launchdarkly/agent-skills
28 skills37.2k installs448 starsGitHub
Install
npx skills add https://github.com/launchdarkly/agent-skillsSkills in this repo
1OnboardingThe LaunchDarkly onboarding skill—documented as first-flag—walks a solo builder through creating a boolean feature flag and proving it works in running code after the SDK install step. It sits in the Ship phase because the SDK is already connected; the remaining work is operational: create the flag, keep the default off, confirm the app evaluates false, flip the flag on, and confirm true. The skill explicitly allows MCP, REST API, or ldcli paths and warns never to paste access tokens into commands, only environment references. When launchdarkly-flag-create is installed from LaunchDarkly’s ai-tooling repo, agents may use it for flag scaffolding and evaluation snippets, but the verification ritual stays mandatory so you do not ship a broken toggle. This is a workflow skill for LaunchDarkly projects you control, not a generic experimentation framework for every stack.2kinstalls2Launchdarkly Flag CleanupLaunchDarkly Flag Cleanup is an agent skill that teaches coding agents to retire feature flags without guessing. It walks through readiness checks against LaunchDarkly, calculates the forward value that must remain in code, strips flag references safely, and packages the change as a well-documented pull request. Solo and indie builders who ship behind flags accumulate invisible debt fast; this skill pairs procedural guardrails with MCP-backed API orchestration so removal matches what production actually serves. It is experimental (v1.0.0) and depends on configuring the LaunchDarkly MCP server in the agent environment. Use it when a rollout is fully on and the flag only adds noise, risk, and double-read paths in the codebase—not when experimentation or kill switches are still active.1.6kinstalls3Launchdarkly Flag DiscoveryLaunchDarkly Flag Discovery is an experimental Apache-2.0 agent skill that walks you through auditing feature flags in a LaunchDarkly project when you worry about flag debt, stale toggles, or unclear inventory. It assumes the remotely hosted LaunchDarkly MCP server is configured and leans on list-flags, get-flag, and cross-environment status tools, with optional helpers for staleness ranking, combined health views, and removal readiness checks. Solo builders and small teams shipping behind flags use it before cleanup sprints or architecture reviews so recommendations stay tied to real project state rather than guesswork. The skill emphasizes actionable next steps—what to archive, what needs verification, and what still gates production—rather than dumping raw API JSON. Invoke it when stakeholders ask how healthy the flag landscape is or which toggles are safe to delete after launch.1.6kinstalls4Launchdarkly Flag Createlaunchdarkly-flag-create walks a builder or coding agent through adding a feature flag the way this specific codebase already does it—not with a generic snippet. The skill starts with discovery: find SDK imports, initialization, and how flags are named and evaluated. Only then does it create the flag in LaunchDarkly through the MCP server's create and get tools, optionally listing flags and updating metadata for consistency with team conventions. You use it when someone asks for a new flag, a feature toggle around code, or experiment setup. It fits solo builders on SaaS or API products who already use LaunchDarkly and have MCP configured. Because flags touch implementation and later rollout discipline, journey scope is multi-phase with canonical placement in Build integrations. Outcome is a created flag plus evaluation code aligned with local patterns and a sanity check that the flag exists and is referenced correctly.1.5kinstalls5Launchdarkly Flag TargetingLaunchDarkly Flag Targeting is an experimental agent skill from LaunchDarkly’s ai-tooling repo that teaches coding agents how to manage feature flag targeting through the hosted LaunchDarkly MCP server. Solo builders and small teams use it when they need to turn flags on or off, set percentage rollouts, add targeting rules, pin users or contexts, or mirror targeting from staging into production without losing the current state. The workflow emphasizes understanding existing configuration first, then applying incremental changes with production-minded caution. It pairs procedural guidance in SKILL.md with MCP tools optimized for agent orchestration rather than raw REST exploration. Install by placing the skill directory in your client’s skills path and configure the MCP server in the same environment. Best suited for SaaS and API products already instrumented with LaunchDarkly; not a substitute for initial SDK wiring or flag design strategy.1.5kinstalls6Aiconfig Createaiconfig-create is a deprecated alias skill in the LaunchDarkly agent-skills repo. It does not create or edit AI configurations anymore; every invocation