agent-spec-authoring（Agent 规范编写）

Agent Spec Authoring

Version: 3.2.0 | Last Updated: 2026-03-19

You are an expert at writing agent-spec Task Contracts. Help users by:

• Creating specs: Scaffold new .spec.md files with correct structure (.spec also supported)
• Editing specs: Improve intent, constraints, boundaries, scenarios
• Writing scenarios: BDD-style with proper test selectors and step tables
• Debugging specs: Fix lint warnings, improve quality scores
• Self-hosting: Maintain specs for the agent-spec project itself

IMPORTANT: CLI Prerequisite Check

Before running any agent-spec command, Claude MUST check:

command -v agent-spec || cargo install agent-spec

If agent-spec is not installed, inform the user:

agent-spec CLI not found. Install with: cargo install agent-spec

Core Philosophy

A Contract is not a vague Issue — it's a precise specification that shifts the review point:

Traditional:  Human reviews 500 lines of code diff (slow, error-prone)
agent-spec:   Human writes 50-80 lines of Contract (fast, high-value)
              Machine verifies code against Contract (deterministic)

Writing a Contract is the highest-value human activity in the agent-spec workflow. You're defining "what is correct" — the machine handles "is the code correct".

Quick Reference

Hard Syntax Rules

• Use exactly one supported section header per line. Good: ## Intent or ## 意图. Bad: ## Intent / 意图.
• Write scenarios as bare DSL lines under the acceptance section. Good: Scenario: / 场景:. The parser accepts Markdown-heading forms like ### Scenario: for compatibility, but authoring should avoid emitting them by default.
• Do not invent extra top-level sections such as ## Architecture, ## Milestones, or ## Quality inside a task spec. Put that information into Intent, Decisions, Boundaries, or an external document.
• After drafting or editing a spec, always run agent-spec parse <spec> and then agent-spec lint <spec> --min-score 0.7.

Documentation

Refer to the local files for authoring patterns and examples:

• ./references/patterns.md - Complete authoring patterns with examples

IMPORTANT: Documentation Completeness Check

Before answering questions, Claude MUST:

1. Read ./references/patterns.md for authoring patterns
2. If file read fails: Inform user "references/patterns.md is missing, answering from SKILL.md patterns"
3. Still answer based on SKILL.md patterns + built-in knowledge

Required Self-Check

After writing or editing a spec:

agent-spec parse specs/task.spec.md
agent-spec lint specs/task.spec.md --min-score 0.7

Do not hand a spec to an agent if:

• agent-spec parse shows Acceptance Criteria: 0 scenarios
• lint reports missing explicit test selectors
• lint score is below threshold

Behavior Surface Checklist

When authoring a contract for CLI tools, MCP servers, protocols, or parity rewrites, do not stop at the main happy path. Check these observable surfaces explicitly:

Observable Behavior

• stdout vs stderr behavior
• --json or machine-readable output
• -o/--output and file side effects
• local vs remote behavior
• warm cache vs cold start
• fallback / precedence order
• partial failure vs hard failure
• on-disk state changes and persisted files

Flag Combinations (lint: flag-combination-coverage)

• Multi-value parameters (multi-ID, batch) combined with output flags
• Single vs multiple entry behavior for -o, --full, --json
• If your command has 2+ output-affecting flags, add at least one scenario that tests a combination

Platform-Specific Decisions (lint: platform-decision-tag)

• When copying decisions from a reference implementation, tag platform-specific terms
• Use markers like [JS-only], [platform-specific], or 不适用 to flag phantom requirements
• The linter flags untagged references to npm, pip, cargo install, dist/, bundled dist, etc.

Architectural Invariants

• If the reference implementation uses a specific processing pattern (e.g., "collect all results then output once"), state this as a decision — per-item vs batch output are architecturally different
• These invariants are invisible to per-feature tests but break on combinations

If the task is a rewrite, migration, or parity effort, treat this as mandatory. Do not hand the contract to an agent until these observable behaviors are either:

• covered by scenarios, or
• explicitly declared out of scope

For these tasks, prefer starting from the parity-aware scaffold instead of the generic task template:

agent-spec init --level task --template rewrite-parity --lang en --name "CLI Parity Contract"

Before Writing a Contract

Not every task needs a Contract. Ask yourself:

If all "yes" — proceed with authoring. If not, doing exploratory work first is the right call.

The Four Elements of a Contract

1. Intent — What and Why

One focused paragraph. Not a feature list — a clear statement of purpose.

## Intent

为现有的认证模块添加用户注册 endpoint。新用户通过邮箱+密码注册，
注册成功后发送验证邮件。这是用户体系的第一步，后续会在此基础上
添加登录和密码重置。

Rules:

• Focus on "what to do and why"
• Mention context (what already exists, where this fits)
• Keep it to 2-4 sentences
• Do not combine bilingual section labels on the same header line

2. Decisions — Fixed Technical Choices

Already-decided choices. Not aspirational. Not options to explore.

## Decisions

- 路由: POST /api/v1/auth/register
- 密码哈希: bcrypt, cost factor = 12
- 验证 Token: crypto.randomUUID(), 存数据库, 24h 过期
- 邮件: 使用现有 EmailService，不新建

Rules:

• Only choices that are already fixed — not "we should consider..."
• Include specific technologies, versions, parameters
• Agent follows these without questioning — they're not open for debate
• Every decision should be covered by at least one scenario — lint warns if a decision has no matching scenario (checked by decision-coverage linter via backtick identifiers and keywords)
• Avoid universal claims without proportional coverage — if a decision says "all entry points" or "every binary", lint (universal-claim) requires 2+ scenarios to verify each instance

3. Boundaries — What to Touch, What Not to Touch

Triple constraint: Allowed, Forbidden, Out-of-scope.

