analyze-value

Analyze Value — Decision Value Analysis (Paper / Event / Technology / Product)

Input a paper URL/PDF/title, event description/news link, technology direction name, or specific product model/series, and execute progressively (locate → parallel collection → cross-synthesis), outputting a decision-oriented value analysis report. If the target cannot be found, exit early without wasting agent calls.

Difference from other skills: this skill focuses ondecision judgment— whether it is worth paying attention to/buying/investing in, why, and how to act.

When to Use

Paper analysis：

• "Is this paper worth reading" / "Analyze the value of this paper"
• "Paper value analysis https://arxiv.org/abs/..."
• "analyze value: Attention Is All You Need"

Event analysis：

• "Analyze the Fed rate cut" / "Is this wave of layoffs worth paying attention to"
• "analyze value: OpenAI releases GPT-5"
• "What impact does this policy/event have on me"

Technology value analysis：

• "Is lithography machine technology worth paying attention to" / "What stage is EUV technology at now"
• "analyze value: RISC-V technology" / "Analyze the value of embodied intelligence"
• Need to judge whether a technology direction is worth investing attention or resources in

Product value analysis：

• "Is the ASML NXE:3800E worth purchasing" / "Is the iPhone 16 Pro worth buying"
• "analyze value: Tesla Model 3" / "Analyze the value of this lithography machine"
• Need to judge whether a specific product (model or series) is worth buying/procuring/adopting

Does not trigger：

• In-depth academic notes → /analyze-paper
• Technology landscape research (how to use, learning path, ecosystem) → /analyze-tech

---

Phase 0: Input classification and target location (must be executed first)

⚠️ Key step: classification must be completed first before entering the corresponding process. It is forbidden to skip classification and directly execute paper location.

Step 1: Determine the analysis type (first step, cannot be skipped)

Based on the user input, determine whether it belongs topaper、event、technologyorproductanalysis.Default rule: inputs without clear paper/technology/product signals go to event analysis.

Signal	Determination
Has an arXiv/DOI/PDF link, explicitly mentions "paper" "paper" "this article"	→ {analysis_type} = paper → Phase 0A
Has brand + model/series name, "is it worth buying" "which to buy" "which model to choose" "procurement" "review"	→ {analysis_type} = product → Phase 0D
Technology field/direction/route, "XX technology" "XX direction" "XX trend", technology terms without brand or model	→ {analysis_type} = tech → Phase 0C
News link, event/policy/release/accident with a clear time point, "XX event" "rate hike" "rate cut" "layoffs"	→ {analysis_type} = event → Phase 0B
Truly unable to determine	Confirm with the user before continuing

Example：

• "ASML NXE:3800E" → product(brand + specific model)
• "Is the iPhone 16 Pro worth buying" → product(brand + model + "worth buying")
• "Tesla Model 3 series" → product(brand + product series)
• "ASML EUV lithography machine series" → product(brand + product series)
• "Yen rate hike event" → event(describes a macroeconomic event)
• "OpenAI releases GPT-5" → event(describes a product launch event)
• "Analyze the Fed rate cut" → event(describes a policy event)
• "Lithography machine technology" → tech(technology direction, no brand)
• "Is embodied intelligence worth paying attention to" → tech(technology field)
• "What stage is RISC-V technology at now" → tech(technology route)
• "arxiv.org/abs/2401.xxxxx" → paper(has an arXiv link)
• "Attention Is All You Need" → paper(explicit paper title)

Key distinguishing points：

• product vs tech: has brand + model/series → product; only a technology field name → tech. "ASML EUV lithography machine series" → product; "EUV lithography technology" → tech.
• product vs event: value judgment about the product itself → product; impact of the product launch as an event → event. "Is the iPhone 16 Pro worth buying" → product; "Apple releases iPhone 16" → event.
• tech vs event: an ongoing technology direction → tech; a specific event with a point in time → event.

After determination, enter the corresponding Phase 0 process.

---

Phase 0A: Paper positioning (only when analysis_type = paper)

If Step 1 determines it is an event,skip this section, and go directly to Phase 0B.

Based on user input, try to locate the paper:

Input type	Processing flow
URL	WebFetch to retrieve the page, extract title, authors, abstract, institution
PDF path	Read to read the PDF, extract title, authors, abstract
Title/keywords	WebSearch to search and locate, then WebFetch to get details

Early exit rules

If any of the following conditions is met,directly inform the user and end, and do not enter subsequent Phases:

• WebSearch still cannot locate a specific paper after 2 searches
• User input is too vague (such as "AI paper"), and a unique target cannot be determined
• The URL is inaccessible and there is no alternative source

At the end, output:未能定位到论文「{用户输入}」，请提供更具体的论文标题、arXiv 链接或 PDF 路径。

Obtain the original paper text (cannot be skipped)

Try in order until content containing the main text is obtained:

