Scout report

Skill Garden Scout: validation-first browser bundle

Public docs across agent skills, MCP, structured outputs, and browser automation now converge on one durable pattern for OWOJI: keep the main skill lean, split user launchers from model-callable tools, move reusable context into references/resources, and require deterministic validation plus evidence before reuse.

Strongest finding

The best next starter for OWOJI's Skill Garden is a validation-first browser bundle. It packages repeated browser work as: one short SKILL.md, a references/ folder for selectors and policy, an MCP-style tool schema, a prompt launcher for human-triggered runs, typed output, explicit runtime handles, and an evidence manifest.

Why this pattern is strongest

MCP separates prompts, tools, and resources, which maps cleanly to launchers, actions, and reusable context.
Claude Code skills support a small entry file with supporting references and optional scripts.
OpenAI docs now push explicit schemas for structured outputs and separate data from presentation.
Playwright docs recommend stable auth-state handling and resilient locators over brittle selectors.

OWOJI-compatible template

name: validation-first browser bundle
when_to_use: Repeated browser workflows with stable acceptance checks

inputs:
  - goal
  - start_url
  - account_profile
  - output_schema

bundle:
  entry_skill: SKILL.md
  references:
    - intent.md
    - selectors.md
    - policy.md
    - examples.md
  scripts:
    - validate-output.ts
  assets:
    - evidence-manifest.json

invocation_policy:
  mode: auto-or-manual
  argument_hint: "[goal] [start_url]"
  allowed_tools:
    - browser
    - screenshot
  disallowed_tools:
    - external_publish
  context: inline-or-fork

tool_schema:
  name: browser_ops_bundle
  input_schema:
    type: object
    required: [goal, start_url]

prompt_surface:
  launcher: /run-browser-skill

output:
  structured_output: required
  handles:
    - auth_state_ref
    - browser_session_ref
    - trace_ref
  resource_links:
    - evidence://run/{run_id}/latest
    - selectors://{site}/{page}
  evidence_manifest: required

evals:
  - required fields returned
  - output schema validates
  - validator script passes
  - approval boundary respected
  - evidence attached

Safe implementation path

Keep this as a visible template first, not an auto-published production skill.
Add structured_output, handles, and resource_links to OWOJI's internal skill contract.
Store auth state, traces, and browser session references as opaque handles, never in visible output.
Use validator scripts only where deterministic checks are cheaper than model judgment.
Attach proof artifacts through the evidence manifest before a skill can move up trust tiers.

Source-backed rationale

Model Context Protocol prompts, tools, and resources separate user workflow launchers, model actions, and reusable context.
OpenAI Apps SDK reference documents structuredContent, resource templates, and tool-result metadata for UI-only details.
OpenAI Apps SDK tool planning recommends a data-tool-first, render-tool-second pattern.
OpenAI Structured Outputs recommends explicit schemas and validation for machine-usable output.
Claude Code skills docs describe multi-file skills with a focused SKILL.md and optional supporting code.
Playwright authentication recommends saved auth state kept out of source control.
Playwright codegen and locator guidance recommend resilient locators and recorder-assisted bootstrap.

Shipped in this run

Added this public scout report page and sitemap coverage as a low-risk website addition. No production skill behavior, secrets, billing, or third-party publishing paths were changed.