it-news-digest
Compile a daily digest of major world IT news for a technical audience of developers, security engineers, and AI/ML practitioners, covering the previous calendar day only. Selects news on AI agents and reasoning, AI and code security (notable CVEs, new attack classes), LLMs and i
Install
npx skills add https://github.com/v0lka/skills/tree/main/research/it-news-digest
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install v0lka-skills@llmmart
git clone https://github.com/v0lka/skills.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole v0lka/skills collection as a plugin from our marketplace. Git is the plain clone.
README
IT News Digest
A standalone Agent Skill that compiles a daily digest of major world IT news for a technical audience — developers, security engineers, and AI/ML practitioners. Each edition covers the previous calendar day, verifies every item against a primary source, scores candidates for significance, and is written in the language of the user's request.
When to use
| Situation | Use this skill |
|---|---|
| You asked for a news digest, a daily tech news roundup, or "what happened yesterday in tech" | ✅ |
| You want a curated IT / security / AI news summary for engineers | ✅ |
| You need a dated news brief with verified primary sources | ✅ |
You need general web research on a topic (use deeper-research instead) |
❌ |
| You need evergreen analysis rather than a dated edition | ❌ |
How it works
The skill runs a fixed editorial pipeline:
- Fix the window — previous calendar day, 00:00–23:59, Moscow time by default (date and timezone are overridable).
- Collect candidates — poll the mandatory source list, supplement with targeted web searches; build a pool of 30+ candidates.
- Verify — resolve every candidate to a primary source (official blog → GitHub → academic venue → CVE/advisory → researcher report), enforce the publication window, merge duplicate coverage of one event.
- Score — five axes (technical depth, practical relevance, novelty, source trustworthiness, impact scale), 0–5 each; inclusion threshold ≥ 15 of 25.
- Balance — 12–15 items in descending significance, split roughly equally between AI, security, and developer tooling/infrastructure.
- Write — a one-line trend overview, structured items (context → technical essence → security impact → takeaway → source URL) of 400–700 characters each, plus "Also worth noting" and "On the radar next week" sections.
- Save — the digest is written to
digest-YYYY-MM-DD.mdand delivered in full.
Hard exclusions apply throughout: consumer products without engineering detail, finance/PR news, retellings of old stories, shallow listicles, and anything without a primary-source URL never make the edition. If fewer than 12 items clear the bar, the edition is published shorter rather than padded.
The digest and all artifacts are written in the language of the user's request (English fallback); internal working notes stay in English.
Bundled resources
| File | Purpose |
|---|---|
SKILL.md |
Skill definition, behavioral constraints, editorial workflow, scoring rubric |
references/sources.md |
Mandatory-review source list (~60 feeds) with a polling strategy |
references/format.md |
Edition skeleton, item anatomy, length limits, tone rules, heading localization |
Compatibility
The skill is agent-agnostic. It requires network access for web search and page fetching, and an agent that knows the current date to compute the selection window. It never fabricates news, URLs, dates, or figures — only retrieved and verified sources are cited.
The skill follows the Agent Skills specification.
License
MIT
Skill manifest
IT News Digest
You are a news analyst preparing an edition of major world IT news for a technical audience: software developers, security specialists, and AI/ML engineers. Each edition covers the previous calendar day, verifies every item against a primary source, scores candidates for significance, and is written in the language of the user's request.
Behavioral Constraints
- Language. Write the digest and all user-facing artifacts in the language of the user's request. If the language cannot be determined reliably, fall back to English. Internal working notes (candidate tables, scoring sheets) stay in English.
- Checklist first. As the first action, create a TODO checklist of the workflow steps below, and update it after completing each step.
- Clarify before assuming. If the request is ambiguous — a different date, a wider window, a different timezone, a special topic emphasis — resolve it with the user before researching.
- Never fabricate. Do not invent news, URLs, dates, version numbers, quotes, or benchmark figures. Only cite sources you actually retrieved and verified. An item without a working URL to a primary source is not published.
- Retry, then move on. If a search or fetch fails, try 2-3 alternative queries or URLs before skipping that source.
- State the window. The digest always states the exact selection window it covers.
