AGENT SKILL

deep-research

Use when the user asks for "deep research", a "research report on", a "comprehensive analysis of", to "look into X in depth", to "write a report on X", or any question needing multi-source synthesis beyond a single search — an autonomous multi-step deep research engine implementing an iterative Think → Search → Extract → Synthesize → Stop loop. The LLM drives every decision: what to search, what's relevant, what's missing, and when to stop. Produces a cited, magazine-quality report with inline citations, category-specific formatting, and research stats. For entity vetting/dossiers use entity-research; for news digests use news-monitoring; for source-grounded Q&A use notebooklm-mode.

License 许可
MIT
最近更新
2026-11-13
Stars
★ 123
Forks
22

01 前言 Front Matter

name
deep-research
license
MIT
metadata
{ "author": "moonlight-lupin", "hermes": { "tags": [ "research", "deep-research", "report", "synthesis", "iterative", "citations", "evidence-basis", "provenance" ], "related_skills": [ "news-monitoring", "entity-research", "notebooklm-mode" ] }, "version": "1.13.0", "platforms": [ "linux", "macos", "windows" ] }
description
Use when the user asks for "deep research", a "research report on", a "comprehensive analysis of", to "look into X in depth", to "write a report on X", or any question needing multi-source synthesis beyond a single search — an autonomous multi-step deep research engine implementing an iterative Think → Search → Extract → Synthesize → Stop loop. The LLM drives every decision: what to search, what's relevant, what's missing, and when to stop. Produces a cited, magazine-quality report with inline citations, category-specific formatting, and research stats. For entity vetting/dossiers use entity-research; for news digests use news-monitoring; for source-grounded Q&A use notebooklm-mode.

02 资源文件 26

03 SKILL.md 正文

Deep Research — Iterative Research Engine

An autonomous, multi-step research engine that performs exhaustive information gathering and synthesis. Unlike a single web_search, this skill implements an iterative loop where the agent plans, searches, extracts, synthesizes, and decides when to stop — producing a cited report with structured evidence, source quality tiers, and explicit gaps/contradictions sections.

See references/structured-evidence-format.md for the evidence.json layer and references/docsgpt-concepts.md for adaptive-depth concept mapping.

For the "is there updated or custom firmware for device X?" question class, follow references/firmware-support-status-research.md — vendor download page + official-forum RSS + GitHub-API triangulation recipe.

When to use

  • User asks for "deep research", "research report", "comprehensive analysis"
  • User wants a written report on a topic (not just a quick answer)
  • Question requires multi-source synthesis with citations
  • User says "look into X in depth" or "write a report on X"

When NOT to use

  • Entity vetting/dossiers → use entity-research (has sanctions screening, structured lenses)
  • Recurring news digests → use news-monitoring (has cron, dedup, multi-language)
  • Source-grounded Q&A from collected sources → use notebooklm-mode (has vault + RAG)
  • Quick factual question → just use web_search directly
  • Single-source extraction → use web_search + web_extract

Architecture

User question
  → Step 0: Clarification (optional — assess if question is too vague)
  → Step 1: Plan (sub-questions, key topics, success criteria, category, language, complexity)
  → Step 2: Date grounding + language anchor
  → Step 3: Iterative loop (max rounds = complexity cap: simple=2, moderate=3, complex=5)
      ├─ 3a: Generate gap-driven queries (incl. refute queries from round 2+, empty-search refinement)
      ├─ 3b: Search (web_search) + fetch (web_extract)
      ├─ 3c: Quality filter + extraction + source quality classification + numbered citations
      ├─ 3d: Synthesize into cumulative research state (with token budget check)
      ├─ 3e: Structured evidence (optional: evidence.json for 5+ sources)
      └─ 3f: Stopping check (LLM evaluates coverage + token budget)
  → Step 4: Final report (overview-first structure, see template)
  → Step 5: Stats summary

Report output order (overview-first): Executive Summary → comparison/overview table → detailed analysis per sub-question → contradictions → gaps → conclusion → source table. Readers get the answer and the at-a-glance comparison before the detailed reasoning.

Step 0 — Clarification (optional)

Arriving with a brief. If the user has already confirmed a task brief (the task-brief skill) for this research, it is the Step 0 output: carry its Goal, Done-when tests, Context, Constraints and Assumptions straight into the Step 1 plan and don't ask its questions again. Still run the referent checks below if the brief leaves a name or acronym ambiguous.

Before planning, assess whether the question is specific enough to research productively. This step prevents wasting rounds on a question that's too vague.

