create-experiment
Thin compatibility entrypoint for creating a new Harness FME experiment. Invokes manage-experiments' Create experiment flow directly - same scope, parent/environment resolution, hypothesis, metrics, treatments, window, and rule - and returns the created experiment id plus verification. Use when asked to create an experiment, set up an A/B test, or design a new experiment. Do NOT use for updating, pausing, resuming, completing, archiving, or deleting an experiment (manage-experiments), or for explaining results (review-experiment-results). Trigger phrases: create experiment, set up an A/B test, design an experiment, launch a test, new experiment for X.
88%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Passed
No findings from the security scan
Create Experiment
Compatibility entrypoint kept for direct /create-experiment invocation. This
skill does not duplicate any workflow - it hands the entire creation flow to
manage-experiments's Create experiment
section and returns once that flow completes.
Tools
Performs no operations of its own; it uses the operations of manage-experiments' Create experiment flow, named in tool-map.md.
Instructions
Step 1: Delegate
Read manage-experiments/SKILL.md and follow
its Phase 1: Establish scope and Create experiment section (Steps 1-8)
exactly as written, using the current request's scope (org/project), parent
flag + environment, hypothesis, metrics, treatments, window
(startAt/endAt), and rule. Do not re-derive or re-interpret any of that
section's logic here - this skill has no independent instructions for
resolving treatments, choosing metrics, or drafting the payload; it only
exists so /create-experiment remains a valid invocation.
If the tool interface you're running under can't follow a markdown link to
load another skill's file on demand, read
skills/fme/manage-experiments/SKILL.md directly (same repo, relative path) and
proceed the same way.
Step 2: Return the result
Once manage-experiments' Step 8 (Verify) completes, report the created
experiment's id, status, and the verified fields (baseline/comparison
treatments, key/supporting metrics, rule) back to the user. Don't add a
second verification pass - manage-experiments' Step 8 is authoritative.
Examples
- "Create an experiment on the new-checkout flag" → delegate to manage-experiments Create flow, return created id
- "Set up an A/B test for the pricing page" → same delegation
- "Pause experiment pricing_test_2" → not this skill; route to manage-experiments (Update)
- "Did the checkout experiment win?" → not this skill; route to review-experiment-results
Performance Notes
- No independent tool calls: every MCP/CLI operation happens inside the delegated manage-experiments flow, so there's no duplicate scope resolution or double-verification to optimize away here.
Troubleshooting
| Issue | Resolution |
|---|---|
| User asks to update/pause/resume/complete/archive/delete an existing experiment | Route to manage-experiments (Update/Delete), not this skill |
| User asks about results, winner, or significance | Route to review-experiment-results |
| manage-experiments' Create flow stops on a missing decision (hypothesis, treatments, metrics) | Surface that stop to the user here too - don't guess on its behalf |
| Delegated flow fails partway (e.g. 404 on parent, 409 duplicate name) | Report the error as manage-experiments' Troubleshooting table describes; don't retry with invented values |
- Repository
harness/harness-ai
- Revision
ca2ab81
- Last updated
- First committed
Is this your skill?
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.