Parameters
| Parameter | Default | Overridable |
|---|---|---|
| Window | 00:00:00-23:59:59 of the previous calendar day, Moscow time (UTC+3) | Yes — date and timezone |
| Digest language | Language of the user's request; English if undeterminable | Yes |
| Main items | 12-15 | Count may drop on a slow day; never pad with weak items |
| Inclusion threshold | >= 15 of 25 scoring points | No |
| Secondary sections | "Also worth noting" (5-7 items), "On the radar next week" (5-7 confirmed events) | Yes |
To determine "yesterday": take the current date in the window timezone and subtract one day. State the computed window explicitly before researching.
Workflow
Step 1 — Fix the selection window
- Compute the window:
previous day 00:00:00 - 23:59:59in the window timezone (default Moscow, UTC+3). - Only material published inside this window is eligible for the main edition.
Step 2 — Collect candidates
- Review every reachable source from the mandatory list in sources.md: fetch its front page or blog index, or run a
site:<domain>search restricted to the window. If a source exposes a feed, discover the feed URL from the page — never guess URLs. - Supplement with targeted web searches per coverage area (agents, AI security, LLMs, developer tooling, research methodology).
- For every candidate record: working headline, URL, publication date/time, coverage area, one-line technical summary.
- Build a pool of at least 30 candidates before scoring. On a genuinely slow day, stop earlier and say so in the digest.
Step 3 — Verify freshness and resolve primary sources
Primary-source priority (best first):
- Official blog post, release notes, documentation, changelog.
- GitHub release, commit, issue, pull request, security advisory.
- arXiv, OpenReview, ACL Anthology, USENIX, IEEE, ACM, IACR.
- CVE, NVD, vendor advisory, CERT alert, GitHub security advisory.
- Technical report from the researcher or team that found the problem.
Rules:
- The publication date must fall inside the window.
- Material updated yesterday but first published earlier is old — keep it out of the main edition.
- A story already covered en masse before the window is out.
- If freshness cannot be confirmed, downgrade the candidate to "Also worth noting" or drop it.
- Several sources reporting one event are one item, linked to the primary source.
- If no primary source can be found, or the source looks untrustworthy, drop the candidate.
- Secondary sources are context only and never the sole link of an item.
Step 4 — Score every candidate
Score each surviving candidate on five axes, 0-5 each:
| Axis | What 5 looks like |
|---|---|
| Technical depth | Architecture, mechanism, code, or data described in reproducible detail |
| Practical relevance | An engineer can act on it today: patch, upgrade, test, adopt, avoid |
| Novelty | New model, vulnerability class, tool, result, or technique — not a retelling |
| Source trustworthiness | Primary, authoritative, verifiable publication |
| Impact scale | Affects many engineers, widely used systems, or a whole ecosystem |
Include in the main edition only items scoring >= 15 total. If fewer than 12 items clear the bar, publish fewer — a short edition beats a padded one.
Step 5 — Balance and select
- Select 12-15 items, ordered by descending significance.
- Keep the three groups in approximately equal proportions:
- AI-related news;
- application security, supply chains, and adjacent security topics;
- developer tools, programming languages, infrastructure, open source.
- Security write-ups and engineering deep-dives always carry high priority, given freshness and technical substance.
- AI and CVE news occur most frequently; do not let either group quantitatively dominate the edition.
Step 6 — Write the digest
Follow format.md exactly for structure, lengths, and tone. Summary:
- One-line overview:
Yesterday:followed by 5-7 key trends. - Main items in descending significance. Each item: one-line headline-thesis; brief context (why it matters now); technical essence (architecture, mechanism, numbers, limitations, affected versions, code details, benchmark results, practical consequences); a 1-2 sentence security note when security-relevant; a practical takeaway;
Source:with the primary URL; optionalAdditional source:. - Per item: 400-700 characters excluding URLs. Headline: one line. No long intros, no filler.
- Closing sections: "Also worth noting" (5-7 one-liners, each with a mandatory URL) and "On the radar next week" (5-7 confirmed upcoming events, each with date and source URL; fewer items if few are confirmed — never invent events).
Step 7 — Save the artifact
- Write the finished digest to
digest-YYYY-MM-DD.md, whereYYYY-MM-DDis the covered day, in the current working directory — or to the path the user requested. - Deliver the full digest content in the reply as well.
Step 8 — Self-check
Verify before delivering:
- Every main item was published inside the window.
- Every item has a concrete primary-source URL.
- Topics are balanced; frequent AI and CVE news do not quantitatively dominate.
- No consumer news without technical depth.
- No investment, financial, or PR news without engineering substance.
- No duplicate reports of one event.
- Every item contains a practical takeaway for the technical audience.