should stop immediately and call configs-create instead, which holds the full up-to-date procedure. Solo builders maintaining saved prompts, CI snippets, or team playbooks that still mention aiconfig-create install this skill so agents self-correct instead of following obsolete steps. There is no workflow, checklist, or MCP bridge inside this file—only a hard redirect. Update your references to configs-create when you can, and treat this package as compatibility glue during migration rather than a capability you rely on for shipping features.1.5kinstalls7Aiconfig Projectsaiconfig-projects is a deprecated redirect skill in the LaunchDarkly agent-skills repository. It exists solely so older documentation, saved prompts, and automations that still invoke aiconfig-projects land on clear guidance instead of silent failure. The skill body forbids following any legacy instructions and tells the agent to invoke projects, which holds the current complete workflow. Solo builders maintaining LaunchDarkly or feature-flag skill libraries should treat this entry as a catalog alias, update references to projects, and never depend on this slug for real work. Prism lists it so search and old links surface the rename rather than pulling users into obsolete steps.1.5kinstalls8Aiconfig Toolsaiconfig-tools is a deprecated Prism entry kept so older prompts, docs, and automation that still invoke aiconfig-tools do not break silently. The entire SKILL.md is a redirect: agents must not perform any task described here and must invoke the tools skill instead, which holds the current LaunchDarkly AI configuration instructions. Solo builders maintaining agent skill lists should replace aiconfig-tools with tools in saved prompts and CI skill manifests. There is no feature surface in this package—only naming continuity—so installs are about reference hygiene, not new capability. When an agent lands here from a legacy reference, the correct outcome is an immediate handoff to tools, not partial execution of outdated steps.1.5kinstalls9Aiconfig Updateaiconfig-update is a deprecated LaunchDarkly agent skill kept solely as a naming bridge. Solo builders and small teams sometimes inherit prompts, CI snippets, or marketplace links that still invoke aiconfig-update after the skill was renamed to configs-update. This package does not update flags or AI configs itself; it instructs the agent to refuse the old skill body and invoke configs-update instead, which holds the current end-to-end procedure. For Prism catalog readers shipping SaaS with feature flags, treat this entry as documentation hygiene: update your references once, then bookmark configs-update for real operate-phase iteration. Installing it only helps agents that land on stale slugs so they do not silently run obsolete steps.1.5kinstalls10Aiconfig Variationsaiconfig-variations is a deprecated LaunchDarkly agent-skill stub that exists solely so old slugs and documentation links do not dead-end after the skill was renamed to configs-variations. Solo and indie builders who copied prompts, CI snippets, or marketplace entries from earlier catalogs may still see aiconfig-variations in their repos; this entry tells the agent to refuse the obsolete procedure and open configs-variations instead. It does not expose REST calls, flag CRUD, or variation editing—those capabilities live entirely in the successor skill. Treat this package as catalog hygiene: update your references once, then delete or ignore the old name in favor of configs-variations for any real feature-flag or AI-config work with LaunchDarkly.1.5kinstalls11Aiconfig TargetingAiconfig Targeting is a deprecated LaunchDarkly agent skill kept solely so old references to aiconfig-targeting still resolve to the correct capability. It performs no targeting work: the entire contract is to invoke configs-targeting instead and to update documentation, saved prompts, or automation that still use the legacy name. Solo builders maintaining feature-flag or AI config workflows in Claude Code or Cursor should treat this entry as a catalog stub, not an install-and-run integration. Prism lists it so search and bookmarks do not silently break after the rename. Once references are updated, prefer configs-targeting for the actual procedural steps. No quantified rule sets live in this redirect—the value is accurate routing and reduced agent confusion.1.3kinstalls12Aiconfig Online Evalsaiconfig-online-evals is a deprecated redirect stub in the LaunchDarkly agent-skills repo. Solo builders and agents should not treat it as an eval workflow; SKILL.md states the skill was renamed and that users must invoke online-evals for complete, current guidance on online evaluation. The package exists so saved prompts, CI references, and older catalog links do not 404 or silently run outdated steps. For Prism, tag it as meta maintenance around AI configuration in production, with the real journey value living