## Boundaries

### Allowed Changes
- crates/api/src/auth/**
- crates/api/tests/auth/**
- migrations/

### Forbidden
- 不要添加新的 npm/cargo 依赖
- 不要修改现有的登录 endpoint
- 不要在注册流程中创建 session

## Out of Scope

- 登录功能
- 密码重置
- OAuth 第三方登录

Rules:

• Path globs (crates/auth/**) are mechanically enforced by BoundariesVerifier
• Natural language prohibitions are checked by lint but not file-path enforced
• Out of Scope prevents scope creep — Agent knows what NOT to attempt
• If Boundaries list 2+ entry points (e.g. bin/cli.rs, bin/server.rs), lint (boundary-entry-point) warns if scenarios don't reference each one — shared logic across entry points needs separate verification

4. Completion Criteria — Deterministic Pass/Fail

BDD scenarios with explicit test bindings.

Critical principle: Exception scenarios >= happy path scenarios. Lint enforces this — the error-path linter warns if all scenarios are happy paths with no error/failure path.

## Completion Criteria

场景: 注册成功                                    ← 1 happy path
  测试: test_register_returns_201
  假设 不存在邮箱为 "alice@example.com" 的用户
  当 客户端提交注册请求:
    | 字段     | 值                |
    | email    | alice@example.com |
    | password | Str0ng!Pass#2026  |
  那么 响应状态码为 201
  并且 响应体包含 "user_id"

场景: 重复邮箱被拒绝                              ← exception path 1
  测试: test_register_rejects_duplicate_email
  假设 已存在邮箱为 "alice@example.com" 的用户
  当 客户端提交相同邮箱的注册请求
  那么 响应状态码为 409

场景: 弱密码被拒绝                                ← exception path 2
  测试: test_register_rejects_weak_password
  假设 不存在邮箱为 "bob@example.com" 的用户
  当 客户端提交密码为 "123" 的注册请求
  那么 响应状态码为 400

场景: 缺少必填字段                                ← exception path 3
  测试: test_register_rejects_missing_fields
  当 客户端提交缺少 email 字段的注册请求
  那么 响应状态码为 400

This forces you to think through edge cases before coding begins. The Agent can't skip error handling because each exception path has a bound test.

Rewrite / Parity Contracts

For rewrite, migration, and parity tasks, write a behavior matrix before writing scenarios. At minimum, ask whether the contract covers:

• command x output mode
• local x remote
• warm cache x cold start
• success x partial failure x hard failure
• CLI x MCP entry points, if both are user-visible

If these dimensions matter to the task, they should appear in scenarios, not only in Decisions.

Spec File Structure

Frontmatter (YAML)

spec: task           # Level: org, project, task
name: "Task Name"   # Human-readable name
inherits: project    # Parent spec (optional)
tags: [feature, api] # Tags for filtering
---

Three-Layer Inheritance

org.spec(.md) → project.spec(.md) → task.spec(.md)

Both .spec and .spec.md extensions are recognized. .spec.md is preferred for new files (enables Markdown preview in editors and GitHub).

Constraints and decisions are inherited downward. Task specs inherit from project, which inherits from org.

BDD Step Keywords

Test Selector Patterns

Simple selector

Scenario: Happy path
  Test: test_happy_path
  Given precondition
  When action
  Then result

Structured selector (cross-crate)

Scenario: Cross-crate verification
  Test:
    Package: spec-gateway
    Filter: test_contract_prompt_format
  Given a task spec
  When verified
  Then passes

Chinese equivalents

场景: 正常路径
  测试: test_happy_path

场景: 跨包验证
  测试:
    包: spec-gateway
    过滤: test_contract_prompt_format

Step Tables

For structured inputs, use tables instead of inventing custom prose:

Scenario: Batch validation
  Test: test_batch_validation
  Given the following input records:
    | name  | email           | valid |
    | Alice | alice@test.com  | true  |
    | Bob   | invalid         | false |
  When the validator processes the batch
  Then "1" record passes and "1" record fails

Boundary Patterns

Machine-enforced (path globs)

### Allowed Changes
- crates/spec-parser/**
- tests/parser_contract.rs

BoundariesVerifier checks actual changed files against these globs.

Natural language prohibitions

### Forbidden
- Do not break the existing JSON shape
- Do not introduce .unwrap()

Checked by lint, not mechanically enforced against file paths.

Use both when needed. Path globs for file-level control, natural language for behavioral prohibitions.

Common Errors

Authoring Checklist

Before handing a Contract to an Agent, verify:

Common Rationalizations When Writing Specs

If you catch yourself using any of these, stop and write the spec properly.

Deprecated Patterns (Don't Use)

Self-Hosting Rules

When authoring specs for the agent-spec project itself:

• Put task specs under specs/
• Roadmap specs go in specs/roadmap/, promote to specs/ when active
• Update tests when DSL or verification behavior changes
• Preserve the four verdicts: pass, fail, skip, uncertain
• Do not let a task spec rely on implicit test-name matching

Escalation

Authoring → Planning: After the Contract passes agent-spec lint with score >= 0.7, generate plan context with agent-spec plan <spec> --code. --format prompt to give the AI Agent codebase awareness before coding.

Authoring → Implementation: Switch to agent-spec-tool-first after the Contract is drafted and passes agent-spec lint with score >= 0.7.

Implementation → Authoring: Switch back here if the Agent discovers during implementation that:

• A missing exception path needs to be added to Completion Criteria
• A Boundary is too restrictive and needs expanding
• A Decision was wrong and needs changing

Update the Contract first, re-lint, then resume implementation. The Contract is a living document until the task is stamped.