Authoring an agent skill

Adding a new agent skill is the second flywheel beyond connectors. Skills sit on the agent platform’s money-touching surface, so we ask for pre-PR alignment before you write code.

Five-step flow

Open a New skill issue first

Use the new-skill issue template. Include slug, runtimes, scopes, policy template, consent summary, and rationale.

Read @glideco/skills-base

The SkillManifest Zod schema is the source of truth for what a skill manifest must contain — slug, runtime compat, required scopes, policy template, consent summary, trust tier, publisher.Closed vocabularies (SkillTrustTier, SkillRuntime, SkillCategory, SkillScope) are CODEOWNERS-protected. Adding a new entry requires @glideco/policy-engine to understand the same vocabulary.

Scaffold the package

packages/skills/<id>/
├── package.json
├── tsconfig.json
├── icon.svg
├── README.md
├── src/
│   ├── manifest.ts            // SkillManifest object
│   ├── policy.ts              // policy-template + per-tier presets
│   ├── index.ts               // re-exports
│   └── __tests__/
│       └── contract.test.ts

Or use the CLI:

glide partner submit ./my-skill --type=skill

The 4-tier policy preset pattern (every skill ships these):

export const POLICY_PRESETS = {
  supervise:  { perTxMaxUsdCents: 10_000, dailyCapUsdCents: 100_000, ... },
  rideAlong:  { perTxMaxUsdCents: 25_000, dailyCapUsdCents: 250_000, ... },
  trust:      { perTxMaxUsdCents: 50_000, dailyCapUsdCents: 500_000, ... }, // package default
  veteran:    { perTxMaxUsdCents: 100_000, dailyCapUsdCents: 1_000_000, ... },
} as const satisfies Record<string, SkillPolicyTemplate>;

Per-tier exposure must increase monotonically (supervise < rideAlong < trust < veteran). The contract test enforces this.

Write the contract test

Extend SkillContractTestSuite from @glideco/skills-base:

import { SkillContractTestSuite } from '@glideco/skills-base';
import { manifest } from '../manifest';

class MySkillSuite extends SkillContractTestSuite {
  readonly manifest = manifest;
}

The suite asserts:

Manifest validity vs SkillManifest v1.
Policy template internal consistency (perTxMax ≤ dailyCap; non-zero velocityCap.maxCount).
Scope sanity (no duplicates).
Consent under-disclosure guard. If requiredScopes includes a money-touching scope (payments:initiate, cards:manage, x402:pay), consentSummary MUST mention payments / money / cards / $ / usd.

Author the consent flow copy honestly

Per the OSS plan §M5 prompt-injection review: a community-tier skill template that under-discloses caps OR overrides displayed envelope behavior is a security finding, not a feature.The contract test catches the obvious cases. Reviewers catch subtler ones during the verified-tier promotion review.Example consent summary that passes the under-disclosure guard:

“Claude reads your QuickBooks Online invoices and drafts USD payments to your existing vendors, capped at $2,500 per transaction and $10,000 per day. Each payment requires your explicit approval before broadcast.”

Compare to one that wouldn’t pass:

“A friendly skill that helps you stay organized.” (no money words; under-discloses)

Reading list

Agent skills catalog — every skill currently in the repo.
Money-safety contracts — F-rules every skill inherits.
Threat model — T10 (malicious skill consent under-disclosure).

Get started

Concepts

Packages

Connectors

Agent skills

Headless MCP

Security

Standards

Legal

Authoring an agent skill

Five-step flow

Reading list