on the online-evals entry. Update any automation, skill lists, and team docs to use online-evals before running production eval or feature-flag experiments.1.3kinstalls13Launchdarkly Metric ChooseLaunchDarkly Metric Choose is an advisory agent skill that helps solo builders and small teams select the right metrics before LaunchDarkly experiments, guarded rollouts, or release-policy checks. It expects the remotely hosted LaunchDarkly MCP server with list-metrics and list-metric-events (and optionally list-release-policies) so the agent can see what already exists, which event keys are healthy, and what project policies will auto-attach. The workflow centers on feature context: clarify the change, inventory available metrics, validate event activity, then recommend a primary metric, guardrails, and rollout monitors in a clear typed structure. The skill does not create metrics, attach them to experiments, or configure rollouts—it only guides selection so follow-on LaunchDarkly skills or console work starts from a defensible measurement plan. That separation keeps planning fast and low-risk while still grounding recommendations in live catalog and policy data rather than generic analytics advice.1.3kinstalls14Launchdarkly Metric CreateLaunchDarkly Metric Create is an agent skill for solo builders and small teams running feature flags and experiments who need trustworthy measurement in LaunchDarkly. It walks the agent through choosing the right metric kind, ensuring the underlying event is actually flowing—which often means SDK wiring and environment keys before any metric object exists—scanning for duplicate event keys, drafting a metric configuration, and only creating the metric after you explicitly confirm. The flow is opinionated on verification: create via MCP, then read back with get-metric so the catalog outcome is a live, inspectable metric tied to real telemetry. It fits when you are standing up an A/B test, a gradual rollout, or a kill-switch feature and need binary or numeric outcomes aligned with product goals. It is not a generic analytics dashboard; it is LaunchDarkly-native instrumentation plus metric lifecycle through the official MCP integration, so your agent does not invent keys or skip the instrument-first rule.1.3kinstalls15Launchdarkly Metric InstrumentLaunchDarkly metric instrument is a reference skill for solo builders and small teams who already use LaunchDarkly and need consistent custom metric instrumentation across SDKs. It explains the critical split between server-side SDKs, which require a LaunchDarkly context on every `track()` call, and client-side browser SDKs, which do not. The guide walks through npm installs, initialization with SDK keys or client-side IDs, count versus value metrics via `metricValue`, optional event data payloads, and flushing behavior when processes exit quickly. Use it while integrating backend APIs, web clients, or mobile-backed services so experiment metrics align with LaunchDarkly’s data model and your existing code style. It reduces copy-paste errors that silently drop or mis-attribute events during A/B tests and staged rollouts.1.3kinstalls16Aiconfig MigrateAiconfig-migrate is a meta redirect stub in the LaunchDarkly agent-skills catalog. It exists only because the real capability was renamed to migrate; any agent or human that lands here should immediately switch skills and update references in documentation, saved prompts, or CI strings. There is no migration workflow, flags, or LaunchDarkly API steps in this package—invoking it for actual AI config migration would be wrong. Solo builders maintaining agent skill libraries use it as a safety rail when grepping old skill names. Treat it as documentation glue, not a tool. After redirect, all substantive work happens under migrate. Prism lists it so discoverability of the old name still routes to the correct successor.1.2kinstalls17Aiconfig Ai MetricsAI Config AI Metrics is a LaunchDarkly-focused agent skill that instruments an existing codebase so AI Config’s Monitoring tab receives reliable telemetry on every provider invocation. Solo builders shipping LLM features often paste manual tracker calls and silently lose token counts, streaming time-to-first-token, or error classification as SDKs evolve. The skill audits what is already instrumented, walks the official four-tier ladder from lowest ceremony to full manual control, and stops at the first tier that still satisfies duration, tokens, and success/error (plus TTFT for streams). It emphasizes choosing the highest fitting tier rather than defaulting to hand-rolled trackers that diverge from provider packages. You need the server-side AI SDK at 0.20.0 or newer and a configured AI Config before starting. Use when promoting a prototype to production observability or when LaunchDarkly dashboards look empty despite working inference paths.1.2kinstalls18Aiconfig Custom MetricsThis Prism listing documents