Hiring signal tracker

Hiring Signal Tracker

Collects job postings from listed tech companies' official ATS APIs, scores every job description against a tiered AI-skill dictionary, and accumulates a per-company time series of AI-investment signals for fundamental (value-investing) research. The output is research input, never a buy/sell recommendation.

Directory layout

hiring-signal-tracker/
├── SKILL.md              this file
├── README.md             user-facing docs (Chinese)
├── config.json           watchlist + dictionaries + classification rules
├── scripts/
│   ├── pipeline.py       collect → score → time series → HTML report (stdlib only)
│   └── probe.py          detect which ATS a new company uses
├── references/
│   └── methodology.md    investment logic, metric definitions, caveats — READ THIS
│                         before interpreting any numbers for the user
├── data/                 accumulating time series (jobs_*.csv, history.csv, state.json)
└── reports/              generated HTML reports

Core operations

Run a collection (the most common request — "跑一次采集 / 更新报告"):

cd <skill-root> && python3 scripts/pipeline.py

Python 3.8+ standard library only; no pip installs. Takes 1–3 minutes for ~27 companies. Same-day reruns overwrite that day's snapshot safely. After running, open reports/report_<date>.html for the user and summarize the notable changes (see "Interpreting results" below).

Add or replace companies ("把 X 加进股票池"):

1. Detect the company's ATS: python3 scripts/probe.py <guess1> <guess2> ... Try the company's short name, product name, and name without spaces (e.g. for HubSpot the working token was hubspotjobs, not hubspot).
2. If found, add an entry to companies in config.json with name, ticker, ats, token, and a theme (used for peer grouping in reports — keep peers together).
3. If all probes miss, the company likely uses Workday or a custom system — tell the user it cannot be tracked in v1 rather than guessing.
4. Run a collection to establish the company's baseline.

Companies known to be unreachable: Snowflake, CrowdStrike, Zscaler, Atlassian, UiPath (Workday/custom). Big-tech (Google/Meta/Microsoft/Amazon) is deliberately excluded: thousands of postings dilute the signal.

Set up recurring collection ("每周自动采集"): prefer the host's scheduling feature if present. Otherwise offer a cron entry (weekly, Monday morning):

0 9 * * 1 cd <skill-root> && /usr/bin/python3 scripts/pipeline.py

Report delivery by email requires the user's own connected mail tool and their explicit approval each time — never wire credentials into scripts.

Edit the skill dictionary (config.json skill_categories): after any dictionary change, warn the user that density scores are no longer comparable with earlier snapshots; offer to archive data/history.csv into data/archive/ and restart the baseline.

Interpreting results

Read references/methodology.md before writing analysis for the user. The non-negotiable rules it explains:

• Within-company time series beats cross-company levels. JD writing styles differ by company and ATS; a company's change against its own baseline is the signal. Cross-company comparison is only meaningful within the same theme.
• Signal quality ranks by role type: demand-validated roles (FDE, solutions, sales engineering — staffed against real contracts) > product engineering > research. Non-tech AI penetration (AI skills required in sales/finance/ops roles) marks AI moving from lab to business infrastructure.
• Negative signals are more reliable than positive ones. Companies fake expansion, never contraction. Lead with red flags (batch job removals) when present.
• Boilerplate suppression: a term hitting ≥90% of one company's JDs is marketing copy, not skill demand — the pipeline removes it automatically and lists it in the boilerplate_suppressed column. Mention it if a company's score shifted because of it.
• Hiring is necessary but not sufficient evidence of value creation: a cash-burning expander and a profitable compounder post the same jobs. Always push toward reconciliation (below) before drawing conclusions.

Advanced analysis (agent-run, not scripted)

These are part of the research method; run them with your own web/search tools when the user asks for deeper work.

B1 — Earnings reconciliation. For companies whose signals moved, pull the latest 10-Q/10-K from SEC EDGAR (https://www.sec.gov/cgi-bin/browse-edgar) and check whether revenue growth, RPO / deferred revenue, and headcount direction confirm or contradict the hiring signal. A rising demand-validated share with accelerating RPO is confirmation; with stagnant RPO it may be firefighting, not expansion.

B2 — Say-do gap. Compare management's AI narrative in the latest earnings call with the hiring data. Specific claims (customer counts, revenue contribution) + matching hires = credible. Vision language + no AI hiring = flag the inconsistency to the user; that gap is itself a research finding.

C — Composite momentum (needs ≥3 snapshots). Rank companies within theme on a blend of Δ total postings, Δ weighted density, Δ demand-validated share, and red-flag count. Present as a research-priority funnel ("deep-dive candidates"), never as a buy list.

Compliance boundaries

Only official, public, unauthenticated job-board APIs. Never scrape LinkedIn, Glassdoor, or Indeed (ToS-prohibited), never bypass rate limits or robots.txt, and never present output as investment advice — it is one input into fundamental research done by a human analyst.