Experiment Launch

experiment-launch

Launch a GrowthBook A/B test end-to-end: create the experiment in draft, prep or reuse the feature flag, add the experiment-ref rule on a fresh draft revision, then call /start to publish the rule and flip the experiment to running. Handles the approval-required and pre-launch-checklist failure paths.

All API calls go through the bundled helper: ${CLAUDE_PLUGIN_ROOT}/scripts/gb-call. It needs GB_API_KEY — set in your shell, or written to ~/.config/growthbook/.env by /growthbook:setup. If it's missing or invalid, gb-call's error message points back at /growthbook:setup.

Required inputs

Collect from the user (or earlier skill output) before starting. Prompt for what's missing.

• Feature flag name — kebab-case key, regex [a-zA-Z0-9_-]
• Experiment name — human-readable
• Variations — array of {name, value}, length ≥ 2. The first entry is the control. Values are serialized as strings on the rule (booleans → "false"/"true", numbers → "42", JSON → JSON-encoded string).
• Project ID (optional) — pins the experiment and flag to a specific project

Optional inputs

• Hypothesis — falsifiable; if/then/because format
• Template description — English description matched against templateMetadata.name/description returned by /v1/experiment-templates. If omitted and templates exist, ask the user to pick one or skip.

If no template is used, also collect (or resolve interactively in step 2):

• Datasource — id or English name
• Hash attribute — the unit of randomization (id, device_id, etc.). Must equal the identifierType of the assignment query selected on the datasource.
• Assignment query — id or English name; lives inside the chosen datasource's assignmentQueries
• Goal metric ID — exactly one (the primary KPI you'd ship or kill on)
• Secondary metric IDs — supporting metrics
• Guardrail metric IDs — defensive metrics that should not regress

Workflow

Track progress with this checklist. Do not skip or reorder.

- [ ] 1. Pick a template (or skip)
- [ ] 2. Resolve hash attribute → datasource → assignment query → metrics (no-template path only)
- [ ] 3. Create the experiment in draft
- [ ] 4. Create or reuse the feature flag
- [ ] 5. Add the experiment-ref rule on a fresh draft revision
- [ ] 6. Prompt user to QA the experiment and feature flag config
- [ ] 7. POST /start; branch to 7a (approval) or 7b (checklist) on 400
- [ ] 8. Report links and state

1. Pick a template (or skip)

gb-call GET /api/v1/experiment-templates

• Zero templates → continue without one; go to step 2.
• One or more → list as name — description plus a final "Skip" option. If a template description was provided, pre-select the best match by templateMetadata.name/description and confirm. Never invent a template.

If chosen, capture: id (becomes templateId), datasource, exposureQueryId, hashAttribute, goalMetrics, statsEngine, targeting. Templates inject all of these; skip step 2 entirely.

If the template's type is "multi-armed-bandit", halt and confirm with the user. Bandits behave very differently from standard A/B tests (dynamic traffic allocation, per-arm probabilities instead of winner/loser, different analysis), and this skill's launch and analysis assumptions are written for type: "standard". Recommend they configure bandits in the UI for now.

2. Resolve hash attribute → datasource → assignment query → metrics

No-template path only. Order matters — pick hash attribute first so you don't trap yourself on a datasource that can't randomize on it.

2a. Pick the hash attribute. Filter to attributes flagged as hashAttribute:

gb-call GET /api/v1/attributes

Surface attributes where hashAttribute === true and archived !== true. Ask the user to pick. If the filtered list is empty, halt — tell the user to mark at least one attribute as a hash attribute under Settings → Attributes in GrowthBook.

2b. Pick the datasource.

gb-call GET /api/v1/data-sources

Resolve English name or ID against dataSources[].name / id. Capture DATASOURCE_ID and keep the full object — 2c reads its assignmentQueries.

2c. Pick the assignment query. Filter dataSources[].assignmentQueries to entries where identifierType === HASH_ATTRIBUTE:

• Exactly one match → auto-select it; print one line stating the choice.
• Zero matches → halt and offer three fixes: pick a different datasource (rerun 2b), change the hash attribute (rerun 2a), or add an assignment query for `` in GrowthBook.
• Two or more → list each as name (identifierType=) — and let the user pick.

If assignmentQueries is empty entirely, halt and tell the user to configure one in GrowthBook before re-running.