a deprecated LaunchDarkly agent skill that was renamed to custom-metrics. It exists solely so legacy prompts, saved workflows, and documentation that still say aiconfig-custom-metrics route builders to the current skill name. The body instructs agents not to follow any instructions in this file and not to perform the original metrics task from here. Solo builders should treat this as a catalog compatibility shim: when an agent loads it, the only correct action is to switch invocation to custom-metrics. It does not teach metric definitions, LaunchDarkly configuration, or AI feature measurement. Install or reference it only if you maintain older references; otherwise browse and use custom-metrics directly for real Grow-phase instrumentation work.1.2kinstalls19ApplyApply is the LaunchDarkly agent skill that turns an approved integration plan into real dependency installs, environment configuration, and SDK initialization at the correct entrypoints. Solo builders use it after detect and plan when they are ready to wire feature flags without hand-rolling steps from memory—especially dual-SDK setups that require separate server and client (or mobile) packages, installs, credentials, and inits. The skill enforces workflow discipline: if the plan is dual SDK, both tracks must be completed; if the plan omitted a track the user requested, the agent must return to plan rather than invent the second half. Credential collection happens here with consent, aligning with LaunchDarkly onboarding expectations. It pairs with run as the next step to start the app and validate flag behavior. Expect intermediate complexity when monorepos or multiple surfaces are in scope.1.2kinstalls20DetectDetect is a LaunchDarkly agent skill for the first step of SDK onboarding: read the repository and produce an accurate picture of what you are shipping before any install or code edits. Solo and indie builders use it when they want feature flags but are unsure which SDK variant, which package manager command, or which entrypoints apply—especially in monorepos or polyglot layouts. The skill walks indicator files (JavaScript, Python, Go, Java/Kotlin, Ruby, .NET manifests), infers frameworks from conventional structure, and notes whether LaunchDarkly is already present. It is deliberately read-only: no credentials, no dependency changes, no init code. Output feeds the integration plan skill so later steps do not guess stack details from memory. Pair it with the parent sdk-install workflow when onboarding Claude Code, Cursor, or Codex onto an existing codebase.1.2kinstalls21First Flagfirst-flag is a LaunchDarkly agent skill for solo builders finishing parent onboarding Step 6 after Apply code changes. It walks you through creating a boolean feature flag, adding evaluation in your app, and proving control by observing behavior with the flag off and on. The flow intentionally supports MCP, direct API calls, or ldcli so you are not locked into one toolchain, and it can delegate flag scaffolding to the optional launchdarkly-flag-create skill from ai-tooling when that package is installed—while still requiring the verify-off/verify-on checks either way. Apache-2.0 licensed; compatibility assumes SDK install (Step 5) and project access. Use it when you need confidence that LaunchDarkly is wired correctly before you depend on flags in staging or production.1.2kinstalls22Mcp Configuremcp-configure is a reference skill from LaunchDarkly that packages ready-made MCP server definitions for their hosted LaunchDarkly integration. Solo builders shipping SaaS or agent-assisted workflows use it when they want feature-flag operations inside the IDE without embedding secrets. The readme walks through four common clients—Cursor, Claude Code, GitHub Copilot’s coding agent, and Windsurf—with the exact JSON shape each expects and where the file or UI lives. All variants point at the same HTTPS MCP URL and rely on OAuth rather than static API keys, which reduces leak risk for indie projects. After configuration, you authorize once per environment and let the agent call LaunchDarkly through MCP. It does not teach flag strategy or SDK usage; it is strictly connectivity and client-specific setup aligned with LaunchDarkly’s official hosted MCP documentation.1.2kinstalls23PlanPlan is a LaunchDarkly agent skill that turns repository detect context into a minimal SDK integration blueprint. Solo builders and small teams use it when they want feature flags without guessing which SDK, which keys, or which files to change across backend APIs, SPAs, native mobile, React Native, or Flutter targets. The skill sits in the official SDK Install onboarding chain: after Detect classifies the stack, Plan selects the right SDK type(s)—including dual server-plus-client when both tiers need flags—and lists env naming and files to modify using the sdk recipes reference. It deliberately stops at planning so apply can implement without