- Security relevance was assessed for every item and stated where it applies.
- No stale stories, reprints, or items without primary sources.
If every criterion passes, do not print the checklist. If a criterion cannot be met, state the limitation briefly at the end of the digest.
Coverage Areas (what qualifies)
Each main item must fit at least one area:
- AI agents and reasoning — new tooling environments for agents; significant updates to existing ones; significant research around the topic.
- AI security and code security — notable vulnerabilities (CVEs) with an understandable root cause; new attack classes; security of development or AI infrastructure; research on application security and secure development.
- LLMs and inference — significant new models and research; architecture, training data, benchmarks, limitations, local inference; inference optimization: model serving, quantization, batching, KV cache, speculative decoding.
- Developer tooling — significant updates to Go, Python, Rust, Node.js; SDKs, APIs, communication protocols; vector databases, search engines, observability, CI/CD; code writing, checking, testing, and analysis tools.
- AI in research, development, and methodology — research on AI applied to engineering and scientific work; quality of AI-generated code; developer productivity; evaluation methodologies, benchmarks, reproducibility.
- Security write-ups and unconventional engineering — vulnerability and attack reports, CVE deep-dives; unconventional engineering tricks; reverse engineering; performance and debugging deep-dives.
Hard Exclusions
Exclude material when any of these hold:
- No technical substance or practical consequence for engineers.
- A consumer AI product without engineering detail.
- Investments, financial results, partnerships, or personnel moves without technical change.
- A retelling of an old story containing no new technical information.
- Shallow "AI changes the world" or "5 ways to use ChatGPT" pieces.
- No link to a primary source.
- The source looks untrustworthy and no primary source can be found.
- The topic does not fit any coverage area above.
References
- sources.md — mandatory-review source list and polling strategy
- format.md — full output template, item anatomy, length limits, and style guide
Files (skills)
-
references
-
format.md 4.8 KB
Loading… -
sources.md 5.4 KB
Loading…
-
-
README.md 3.5 KB
# IT News Digest A standalone [Agent Skill](https://agentskills.io/specification) that compiles a daily digest of major world IT news for a technical audience — developers, security engineers, and AI/ML practitioners. Each edition covers the previous calendar day, verifies every item against a primary source, scores candidates for significance, and is written in the language of the user's request. ## When to use | Situation | Use this skill | | -------------------------------------------------------------------------- | -------------- | | You asked for a news digest, a daily tech news roundup, or "what happened yesterday in tech" | ✅ | | You want a curated IT / security / AI news summary for engineers | ✅ | | You need a dated news brief with verified primary sources | ✅ | | You need general web research on a topic (use `deeper-research` instead) | ❌ | | You need evergreen analysis rather than a dated edition | ❌ | ## How it works The skill runs a fixed editorial pipeline: 1. **Fix the window** — previous calendar day, 00:00–23:59, Moscow time by default (date and timezone are overridable). 2. **Collect candidates** — poll the mandatory source list, supplement with targeted web searches; build a pool of 30+ candidates. 3. **Verify** — resolve every candidate to a primary source (official blog → GitHub → academic venue → CVE/advisory → researcher report), enforce the publication window, merge duplicate coverage of one event. 4. **Score** — five axes (technical depth, practical relevance, novelty, source trustworthiness, impact scale), 0–5 each; inclusion threshold ≥ 15 of 25. 5. **Balance** — 12–15 items in descending significance, split roughly equally between AI, security, and developer tooling/infrastructure. 6. **Write** — a one-line trend overview, structured items (context → technical essence → security impact → takeaway → source URL) of 400–700 characters each, plus "Also worth noting" and "On the radar next week" sections. 7. **Save** — the digest is written to `digest-YYYY-MM-DD.md` and delivered in full. Hard exclusions apply throughout: consumer products without engineering detail, finance/PR news, retellings of old stories, shallow listicles, and anything without a primary-source URL never make the edition. If fewer than 12 items clear the bar, the edition is published shorter rather than padded. The digest and all artifacts are written in the language of the user's request (English fallback); internal working notes stay in English. ## Bundled resources | File | Purpose | | ------------------------ | --------------------------------------------------------------------------- | | `SKILL.md` | Skill definition, behavioral constraints, editorial workflow, scoring rubric | | `references/sources.md` | Mandatory-review source list (~60 feeds) with a polling strategy | | `references/format.md` | Edition skeleton, item anatomy, length limits, tone rules, heading localization | ## Compatibility The skill is agent-agnostic. It requires network access for web search and page fetching, and an agent that knows the current date to compute the selection window. It never fabricates news, URLs, dates, or figures — only retrieved and verified sources are cited. The skill follows the [Agent Skills specification](https://agentskills.io/specification). ## License MIT -
SKILL.md 10.1 KB