2d. Pick metrics, filtered by datasource. The API rejects metrics from a different datasource than the experiment's:

gb-call GET '/api/v1/fact-metrics?datasourceId=&limit=100'
gb-call GET '/api/v1/metrics?datasourceId=&limit=100'

Help the user pick:

• Goal metric(s) (GOAL_METRIC_IDS). Ideally one, two max — push back at three or more and demote the rest to secondary or guardrail.
• Secondary metrics (SECONDARY_METRIC_IDS) — supporting context.
• Guardrail metrics (GUARDRAIL_METRIC_IDS) — defensive. Push back if they name none; every experiment needs at least one. Guardrails are excluded from multiple-comparison correction by design.

3. Create the experiment in draft

Set trackingKey to the feature flag name so the SDK ties exposures to the flag. Each variation needs a stable string key ("0", "1", ...) and name. Variation values live on the flag rule (step 5), not on the experiment payload.

Template path — do NOT also send datasourceId / assignmentQueryId; the template provides them and the API rejects the combination.

{
  "templateId": "",
  "trackingKey": "",
  "name": "",
  "hypothesis": "",
  "variations": [
    { "key": "0", "name": "Control" },
    { "key": "1", "name": "Treatment" }
  ],
  "project": ""
}

No-template path — send everything from step 2 explicitly:

{
  "datasourceId": "",
  "assignmentQueryId": "",
  "hashAttribute": "",
  "trackingKey": "",
  "name": "",
  "hypothesis": "",
  "variations": [
    {"key": "0", "name": "Control"},
    {"key": "1", "name": "Treatment"}
  ],
  "metrics": [""],
  "secondaryMetrics": [],
  "guardrailMetrics": [],
  "project": ""
}

Notes:

• metrics is the goal-metric array; with the one-goal rule it should always be length 1.
• Omit secondaryMetrics / guardrailMetrics entirely if the user picked none. Don't send empty arrays.

Then POST:

echo '' | gb-call POST /api/v1/experiments -

Capture from the response:

• experiment.id — used in steps 5 and 6.
• experiment.variations[].variationId — the string ID for each variation (e.g. var_abc123). You need these in step 5; they are required on the experiment-ref rule.

4. Create or reuse the feature flag

Try create first unless the user said the flag already exists. The flag must default to the control value (variation 0's value), serialized as a string. Default all environments to off as well.

{
  "id": "",
  "valueType": "",
  "defaultValue": "",
  "description": "Drives experiment:  ()",
  "project": ""
}

echo '' | gb-call POST /api/v2/features -

• Success → go to step 5.
• 409 Conflict (flag exists) → fall through to the reuse compatibility checks below.

Reuse path — fetch and validate:

gb-call GET /api/v2/features/

Run these compatibility checks against the response. Each row says what to do on failure:

| Check | Action on failure | | -------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------- | | archived === false | Halt. Tell the user to un-archive the flag in the UI before re-running. | | valueType matches the experiment's ` | **Halt.** Surface both values; do not silently change types. | | project matches the experiment's project (when set) | **Halt.** Reusing a flag from another project misroutes the experiment. | | defaultValue equals the control value (string-compared) | **Warn**, do not halt. The experiment rule supplies the variation values; the existing default applies only when the rule doesn't match. Ask "continue?" | | No existing rule with type === "experiment-ref" AND experimentId === already attached | If one exists → this experiment is already wired up. Skip step 5 and jump to step 6. | | No conflicting experiment-ref rule for another _running_ experiment in the same environments | **Warn**, do not halt. Ask "this flag is currently driving experiment `; add another rule alongside it?" |

When all checks pass, capture the flag's identity and proceed to step 5. Do not mutate the existing flag here — all mutations go through the draft revision in step 5.

5. Add the experiment-ref rule on a fresh draft revision

Use the literal version new to create a draft and add the rule in one call. The path segment new is a magic value that creates a draft branched off the live revision atomically. If the feature has multiple environments, prompt the user for which environments to go live in, and ensure this rule turns on the feature flag for those environments in addition to adding the rule to those environments.

{
  "rule": {
    "type": "experiment-ref",
    "experimentId": "",
    "enabled": true,
    "allEnvironments": true,
    "variations": [
      {
        "value": "",
        "variationId": ""
      },
      {
        "value": "",
        "variationId": ""
      }
    ],
    "description": "Experiment: "
  }
}

echo '' | gb-call POST /api/v2/features//revisions/new/rules -

variations[] must have one entry per experiment variation, in the same order as step 3. Each entry needs both value (serialized as a string) and variationId (the string ID captured in step 3) — variationId is required by the v2 features validator and omitting it returns a 400. Capture the returned version for the draft revision.

