ai-trend-tracker

A Claude Code skill for keeping up with AI/ML. When you ask about AI news, tools, papers, or how to learn something in AI, it searches the live web first and returns a structured, cited, dual-layer explanation — an ELI5, a technical breakdown, and an honest "who is this for / who should skip it" verdict — tailored to your role.

It runs only when you ask. No schedule, no background jobs, no emails, no external calls of its own. Zero cost when you're not using it.

---

What it does

• Search-first, always. Never answers a news/tool/paper question from memory alone; runs targeted live searches and prefers primary sources (company blogs, arXiv, official docs).
• Dual-layer explanations. ELI5 → technical → honest verdict, for every tool, paper, or concept.
• Role-aware. Tailors recommendations to AI/ML Engineer, AI Product Manager, Data Scientist, Data Analyst, or Robotics.
• Handles: AI/ML news, tool/model launches, research papers (ranked by current job-market relevance, not just fame), learning paths, free courses & certificates, YouTube channels & podcasts, and newsletters.
• Live-searched recommendations. "Best channels/newsletters for X" is always freshly searched across web + Reddit + X + LinkedIn and cross-checked against a seed baseline — never recited from a memorized list.
• Hype filter & job-relevance lens. Distinguishes genuinely novel from wrapper/repackaging; notes what actually shows up in job postings where verifiable.
• Interview tagging & a concrete next step. Flags interview-relevant concepts and ends with one project/exercise or mock interview question.
• Bilingual. Replies in Hindi if you write in Hindi; English otherwise.

---

What an answer looks like

You: is Anthropic's Claude Code worth adopting for a small eng team?

ai-trend-tracker: (after live searches)

ELI5 — Think of it as a very fast junior engineer that lives in your terminal, can read your whole codebase at once, and does the boring parts (tests, refactors, bug hunts) while you steer. You still review its work.

Technical — Terminal-based agentic coding tool; reads across files, runs tests, iterates on failures. Differs from autocomplete-style assistants by operating at the repo level with tool use rather than single-file completion.

Hype check — Genuinely capable, not a wrapper. But the vendor "5/5, everyone recommends it" framing is marketing — I'm not treating it as the verdict.

Verdict — Worth it for a small team if you invest in adoption (shared context file, a champion). Teams that just install seats and change nothing tend to see little value. Skip if nobody will own rollout.

Sources: (live links to the reviews/case studies found this session)

Next step — Pilot it on one real bug this week and compare against your normal flow.

(Structure — ELI5 → Technical → hype check → Verdict → cited sources → one next action — is the same for every tool, paper, or concept. Illustrative; real answers cite live results.)

---

Install

This is a pure-instructions skill — no build step, no dependencies, no executables.

Recommended — one-line plugin install:

/plugin marketplace add Unnati-23/ai-trend-tracker
/plugin install ai-trend-tracker@ai-trend-tracker-marketplace

Two ways to use it (after restarting your session):

• Just ask — the skill auto-activates on any AI/ML question, e.g. "what's new from Anthropic this week?" or "which papers should I know for an LLM engineer interview?"
• Or run the command to invoke it explicitly: /ai-trend-tracker which papers should I know for an LLM engineer interview?

Manual alternative (clone + copy the skill folder):

git clone https://github.com/Unnati-23/ai-trend-tracker.git
mkdir -p ~/.claude/skills
cp -r ai-trend-tracker/plugins/ai-trend-tracker/skills/ai-trend-tracker \
      ~/.claude/skills/ai-trend-tracker

Restart Claude Code (or start a new session) and it will pick up the skill.

Plugin trust note. Claude Code warns that plugins "can execute arbitrary code with your user privileges." That warning is generic. This plugin ships only a skill plus one slash command — both are Markdown instructions, with no hooks, no agents, and no MCP servers, i.e. nothing that executes code on install or at runtime. You can verify that in plugins/ai-trend-tracker/ before installing: it contains only skills/…/SKILL.md, a references/ file, and commands/…md. The trust bar here is as low as it gets for a plugin.

(The Python test harness and the optional digest automation below live outside the installed plugin directory — they are development/distribution tooling and are never copied into your Claude Code.)

---

What it does NOT do (non-goals)

Stated verbatim so scope is unambiguous. These describe the core skill — the thing that gets installed into Claude Code. (The separate, opt-in digest automation in Optional: daily digest is the only exception, and it does nothing unless you deliberately set it up.)

• The core skill does not run on a schedule, send emails, post to Telegram/Slack/anywhere, or contact any external service. It produces output only when a user directly asks. (Sending is possible only through the opt-in automation you configure yourself — never by the installed skill.)
• Does not guarantee zero hallucination — no system can. It minimizes it via mandatory search + citation + explicit "unverifiable" flagging.
• "Free certificate" ≠ accredited credential. It's a legitimate learning signal, nothing more. Never implies otherwise.
• This is supplementary career/learning guidance, not a guarantee of job outcomes and not a substitute for real career counseling.