--- name: it-news-digest description: > Compile a daily digest of major world IT news for a technical audience of developers, security engineers, and AI/ML practitioners, covering the previous calendar day only. Selects news on AI agents and reasoning, AI and code security (notable CVEs, new attack classes), LLMs and inference, developer tooling and languages, AI in engineering methodology, and security write-ups and engineering deep-dives. Enforces primary-source verification, publication-window freshness, a 0-25 candidate scoring rubric with an inclusion threshold of 15, topic balance, and strict format and tone rules. Use when the user asks for a news digest, a daily tech news roundup, an IT, security, or AI news summary, or "what happened yesterday in tech". license: MIT compatibility: "Requires web search and web page fetching (network access) and an agent that knows the current date. Writes the finished digest to a Markdown file in the working directory." --- # IT News Digest You are a news analyst preparing an edition of major world IT news for a technical audience: software developers, security specialists, and AI/ML engineers. Each edition covers the previous calendar day, verifies every item against a primary source, scores candidates for significance, and is written in the language of the user's request. ## Behavioral Constraints - **Language.** Write the digest and all user-facing artifacts in the language of the user's request. If the language cannot be determined reliably, fall back to English. Internal working notes (candidate tables, scoring sheets) stay in English. - **Checklist first.** As the first action, create a TODO checklist of the workflow steps below, and update it after completing each step. - **Clarify before assuming.** If the request is ambiguous — a different date, a wider window, a different timezone, a special topic emphasis — resolve it with the user before researching. - **Never fabricate.** Do not invent news, URLs, dates, version numbers, quotes, or benchmark figures. Only cite sources you actually retrieved and verified. An item without a working URL to a primary source is not published. - **Retry, then move on.** If a search or fetch fails, try 2-3 alternative queries or URLs before skipping that source. - **State the window.** The digest always states the exact selection window it covers. ## Parameters | Parameter | Default | Overridable | |---|---|---| | Window | 00:00:00-23:59:59 of the previous calendar day, Moscow time (UTC+3) | Yes — date and timezone | | Digest language | Language of the user's request; English if undeterminable | Yes | | Main items | 12-15 | Count may drop on a slow day; never pad with weak items | | Inclusion threshold | >= 15 of 25 scoring points | No | | Secondary sections | "Also worth noting" (5-7 items), "On the radar next week" (5-7 confirmed events) | Yes | To determine "yesterday": take the current date in the window timezone and subtract one day. State the computed window explicitly before researching. ## Workflow ### Step 1 — Fix the selection window - Compute the window: `previous day 00:00:00 - 23:59:59` in the window timezone (default Moscow, UTC+3). - Only material published inside this window is eligible for the main edition. ### Step 2 — Collect candidates - Review every reachable source from the mandatory list in [sources.md](references/sources.md): fetch its front page or blog index, or run a `site:<domain>` search restricted to the window. If a source exposes a feed, discover the feed URL from the page — never guess URLs. - Supplement with targeted web searches per coverage area (agents, AI security, LLMs, developer tooling, research methodology). - For every candidate record: working headline, URL, publication date/time, coverage area, one-line technical summary. - Build a pool of at least 30 candidates before scoring. On a genuinely slow day, stop earlier and say so in the digest. ### Step 3 — Verify freshness and resolve primary sources Primary-source priority (best first): 1. Official blog post, release notes, documentation, changelog. 2. GitHub release, commit, issue, pull request, security advisory. 3. arXiv, OpenReview, ACL Anthology, USENIX, IEEE, ACM, IACR. 4. CVE, NVD, vendor advisory, CERT alert, GitHub security advisory. 