Agent Skills

Skills are workflow instructions that Agents can invoke.

Skill List

Skill	Default Install	When to Use
consumer-h5-harness	Yes	Master workflow for creating, connecting, implementing, and verifying Consumer H5 projects
fe-harness-create	No	Create Consumer H5 project from scratch, handle input collection and first task
fe-harness-init	No	Safely connect to existing projects, handle conflicts and legacy token discovery
fe-harness-inspect	No	Read project facts, inputs, tokens, verify modes, and Agent workflow states
fe-harness-plan	No	View create/init plans before writing files
fe-harness-doctor	No	Read-only diagnostics for Node, pnpm, scripts, pages, inputs, tokens, Agents, etc.
fe-harness-inputs	No	Register, check, and analyze PRD/RP/UI/API/assets inputs
fe-harness-task	No	Create task numbers, view history, and create immutable snapshots
fe-harness-verify	No	Choose quick/feature/visual/audit verification based on change risk
fe-harness-api	No	Check OpenAPI JSON, generate types and request wrappers by task operationIds
fe-harness-design-tokens	No	Extract, check, and diff unique Design Token ground truth
fe-harness-skills	No	List and install project-level or global Skills
fe-harness-version	No	Check if CLI is available and version matches

Default Installation Strategy

New projects by default only install:

consumer-h5-harness

Installation locations:

.agents/skills/consumer-h5-harness/
.claude/skills/consumer-h5-harness/

Codex and Cursor share .agents/skills, Claude Code uses .claude/skills.

If you need to install default aggregation Skill for all providers:

fe-harness skills install --project --provider all --name consumer-h5-harness

If a specific task requires specialized capability, install the corresponding Skill:

fe-harness skills install --project --name fe-harness-api
fe-harness skills install --project --name fe-harness-design-tokens

Why Only Install One by Default

Command-level Skills are useful, but should not become default context. Installing all command-level Skills by default would make Agents read irrelevant rules about API, Design Token, UI System, visual baseline, etc. during ordinary business tasks.

The aggregation Skill is responsible for stable workflows; command-level Skills are installed on demand through explicit commands:

fe-harness skills install --project --name fe-harness-api
fe-harness skills install --project --name fe-harness-design-tokens

Single Constraint Ontology

AGENTS.md is the only project constraint document. CLAUDE.md and Cursor rules only point to it.

Skills describe workflow procedures, do not replace project constraints, and do not duplicate complete constraint text. This reduces multi-provider rule drift.

What Humans Still Need to Do

Skills can make Agents more stable, but cannot replace human authority judgment. Humans still need to:

• Provide or confirm valid versions of PRD, RP, UI, API, and assets.
• Decide which source is valid when input conflicts occur.
• Confirm Design Token authoritative sources and user ad-hoc overrides.
• Choose which operationId is truly needed among multiple candidates.
• Decide on requirement deferral, external blocking, or acceptance exemption.
• Approve global installation, dependency installation, publishing, and other cross-boundary operations.