# Labeling instructions

You are labeling sentences taken from public software engineering job postings
(September 2026). Each item has: sid, company (board slug), title (job title),
heading (the section heading the sentence sat under, may be empty), sentence.

Question being studied: does the posting expect the ENGINEER to USE AI coding
tools (Cursor, Claude Code, Copilot, Codex, Windsurf, coding agents, "AI-assisted
development") in their own work? Distinguish that from the job being about
BUILDING AI products, and from company boilerplate.

Give each sentence exactly one label:

- usage_required: the candidate is expected to already use / be proficient or
  fluent with AI coding tools or AI-assisted development. Typical: under
  requirements, qualifications, "what you bring", "must have", "you have",
  "about you", or wording like "proficiency with", "active daily use",
  "experience using ... as part of your workflow". Also count "AI fluency:
  you use AI tools in your daily workflow" style lines in a requirements list.
- usage_preferred: same, but explicitly nice-to-have, bonus, plus, preferred,
  "ideal candidate also has", "not required".
- usage_expected: the company or role states that engineers use AI tools in
  their work, or that the hire will use them ("Use AI tools throughout
  development", "we expect all team members to engage with AI tools", "we are
  AI-first: we use AI to move faster", "we encourage AI-assisted development").
  An expectation or culture statement, not phrased as a qualification.
- tools_provided: the company provides the tools: access, budget, stipend,
  subscriptions, licenses ("access to Claude, Cursor and Codex", "$250/month
  tool budget", "access to frontier models").
- hiring_policy: rules or disclosures about AI in the hiring process itself:
  candidates may / may not use AI in interviews or applications, or the
  employer uses AI to screen. Includes NYC Local Law 144 notices.
- building_ai: the sentence is about building or shipping AI products, agents,
  LLM features, models, evals, MCP servers as the JOB'S DOMAIN, or experience
  building such things. Not about using coding tools to write code.
- context: company description, customers ("trusted by Cursor"), product
  marketing, team mission, benefits unrelated to AI tools, or anything else.

Rules:
- Judge the sentence with its heading. A sentence under "Nice to have" is
  usage_preferred even if it says "experience with". A sentence under
  "Minimum requirements" is usage_required even if softly worded.
- If a sentence mixes using coding tools and building AI ("experience with AI
  tooling including OpenAI, Claude, MCP, agent frameworks, vector databases"),
  it is building_ai unless it clearly says the tools are used to write code.
- "AI-native" or "AI-first" describing the COMPANY or its PRODUCT is context.
  Describing how ENGINEERS WORK ("we are an AI-native engineering org, everyone
  codes with agents") is usage_expected.
- Mentions of Claude / ChatGPT / Gemini / Codex as products the company builds
  or sells (Anthropic, OpenAI, Google postings) are context or building_ai.

Also fill tools_named: the coding tools the sentence names AS TOOLS THE ENGINEER
WOULD USE (not products built, not customers). Use these exact names when they
apply: Cursor, Claude Code, GitHub Copilot, Codex, Windsurf, Devin, Cline,
Aider, Kiro, Jules, Gemini CLI, Junie, OpenCode, Cody, Augment, Lovable, v0,
Bolt, Replit Agent, Zed, Warp, ChatGPT, Claude, Gemini, Copilot. Use
"Copilot" only when it is not clearly GitHub Copilot. Do not list MCP. Empty
list if none.

For hiring_policy also fill policy: "forbidden" (candidates must not use AI in
interviews/assessments), "allowed" (candidates may or are encouraged to use AI),
"employer_uses" (company uses AI to screen/review or discloses it), "other".

Output: write a JSON array to the output path given to you, one object per
input item, in the same order, with keys: sid, label, tools_named, policy (null
unless hiring_policy), note (at most 12 words on why, optional). Nothing else.
Do not skip items. Do not modify any other file.