Home/AI Agent Core Development/momentic-explore-prompt
M

momentic-explore-prompt

by @momentic-aiv
4.5(120)

This skill generates the explore-prompt.md file for Momentic's explore agent, providing repo-specific context including applications under test, target URLs, authentication, and test placement. It enables the agent to automatically identify changed user journeys and build tests, reducing manual setup and errors in CI/CD pipelines.

explorationprompt-generationtest-automationai-agentGitHub
Installation
npx skills add https://github.com/momentic-ai/skills --skill momentic-explore-prompt
compare_arrows

Before / After Comparison

1 组
Before

Developers manually inspect the repo structure, confirm app URLs, auth modules, and test paths, then write the explore-prompt.md file. Each configuration takes about 30 minutes and is prone to errors causing explorer agent failures.

After

This skill automatically analyzes repo configs, code, and CI workflows to generate an accurate, complete explore-prompt.md file. Configuration time drops to 2 minutes, significantly reducing human errors and ensuring smooth explorer agent runs.

SKILL.md

Generating an explore-prompt.md

What this file is

The explore agent (momentic ai explore diff over a git range, or momentic ai explore latest over the whole app) identifies changed user journeys and builds Momentic tests for them. Both accept extra instructions appended to the explorer's system prompt via two repeatable options:

  • --prompt <text> — inline text. Repeatable.
  • --prompt-file <path> — a file whose contents are appended. Repeatable.

Passing any --prompt or --prompt-file replaces the project's cloud custom prompt. The combined instructions are all --prompt-file contents followed by all --prompt values, each in the order passed — so multiple flags stack rather than one overriding the other.

explore-prompt.md is the conventional file passed to --prompt-file. It is repo-specific: it tells the explorer which apps exist in this repo, the URLs its tests must target, how to sign in, exactly where to save generated tests, and any quirks it must respect. It is NOT a place to restate general Momentic behavior — that already lives in the momentic-test skill.

Usage:

momentic ai explore diff --prompt-file ./explore-prompt.md

Gather the repo facts first

Do not write the prompt from memory. Inspect the repo and confirm each fact:

  1. Applications under test. Enumerate every frontend/app the generated tests will drive. For each, determine the URL the tests must hit. This is the served/deployed URL used at run time, which is often NOT the Vite/dev URL used while authoring locally. Cross-check momentic.config.yaml environments[].baseUrl and how each app is started in CI (the workflow that runs momentic ai explore diff).
  2. Authentication. How should the explorer sign in? Prefer a reusable login module — record its exact name and id. Note access assumptions (e.g. "this user can reach all parts of the app").
  3. Test placement. The folder each app's tests belong in, expressed relative to the workspace, including any mandatory path prefix. Make the constraints unambiguous.
  4. Throwaway / nested tests. Where tests created as a side effect of an explore run must be saved (usually the single gitignored momentic/junk/ folder). Confirm which path is gitignored.
  5. Quirks. Externally-persisted data that survives runs (must be randomized per run to avoid collisions), auth peculiarities, flaky areas, base-URL caching flags, etc.

Structure to produce

Write concise, imperative markdown. Recommended sections:

  • Applications and target URLs — one bullet per app: name — the URL tests must target — purpose — how it is served/run — the login module (name + id) if auth is required.
  • Journey → application mapping — how to map an in-scope diff to the app(s) it affects so the right editor/port is driven.
  • Test placement rules — the destination folder per app, with explicit HARD RULE callouts for any mandatory prefix. The explorer will otherwise save tests in the wrong place.
  • Throwaway / nested test rule — the gitignored junk folder that every nested test must land in, and that naming a file junk-* does NOT count.
  • Quirks to watch out for — the repo-specific gotchas gathered above.

Rules for a good prompt

  • Target URLs must be the URLs tests actually run against, never the authoring dev-server URLs.
  • Be explicit about path constraints and use HARD RULE: callouts — path mistakes are the most common and most costly explorer failure.
  • Keep it repo-specific and short. Omit anything already covered by general Momentic docs/skills.
  • Reference login modules by both name and id so the explorer can pin the exact module.

Reference template

Below is a placeholder-only skeleton with the sections a good prompt has. Fill in the bracketed values with your repo's real apps, URLs, modules, and folders — do not ship the placeholders.

This repo has <N> frontend application(s). When tests run, each app is served at
the URL below — target these, not the dev-server URLs used while authoring.

- <app-name> — <one-line purpose> at <https://url-tests-must-target>. Sign in
  first using the module named `<login-module-name>` (id: <login-module-id>).
  <access assumptions, e.g. this user can reach all parts of the app>.
- <app-name-2> — <purpose> at <https://url-2>. Tests MUST use <https://url-2>
  as their base URL.

When identifying changed user journeys, map each in-scope diff to the
application(s) above so the right app/port is driven.

Place new tests in the folder matching the app they drive. Every target path is
relative to your workspace and the `<mandatory-prefix>/` segment is MANDATORY —
never drop it:

- <app-name> journeys → <mandatory-prefix>/<app-folder>/.
- <app-name-2> journeys → <mandatory-prefix>/<app-2-folder>/.

HARD RULE — keep the `<mandatory-prefix>/` prefix on the authored test. Saving
to any path that omits it is a hard failure.

HARD RULE: any test created as a side effect of building another test is
throwaway and MUST be saved into the gitignored `<junk-folder>/` directory.
Naming a file `junk-*` does NOT count — only that directory is gitignored.

Some quirks to watch out for:

- <quirk, e.g. externally-synced data persists across runs — randomize any data
  you write each run so runs don't conflict>.

User Reviews (0)

Write a Review

Effect
Usability
Docs
Compatibility

No reviews yet

Statistics

Installs14.7K
Rating4.5 / 5.0
Version
Updated2026年9月2日
Comparisons1

User Rating

4.5(120)
5
37%
4
43%
3
13%
2
5%
1
2%

Rate this Skill

0.0

Compatible Platforms

🤖claude-code

Timeline

Created2026年7月26日
Last Updated2026年9月2日
🎁 Agent Knowledge Cards
Survey