1. arXiv HTML(preferred):WebFetch https://arxiv.org/html/{id}
2. PDF path provided by the user：Read {pdf_path} read the full text
3. Conference/journal page：WebFetch {conference_url}
4. Zenodo / non-arXiv preprint: first WebFetch the homepage to get the file list, then WebFetch the PDF download URL (Zenodo format:https://zenodo.org/records/{id}/files/{filename}.pdf, read the filename from the files section of the page)

Store the original paper text in {paper_content}。

Confirm the corresponding author (must be completed, cannot be skipped)

From {paper_content} find the corresponding author annotation:

• Common markers:* † ‡ superscript,✉、Corresponding author, email footnote
• If the above markers are not found in the current {paper_content} ,you must additionally obtain the original PDF text(retry step 2 or 4 above), because the corresponding author annotation is usually only on the first page of the PDF

After confirmation, save to {corresponding_authors}; if the PDF truly cannot be obtained, record "未在可获取原文中找到（仅限 PDF 第一页）"，Do not infer from author order or search results。

Determine key variables

• {paper_title} — Paper title
• {authors} — Author list
• {corresponding_authors} — Corresponding author list (confirmed from annotations in the original paper, cannot be inferred)
• {institution} — Institution
• {year} — Year
• {venue} — Publication conference/journal (mark unpublished as preprint)
• {abstract} — Abstract
• {paper_content} — Content of the original paper
• {core_problem} — Core problem extracted from the abstract/Introduction (a few keywords)
• {lang} — Output language (default Chinese)
• {date} — Analysis date YYYY-MM-DD
• {name} — English filename (lowercase hyphenated, e.g. msa-memory-sparse-attention）

→ Enter the Phase 1-3 paper workflow.

---

Phase 0B: Event localization (analysis_type = event)

Based on user input, attempt to localize the event:

Input type	Processing workflow
News URL	WebFetch retrieves the page, extracts event title, time, subjects, and core content
Event description	WebSearch for localization (at least 2 times with different keywords), WebFetch to obtain 2-3 first-hand reports
Vague description	WebSearch to confirm the specific event, then continue

Early exit rules

If any of the following conditions is met,directly inform the user and end：

• WebSearch 2 times still cannot localize a specific event
• The event description is too vague (e.g. "recent news"), cannot determine a unique target
• All news sources are inaccessible

At the end, output:未能定位到事件「{用户输入}」，请提供更具体的事件描述、新闻链接或关键词。

Collect first-hand information (cannot be skipped)

1. WebSearch Search 2-3 times (different keyword combinations), collect first-hand reports and authoritative analysis
2. WebFetch Obtain at least 2 first-hand sources (official statements, authoritative media reports)
3. Extract the 5W basic elements of the event: Who / What / When / Where / Why

Save the collected first-hand information to {event_sources}。

Determine key variables

• {event_title} — Event title (summarized in one sentence)
• {event_time} — Event occurrence/expected occurrence time
• {event_status} — Status (occurred / ongoing / upcoming / rumored unconfirmed)
• {key_actors} — Key subjects (decision-makers, affected parties)
• {core_content} — Summary of the event's core content (within 200 characters)
• {event_sources} — Content of first-hand information sources
• {core_problem} — Core issues/contradictions involved in the event (a few keywords)
• {impact_scope} — Scope of impact (individual/industry/country/global)
• {lang} — Output language (default Chinese)
• {date} — Analysis date YYYY-MM-DD
• {name} — English filename (lowercase hyphenated, e.g. fed-rate-cut-2026）

→ Enter the Phase 1-3 event workflow.

---

Phase 0C: Technical topic scoping (only when analysis_type = tech)

If Step 1 determines it is paper or event,skip this section。

Granularity check

Technical analysis requires appropriate granularity. Too broad and it cannot give useful decision judgments; too narrow and it degenerates into a single product/paper analysis.

Granularity	Example	Processing
Too broad	"AI" "semiconductor"	Confirm the sub-direction with the user (e.g., "large model inference optimization" or "AI chips"?)
Appropriate	"EUV lithography machines" "RISC-V processors" "embodied intelligence"	Continue directly
Too narrow	A specific product/single paper	Suggest switching to paper or event analysis

Collect first-hand information (cannot be skipped)

1. WebSearch Search 2-3 times (different keyword combinations), collect technology reviews, industry reports, authoritative analyses
   • "{tech_title} technology roadmap status 2025 2026"
   • "{tech_title} technology landscape players"
   • "{tech_title} development history milestones breakthrough"
2. WebFetch Obtain at least 2 high-quality sources (industry reports, review articles, in-depth coverage by authoritative media)
3. Extract the basic outline of the technology: core problem, main routes, key players, current stage