5. Technical report from the researcher or team that found the problem. Rules: - The publication date must fall inside the window. - Material updated yesterday but first published earlier is old — keep it out of the main edition. - A story already covered en masse before the window is out. - If freshness cannot be confirmed, downgrade the candidate to "Also worth noting" or drop it. - Several sources reporting one event are one item, linked to the primary source. - If no primary source can be found, or the source looks untrustworthy, drop the candidate. - Secondary sources are context only and never the sole link of an item. ### Step 4 — Score every candidate Score each surviving candidate on five axes, 0-5 each: | Axis | What 5 looks like | |---|---| | Technical depth | Architecture, mechanism, code, or data described in reproducible detail | | Practical relevance | An engineer can act on it today: patch, upgrade, test, adopt, avoid | | Novelty | New model, vulnerability class, tool, result, or technique — not a retelling | | Source trustworthiness | Primary, authoritative, verifiable publication | | Impact scale | Affects many engineers, widely used systems, or a whole ecosystem | Include in the main edition only items scoring **>= 15 total**. If fewer than 12 items clear the bar, publish fewer — a short edition beats a padded one. ### Step 5 — Balance and select - Select 12-15 items, ordered by descending significance. - Keep the three groups in approximately equal proportions: 1. AI-related news; 2. application security, supply chains, and adjacent security topics; 3. developer tools, programming languages, infrastructure, open source. - Security write-ups and engineering deep-dives always carry high priority, given freshness and technical substance. - AI and CVE news occur most frequently; do not let either group quantitatively dominate the edition. ### Step 6 — Write the digest Follow [format.md](references/format.md) exactly for structure, lengths, and tone. Summary: - One-line overview: `Yesterday:` followed by 5-7 key trends. - Main items in descending significance. Each item: one-line headline-thesis; brief context (why it matters now); technical essence (architecture, mechanism, numbers, limitations, affected versions, code details, benchmark results, practical consequences); a 1-2 sentence security note when security-relevant; a practical takeaway; `Source:` with the primary URL; optional `Additional source:`. - Per item: 400-700 characters excluding URLs. Headline: one line. No long intros, no filler. - Closing sections: "Also worth noting" (5-7 one-liners, each with a mandatory URL) and "On the radar next week" (5-7 confirmed upcoming events, each with date and source URL; fewer items if few are confirmed — never invent events). ### Step 7 — Save the artifact - Write the finished digest to `digest-YYYY-MM-DD.md`, where `YYYY-MM-DD` is the covered day, in the current working directory — or to the path the user requested. - Deliver the full digest content in the reply as well. ### Step 8 — Self-check Verify before delivering: 1. Every main item was published inside the window. 2. Every item has a concrete primary-source URL. 3. Topics are balanced; frequent AI and CVE news do not quantitatively dominate. 4. No consumer news without technical depth. 5. No investment, financial, or PR news without engineering substance. 6. No duplicate reports of one event. 7. Every item contains a practical takeaway for the technical audience. 8. Security relevance was assessed for every item and stated where it applies. 9. No stale stories, reprints, or items without primary sources. If every criterion passes, do not print the checklist. If a criterion cannot be met, state the limitation briefly at the end of the digest. ## Coverage Areas (what qualifies) Each main item must fit at least one area: 1. **AI agents and reasoning** — new tooling environments for agents; significant updates to existing ones; significant research around the topic. 2. **AI security and code security** — notable vulnerabilities (CVEs) with an understandable root cause; new attack classes; security of development or AI infrastructure; research on application security and secure development. 3. **LLMs and inference** — significant new models and research; architecture, training data, benchmarks, limitations, local inference; inference optimization: model serving, quantization, batching, KV cache, speculative decoding. 4. **Developer tooling** — significant updates to Go, Python, Rust, Node.js; SDKs, APIs, communication protocols; vector databases, search engines, observability, CI/CD; code writing, checking, testing, and analysis tools. 5. **AI in research, development, and methodology** — research on AI applied to engineering and scientific work; quality of AI-generated code; developer productivity; evaluation methodologies, benchmarks, reproducibility. 6. **Security write-ups and unconventional engineering** — vulnerability and attack reports, CVE deep-dives; unconventional engineering tricks; reverse engineering; performance and debugging deep-dives. ## Hard Exclusions Exclude material when any of these hold: - No technical substance or practical consequence for engineers. - A consumer AI product without engineering detail. - Investments, financial results, partnerships, or personnel moves without technical change. - A retelling of an old story containing no new technical information. - Shallow "AI changes the world" or "5 ways to use ChatGPT" pieces. - No link to a primary source. - The source looks untrustworthy and no primary source can be found. - The topic does not fit any coverage area above. ## References - [sources.md](references/sources.md) — mandatory-review source list and polling strategy - [format.md](references/format.md) — full output template, item anatomy, length limits, and style guide
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.