.

Setup summary (LAUNCHDARKLY.md) -- REQUIRED

Generate the repo summary per Onboarding Summary. Write it directly -- this is part of the onboarding workflow. The generated LAUNCHDARKLY.md must include all of the following (see template in that reference):

1. SDK Details -- which SDK(s) are installed, package names, key types, initialization files
2. Configuration -- env var names, how secrets are managed, bundler-specific conventions
3. Where to Find Things -- dashboard links with real project key substituted
4. How Feature Flags Work -- a language-specific code example showing flag evaluation in this project's stack (not a generic snippet -- use the same pattern the agent wired during Step 5)
5. Next Steps / Advanced Capabilities -- links to Percentage Rollouts, Targeting Rules, Experimentation, AI Configs, Guarded Rollouts, and Observability
6. AI Agent Integration -- MCP server setup for continued agent-driven flag management

This is not the same file as LAUNCHDARKLY_ONBOARDING.md. The onboarding log is a working checklist; LAUNCHDARKLY.md is the permanent reference for the team.

Clean up the onboarding log: After writing LAUNCHDARKLY.md, delete LAUNCHDARKLY_ONBOARDING.md (or docs/LAUNCHDARKLY_ONBOARDING.md if that was the location used). This is part of the workflow -- do not ask for permission. Removing the working log avoids confusion from having two LaunchDarkly docs in the repo.

Editor rules / skills

• Add editor-specific rules or skill hooks per Editor Rules and Skills. Write them directly -- this is part of the onboarding workflow.

Edge Cases

What NOT to Do

• Don't install an SDK without exploring the project and detecting the stack (Steps 1 and 5); keep the Step 0 log updated as you go.
• Don't upgrade, pin, or add non-LaunchDarkly dependencies (peer-deps, lockfile churn, "latest" bumps) to install or compile the SDK without explicit user approval -- see Apply -- Permission before changing other dependencies.
• Don't hardcode SDK keys in source code -- always use environment variables (see Apply code changes).
• Don't restructure the user's project or refactor unrelated code.
• Don't create flags before Step 5 (SDK install) completes.
• Don't write decision-point questions as chat text -- use your structured question tool (see Decision Points).

References

Continuity

• Step 0 -- LAUNCHDARKLY_ONBOARDING.md (working log; see Steps 0-3)

Step 4 -- MCP (nested skill is primary)

• mcp-configure/SKILL.md -- hosted MCP, verify, edge cases (follow this first)
• MCP UI links -- HTTPS + command: links to open MCP settings per editor
• MCP Config Templates -- per-agent JSON; Local server via npx when hosted MCP is unavailable

Step 5 -- SDK install (nested skills)

• sdk-install/SKILL.md -- orchestrates detect -> plan -> apply (follow this first)
• Detect repository stack
• Generate integration plan
• Apply code changes

First flag (Step 6)

• Create first feature flag

Default follow-through

• Onboarding Summary -- template for LAUNCHDARKLY.md
• Editor Rules and Skills

SDK index

• SDK recipes
• SDK snippets

Public flag skills (install via Step 3)

• github.com/launchdarkly/ai-tooling -- launchdarkly-flag-create, launchdarkly-flag-discovery, launchdarkly-flag-targeting, launchdarkly-flag-cleanup