EvoX ドキュメント
  • 製品
  • 料金
  • ドキュメント
  • マーケット
EvoMap に戻る
ログインサインアップ
Overview
Practice cases
Features
GenesCapsulesEvolver skillsEvolution events.gepx evolution archives
Configuration
Developers
Security
Administration
Overview
Practice cases
Features
GenesCapsulesEvolver skillsEvolution events.gepx evolution archives
Configuration
Developers
Security
Administration
EvoX ドキュメント/Product/Evolver skills

Evolver skills

Turn repeatedly verified experience into instructions an Agent can follow.

Skills describe workflows with inputs, steps, outputs, and acceptance checks. They are not conversation summaries or Gene files renamed to SKILL.md. Evolver can help extract strategies; an application-ready skill still needs explicit tools, permissions, and prerequisites.

When to create a skill

Skills help when similar tasks recur and a reliable method has emerged. A weekly project update can define source material, output structure, handling of missing information, and review requirements.

If an approach has succeeded only once, retain the result and limitations first. Test it with different inputs before claiming broad applicability.

What a skill should contain

PartWeekly-update example
TriggerA project update is needed from existing notes
InputsNotes, reporting period, and audience
StepsRead sources, verify provenance, organize findings, flag open questions
OutputProgress, blockers, and next steps without invented data
ValidationClaims are supported and dates match the reporting period
BoundariesDraft first; sending or publishing needs separate confirmation

This illustrates content structure, not a preinstalled skill.

Import skills into EvoX

After preparing and validating the skill, choose its source through Add on the Skills page. The illustration changes example state only; it does not read folders or install content. See Import experience for local scanning, ZIP, and GitHub import steps.

EvoX

Create skill also appears in the Add menu; this mock focuses on the three import paths.

Import

Review and install skills from a ZIP file or public ZIP URL.

Drop a skill ZIP here

You can also paste a ZIP file or public URL.

One archive supports multiple skills, up to 64 MB.

A skill package is not .gepx. Use Evolution archives to migrate evolution history.

From experience to a workflow

For a dependency upgrade, a Gene can state “identify dependency changes, then narrow the problem with failing tests.” A Capsule records affected versions, the fix, and passing checks from one upgrade. The skill should extract stable steps rather than copy all paths and logs from that project.

  1. Identify shared inputs across related outcomes: dependency list, upgrade scope, failing tests, and allowed folders.
  2. Write the sequence: inspect version changes → reproduce failure → make a minimal fix → run relevant tests → report remaining risks.
  3. Turn version numbers, project names, and personal paths into inputs. Define stop conditions for unavailable services or failures that cannot be reproduced.
  4. Try a different upgrade task and check that the workflow works without hidden context from the original project.

For example: “If the upgrade only affects test tooling, start with its related tests. Preserve logs when reproduction fails; do not change unrelated business logic.” That is a reusable instruction. A single successful Capsule remains evidence, not an automatically general skill.

Genes describe methods, Capsules retain outcomes, Skills instruct a host, and Recipes order assets. See the distillation guide for commands and arguments.

Verify that the skill transfers to another task

For the next upgrade, supply its new dependency scope, failure logs, and allowed folders without the previous task's full conversation. Have the skill reproduce the failure, fix it, and report diffs, validation commands, and results. Check that it finds the relevant files in the new project rather than copying old paths or version numbers.

Cover normal input, missing failure evidence, an unreproducible failure or unavailable external service, and a request outside the authorized folders. The skill should complete valid work, request missing evidence, stop with a clear limitation when needed, and respect scope. Revise failed steps and retest that case; successful import is not proof of usability.

Keep environment-specific fixes in the Capsule and reusable decisions and steps in the skill. This preserves the method rather than the previous task's incidental conditions.

Updates and rollback

Preserve the current skill and provenance before updating. Compare instructions and scripts rather than treating same-name replacement as an automatic upgrade. Expanded permissions, network access, or deletion steps require renewed review.

If an update fails, stop using it, restore the last verified version, and repeat the original checks. Do not assume a general recycle-bin recovery mechanism exists. See Skill versions.

Troubleshooting

  • Imported but unusable: Check dependencies, tools, and inputs, not just file presence.
  • Additional permissions requested: Instructions do not grant authorization.
  • Same name, different content: Compare provenance, versions, and actual contents before resolving conflicts.
  • Sharing with a team: Remove personal paths, secrets, and internal material; include requirements and validation steps.

Related docs

  • Skills and plugins
  • Skill review
前のページCapsules次のページEvolution events