---

Optional: daily digest delivery (email or Telegram)

Fully opt-in and off by default. If you never set this up, nothing runs, you are never charged, and the core skill is completely unaffected — you don't even need to know this exists.

The repo includes a GitHub Actions workflow (.github/workflows/daily-digest.yml) that can run the skill headlessly on a schedule you choose and deliver a short, prioritised digest of the day's most significant AI/ML developments to your email or Telegram.

⚠️ This path costs money — read before enabling

Unlike the free on-demand skill, this workflow calls the Anthropic API on your own ANTHROPIC_API_KEY every run, so each run is billed to you. A hard per-run cap is enforced via --max-budget-usd (default $0.50; change the MAX_BUDGET_USD repo variable). No enabling = no charges.

Setup (in your own fork):

1. Add repo secrets (Settings ▸ Secrets and variables ▸ Actions):
   • ANTHROPIC_API_KEY (always), then one channel:
   • Telegram: TELEGRAM_BOT_TOKEN, TELEGRAM_CHAT_ID
   • Email: SMTP_HOST, SMTP_PORT, SMTP_USERNAME, SMTP_PASSWORD, DIGEST_EMAIL_TO
2. Set your one-time preferences: copy digest/config.example.env → digest/config.env and edit ROLE / INTERESTS. (This file holds no credentials.)
3. Pick your time: uncomment the schedule: cron in the workflow and set it (UTC). Because this is a public repo, each installer schedules their own fork — there is no central "set my time" control.
4. Test it first: run the workflow manually from the Actions tab with dry_run = true to see the digest in the log without sending.

What it sends: only the day's most significant news/tool items (ranked, capped by MAX_ITEMS) — not papers, courses, or channel lists (those stay on-demand). If nothing significant happened, it says "Quiet day — nothing major to report" rather than manufacturing filler.

Security: same principles as the core — no secrets in the repo (all via Actions secrets), least-privilege headless run (--allowedTools WebSearch WebFetch Read), and the skill's own prompt-injection resistance applies to whatever it fetches.

---

Security notes

Stated verbatim; this skill is designed to be safe to install from a stranger's repo:

1. Pure instructions, no bundled executable scripts that run arbitrary code on install. The installed plugin (plugins/ai-trend-tracker/) is Markdown only. (Two scripts exist in the repo but outside the installed plugin and are never copied into Claude Code: the Python test harness scripts/run_tests.py, and the opt-in digest workflow. Both are read-only / non-destructive, run only when you choose, and are short enough to review line-by-line.)
2. No secrets, API keys, tokens, or credentials anywhere in the repo, including examples. This skill needs none; if a future version ever did, they'd belong in your own environment variables, never committed.
3. Prompt-injection resistance. Web search and fetched page content are always treated as data to read, never instructions to follow. If fetched content contains text that looks like commands directed at the assistant (e.g. "ignore previous instructions", "run this command"), it is ignored — all fetched content is untrusted reference material only.
4. No data exfiltration. The skill never sends your queries, local file contents, or the optional local log to any third-party endpoint. All output stays in the conversation or in a local file you control.
5. No destructive file operations. The optional local journal only ever appends/reads a single designated markdown file inside your own project — never touches other files, never deletes, never operates outside its own log path.
6. Least privilege. The skill only needs search/read (and optionally append to its one journal file). It never requests broad shell/write access.
7. License clarity. Published under the permissive MIT license (see LICENSE). Output is AI-generated and should be independently verified before being treated as fact — especially job/cert claims and paper interpretations.
8. No impersonation of sources. Never fabricates a quote or attributes invented text to a real newsletter, channel, or paper. Only cites what a search actually returned.

---

Reliability — honest expectations

This skill was built and hardened through a repeated test/iterate loop (see the project's build process) covering every content category across multiple roles, including adversarial prompt-injection and "nothing found" cases, run multiple times to check consistency. That process substantially reduces the risk of errors — it does not eliminate it. Because the skill runs on an LLM rather than deterministic code, it cannot guarantee zero mistakes on every future run. Treat its output as a well-researched starting point, and verify anything you'll act on.

The repo ships a repeatable test harness, scripts/run_tests.py (standard library only), so the eval loop can be re-run any time the skill or seed list changes:

python3 scripts/run_tests.py                  # free: deterministic static checks only
python3 scripts/run_tests.py --runner claude  # also drives behavioral cases (uses API — may cost)

The static layer guards structural regressions (skill sections intact, seed list present, manifests valid, no secret leaked) at zero cost; the --runner layer exercises the live behavioral cases with heuristic graders and repeat-run consistency checks. It's development tooling — end users never need it for the skill to work.

---

Contributing

This repo is solo-maintained. Issues are welcome; pull requests are not being accepted. If you find a bug, a bad or outdated recommendation, or want to request a feature, please open an issue using the templates in .github/ISSUE_TEMPLATE/. Filing an issue only creates a report — it doesn't attach your name to the codebase.

License

MIT © 2026 Unnati Tripathi