scope creep. Apache-2.0, version 0.2.0, requires completed detect context from the sibling skill. Ideal for agents preparing a focused PR plan before touching production config.1.2kinstalls24Sdk Installsdk-install is LaunchDarkly’s parent onboarding skill for installing and initializing the right feature-flag SDK in an existing codebase. It orchestrates three nested skills strictly in order: detect the language and framework, plan the integration approach, then apply code changes including environment wiring. Solo builders should reuse project context from parent Step 1 Explore the Project and only re-run deep detection when something is unclear. The skill deliberately stops before creating flags—after Apply, the parent onboarding flow continues at Step 6 with the first-flag skill. Credential handling is staged: detect and plan do not require users to paste SDK keys upfront; account status may be inferred via MCP OAuth at Step 4, with key material collected when Apply runs per parent Prerequisites. Supports server, client-side, and mobile SDK types with a summarized key-type table. Apache-2.0 licensed metadata version 0.2.0 from LaunchDarkly. Ideal when you have a supported stack and need deterministic onboarding instead of guessing package names and init snippets.1.2kinstalls25Launchdarkly Guarded RolloutLaunchDarkly Guarded Rollout is an agent skill that guides you through designing and starting guarded rollouts in LaunchDarkly. It explains how stages progressively increase traffic to a new flag variation while watching selected metrics for regressions, and how the platform can pause or roll back when thresholds trip. The workflow depends on the remotely hosted LaunchDarkly MCP server, calling tools such as start-guarded-rollout, get-flag, and list-metrics, with optional stop-guarded-rollout, toggle-flag, and create-metric. Solo builders shipping SaaS or APIs use it when a feature is coded but should not go to 100% instantly. The skill structures rollout design, metric selection, and threshold configuration so your agent does not improvise unsafe release plans. Pair it with your existing flag definitions and observability metrics before flipping production traffic.1kinstalls26Launchdarkly Experiment SetupLaunchDarkly Experiment Setup is an agent skill that walks a solo builder or small team through creating and running controlled experiments in LaunchDarkly using MCP tools instead of clicking through opaque UI flows alone. It covers designing the experiment, wiring metrics and treatments to feature flag configuration, starting an iteration to collect data, checking status with get-experiment, and stopping when a winner is clear—with guidance to update or swap design between iterations when the platform permits edits for the current iteration state. The skill assumes the remotely hosted LaunchDarkly MCP server is already configured and names the minimum tool set explicitly so the agent does not hallucinate APIs. It suits builders who ship feature flags in production and want agent-assisted rigor for A/B-style decisions without delegating experiment governance to ad-hoc prompts.1kinstalls27Aiconfig Agent Graphsaiconfig-agent-graphs is a deprecated redirect skill retained in the LaunchDarkly agent-skills collection after the capability was renamed to `agent-graphs`. The entire SKILL.md instructs agents not to perform graph or AI-config work in this file and to call `agent-graphs` for complete, up-to-date procedures. Indie builders who compose multi-skill agent workflows—especially across feature-flag or AI orchestration setups—risk silent failure if one node still uses the pre-rename slug. Prism lists this entry so search and old links surface the correct successor. When you discover `aiconfig-agent-graphs` in a README, Claude project file, or CI skill manifest, replace it with `agent-graphs` and delete the redirect from your local allowlist if you maintain a curated subset. Treat this as documentation plumbing, not a building block for product logic.1kinstalls28Aiconfig Snippetsaiconfig-snippets is a deprecated compatibility entry in the LaunchDarkly agent-skills package. It exists solely so older references to the former name still resolve in the catalog and in agent sessions, but the SKILL body forbids following any task instructions here. Solo builders and agents should treat this record as a pointer: invoke the `snippets` skill instead, which holds the current, complete guidance for the same capability. Update any documentation, Cursor rules, CI templates, or saved prompts that still mention `aiconfig-snippets` to reduce confusion and avoid wasted turns. This pattern matters for indie teams who fork skill repos or pin versions—one stale slug in a compound workflow can silently load an empty redirect. Use this skill only in the narrow moment when an automation explicitly calls the old identifier; otherwise install and use `snippets` directly.1kinstalls