Do not publish the revision here. Step 7's /start call auto-publishes the draft when it flips the experiment to running.

6. Pause for QA before /start

The draft revision and experiment are reversible up to this point; step 7's /start publishes the rule and flips the experiment to running.

If the user seems to want you to also embed the feature flag in the codebase, now is a good time to do that.

Stop and surface the UI links so the user can QA the flag default, the rule's variation values, and the experiment's targeting and metrics in the GrowthBook UI:

• Experiment: /experiment/
• Feature: /features/ (draft revision ``)

Derive ` from GBAPIURL by swapping api. → app.`, as in step 8. Wait for the user's explicit go-ahead before proceeding to step 7.

7. Start the experiment

gb-call POST /api/v1/experiments//start '{}'

The /start endpoint does two things server-side:

1. Publishes the draft feature revision from step 5 (publishPendingFeatureDraftsForExperiment).
2. Enforces the org's pre-launch checklist.

Either can fail with a 400. Branch:

• Body starts with "This revision requires approval before publishing" → step 7a.
• Body lists incomplete checklist items → step 7b.
• 2xx → step 8.

7a. Approval required

The experiment and flag exist; only the rule revision is stuck in draft. Halt and offer the user three concrete paths:

> Your org requires approval before this feature flag rule can go live, and /start will not flip the experiment to running until the rule is published. Revision ` on is in draft state. Pick one: > > **A. Standard review flow** (recommended) — I'll request a review now. A teammate (not you, since you created the draft) approves it in the GrowthBook UI at /features/, then you re-run me and I'll resume from /start. > > **B. Org-wide bypass** — an admin enables "REST API always bypasses approval requirements" in **Settings → General → Approvals**. After that, re-run me. > > **C. Per-token bypass** — use a Personal Access Token whose role grants bypassApprovalChecks on this project (Admin or custom role). Update GBAPIKEY`, then re-run me.

If the user picks A, request review on the draft and stop:

echo '{"comment":"Auto-requested by experiment-launch for "}' \
  | gb-call POST /api/v2/features//revisions//request-review -

Do not attempt submit-review yourself — the API rejects self-approval on a draft you created. Stop and tell the user to re-run after approval.

If the user picks B or C, stop with a one-line note. The existing draft will pick up the new permission and publish on retry.

Do not silently retry /start, ignore the error, or discard and recreate the draft to work around the policy.

7b. Checklist incomplete

The REST API does not expose a separate start-checklist endpoint — the failure body from /start is the canonical source. Parse it and surface the incomplete items verbatim:

> The pre-launch checklist isn't complete. The /start call returned: > > ` > > Fix the listed items in the GrowthBook UI at /experiment/, then re-run me — I'll jump straight back to /start`.

Only retry /start with {"skipChecklist": true} in the body if the user explicitly asks to bypass. Never default to bypassing; the checklist is intentional friction.

8. Report

Print a summary:

• Experiment name and id
• Feature flag ID and published revision version
• Template used (name + id) if any
• Unit of randomization (hashAttribute)
• Variations and their values
• Pre-launch checklist status (should be allRequiredComplete=true)
• Experiment status (should be running after a clean /start)
• Direct UI links (derive the host from GB_API_URL by swapping api. → app.):
• Experiment: /experiment/
• Feature: /features/

Guardrails

• Ideally one goal metric, two max. GrowthBook's decision framework treats goal metrics as plural by design and the power calculator supports up to five, but each additional goal dilutes power and complicates the ship/kill decision. Push back at three or more; demote the rest to secondary.
• At least one guardrail. Push back if the user skips guardrails.
• hashAttribute and assignmentQuery.identifierType must match. Mismatch is a real and recoverable error; surface the fix paths in step 2c.
• Metrics must live on the experiment's datasource. Filter /v1/metrics and /v1/fact-metrics by datasourceId in step 2d.
• Do NOT mix templateId with datasourceId/assignmentQueryId. The template path supplies those; the no-template path supplies them explicitly. Mixing yields a 400.
• Flag default = control value. Variation values for flag-linked experiments are strings on the rule — "false"/"true"

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: growthbook
• Source: growthbook/skills
• License: MIT

Install and usage instructions live in the source repository linked above.