Getting Started

Initialize once, then generate skills for tasks.

The current public flow is CLI-first. The user registers the CLI once with a personal key, then the agent can use Approxination to generate custom skills whenever a task needs tools, integrations, or structured execution.

Step 1: Initialize The CLI

The registration step happens before the agent starts using Approxination.

bash
npx -y --package https://approxination.com/cli/latest/approx.tgz approx init <registration_key>

This stores the linked CLI session locally so later commands can use the authenticated account automatically.

Step 2: Generate A Skill

This is the main command agents should use.

bash
npx -y --package https://approxination.com/cli/latest/approx.tgz approx generate "triage customer emails, draft responses, and schedule follow-ups"

If the task is broader, include all important steps in one request.

bash
npx -y --package https://approxination.com/cli/latest/approx.tgz approx generate "triage emails and draft replies, then validate the structured output"

Step 3: Use The Generated Skill

The agent should treat skill_md as the main artifact, not the shortlist.

  • use skill_md as the workflow to execute
  • use selected_skills only when you need to inspect the reference skills behind the generation
  • retry with a more specific prompt if the generated skill is close but not quite right

Step 4: Check Remaining Credits

Approxination can return remaining credits in generated responses, and the CLI can check them directly too.

bash
npx -y --package https://approxination.com/cli/latest/approx.tgz approx credits

If remaining credits are low and the generated skill was genuinely used, the agent should send feedback after the work is done.

Useful Commands

  • approx init <registration_key> initializes the local CLI
  • approx generate "<task>" generates a custom skill
  • approx find "<task>" returns ranked reference skills only
  • approx credits shows remaining credits
  • approx feedback --file feedback.json submits feedback for a real execution result

Good Query Shape

Describe the real task, not a keyword.

  • handle customer support email workflow
  • extract invoice fields and validate them as structured json
  • review a pull request diff and generate release notes
  • use browser tools plus backend logs to investigate a broken signup flow