Store the collected first-hand information in {tech_sources}。

Early exit rules

If any of the following conditions is met,inform the user directly and end：

• After 2 WebSearch searches, substantial information about the technology still cannot be found
• The technology is too niche and lacks sufficient public analysis sources

At the end, output:未能找到关于「{用户输入}」的足够信息，请确认技术名称或提供更多上下文。

Determine key variables

• {tech_title} — Technology name
• {tech_scope} — Analysis granularity (core technology / technology route / industry chain segment)
• {maturity_stage} — Preliminary maturity assessment (early research / engineering / mass production / mature)
• {key_players} — Key players (companies/research institutions/countries)
• {core_problem} — Core technical challenges (a few keywords)
• {tech_sources} — Collected first-hand information
• {impact_scope} — Scope of impact (niche field / industry / cross-industry / global)
• {lang} — Output language (default Chinese)
• {date} — Analysis date YYYY-MM-DD
• {name} — English file name (lowercase hyphenated, e.g., euv-lithography）

→ Enter the Phase 1-3 technical process.

---

Phase 0D: Product positioning (only when analysis_type = product)

If Step 1 determines it is paper, event, or tech,skip this section。

Granularity check

Granularity	Example	Handling
Specific model	"ASML NXE:3800E""iPhone 16 Pro"	Continue directly
Product series	"Tesla Model 3 series" "ASML EUV lithography machine series"	Continue, covering comparisons within the series
Too broad (brand level)	"ASML's products" "Apple phones"	Confirm the specific model or series with the user

Collect first-hand information (cannot be skipped)

1. WebSearch Search 2-3 times (different keyword combinations), collect official specifications, reviews, pricing
   • "{product_name} specifications specs official"
   • "{product_name} review 评测 对比"
   • "{product_name} price 价格 competitors alternatives"
2. WebFetch Obtain at least 2 high-quality sources (official product pages, authoritative reviews, industry analyses)
3. Extract core product information: specifications, pricing, competitors, target users

Store the collected first-hand information in {product_sources}。

Early exit rules

If any of the following conditions is met,Directly inform the user and end：

• After 2 WebSearch searches, still unable to find substantive information about the product
• The product has not been officially released and reliable leaked information is lacking

At the end, output:未能找到关于「{用户输入}」的足够信息，请确认产品名称或提供更多上下文。

Determine key variables

• {product_name} — Full product name (brand + model/series)
• {manufacturer} — Brand/manufacturer
• {product_category} — Category (used to determine the comparison scope, e.g., "EUV lithography machine", "flagship phone", "mid-size SUV")
• {product_tier} — Positioning (flagship/mid-range/entry-level or industry tier)
• {price_range} — Price range
• {release_status} — Status (on sale / announced but not on sale / rumored)
• {core_specs} — Core specification summary
• {product_sources} — First-hand information collected
• {lang} — Output language (default Chinese)
• {date} — Analysis date YYYY-MM-DD
• {name} — English filename (lowercase hyphenated, e.g., asml-nxe-3800e）

→ Enter the Phase 1-3 product workflow.

---

Phase 1-3: Load detailed workflow

Based on {analysis_type} load the corresponding detailed workflow:

• Paper analysis: Use the Read tool to load $CLAUDE_SKILL_DIR/analysis-workflow.md
• Event analysis: Use the Read tool to load $CLAUDE_SKILL_DIR/event-workflow.md
• Technical analysis: Use the Read tool to load $CLAUDE_SKILL_DIR/tech-workflow.md
• Product analysis: Use the Read tool to load $CLAUDE_SKILL_DIR/product-workflow.md

Execute in the order of Phase 1 (parallel collection) → Phase 2 (comprehensive analysis) → Phase 3 (output report). Pass all variables determined in Phase 0 to the subsequent workflow.

---

Error Handling

Situation	Handling
Paper/event/technology/product cannot be located	End directly, output a not-found notice, and do not enter Phase 1
Agent timeout/failure	Mark the corresponding section as "insufficient data"; the rest proceeds as usual
Unable to obtain the original paper text	Mark at the beginning of the report: "⚠️ Full text not obtained; the following is based on the abstract and search", and lower the confidence level
Insufficient information sources for event/technology/product	Mark at the beginning of the report: "⚠️ Limited first-hand information; the following is based on limited sources", and lower the confidence level
Insufficient information on new papers/recent events/emerging technologies/unreleased products	Merge related sections and mark "information is relatively new, data is limited"

Notes

• Decision judgments must be clear, not fence-sitting ("worth watching but note X" is better than "may be worth it or may not be")
• In event analysis, distinguish facts from predictions; predictions must be marked "analysis suggests"
• If an event has multiple parties' positions, all perspectives must be presented, without favoring either side
• In technical analysis, route comparisons must be based on objective indicators, avoiding subjective preference
• In product analysis, competitor comparisons must be based on verifiable specifications and prices, avoiding marketing rhetoric
• The output language follows {lang}, default Chinese