When to clarify:

  • The question is 1-5 words with no scope ("research AI", "tell me about quantum computing")
  • The question could mean very different things depending on intent ("is X good?" — good for what? for whom?)
  • The question lacks a clear angle ("look into climate change" — scientific? policy? economic?)

When NOT to clarify:

  • The question already has sub-questions or specific parameters
  • The user explicitly says "just research it" or "don't ask, just go"
  • The question is specific enough to generate productive search queries immediately
  • The question is a simple factual lookup ("what is the latest version of X?", "when did Y happen?") — even if short, if the intent is clear, proceed
  • The user is responding to a prior clarification (don't loop)

How: Present 1-3 short questions to the user that would narrow the scope. Keep it brief — this is a scope check, not an interview:

"Before I start researching, I'd like to clarify:

  1. [narrowing question 1]
  2. [narrowing question 2]

Please provide these details and I'll begin."

If the question is already specific, skip this step entirely and proceed to Step 1.

Proper-noun referent check (before Round 1): when the question names a term that could refer to either the user's own tooling/environment or a public product (a bot name, a codename, a vendor feature, a person), disambiguate the referent BEFORE running full rounds. If the user is available, ask once with the plausible readings as options; if the user does not answer, proceed on the best interpretation, say which one you chose in the plan, and let the user correct course cheaply. A wrong referent costs entire rounds of searching and a full refocus — it is the most expensive research mistake this loop can make, more so than any vague scope.

Acronym referent probe (cheap, before Round 1): the same check applies to short acronyms that collide with unrelated high-ranking terms — a bare "UI", "MAS", or "RAG" matches medical scales, entertainment companies, or storage tech at the same rank as the intended referent. Run a 2-3 query precision probe (one query per candidate reading) before committing rounds. If the probe is ambiguous, state the chosen reading in the plan header AND as an interpretation note at the top of the final report, so the user can correct course cheaply instead of re-running the loop.

Step 1 — Research Plan

Before searching, break the question into a research plan. Output:

## Research Plan
Question: [user's question]
Date: [current date]
Language: [BCP 47 tag — e.g. en, zh-Hans, ja. Detect from query; user's explicit language preference overrides. Use consistently throughout the report.]

Sub-questions:
1. [specific sub-question 1]
2. [specific sub-question 2]
3. [specific sub-question 3]
4. [up to 6 total]

Key topics:
- [topic 1]
- [topic 2]

Success criteria:
- [what would comprehensive coverage look like?]
- [minimum: each sub-question has ≥1 source]

Done when (2–4 checkable tests, one per line):
- [e.g. "each company row has a dated figure from its own disclosure, or says not found"]
- [e.g. "every deal value states its basis (offer / completion)"]

Load-bearing claims (the 2–4 facts the conclusion rests on):
- [claim] — who could observe it: [registry / regulator / company accounts / statistics body]
  — what would disprove it: [the finding that would overturn it]

Mode: [standard | comparison (items × fields)]   Recipe: [none | research-recipes/<name>.md]

Report category: [factual | comparison | product | how-to | fact-check | explainer]

Complexity: [simple | moderate | complex]

Done when. Turn "a good answer" into 2–4 tests someone could tick or cross on delivery — about the substance, not the format: the figures that must be current, the basis every amount must state, the question each section must answer. Swap vague words for checks ("latest" → the period; "key players" → the named list). They are run at Step 5.5. From a task brief, use its Done-when tests.

Load-bearing claims. Name the few facts the answer depends on, and for each the party that could actually observe it (the evidence route) and the finding that would disprove it. Round 1 queries go to those owners first — the registry, the regulator, the company's own accounts, the statistics series — and round 2's refute query tests the disprove condition.

Comparison mode and recipes. A "compare these N things on these M points" question runs as an items × fields grid whose outline the user approves before any search, with a search budget of at least one query per item plus two. Each item starts from its own latest disclosure, cells take the newest figure in the run, and queries off the grid are dropped. A repeated job can start from a saved recipe. Both: references/comparison-mode-and-recipes.md.

Complexity assessment and adaptive depth: Classify the question's complexity to determine the maximum number of research rounds. This prevents over-searching simple questions and under-searching complex ones.

Complexity Max rounds Per-round breadth When to use
simple 2 2 Single-factual question, narrow scope, 1-2 sub-questions. Example: "What is the latest version of Python?"
moderate 3 3 Multi-faceted question, 3-4 sub-questions, comparison or how-to. Example: "Compare Qdrant vs pgvector for RAG"
complex 5 4 Broad scope, 5-6 sub-questions, requires deep synthesis across domains. Example: "Comprehensive analysis of the self-hosting vs managed API tradeoff for AI infrastructure"

The complexity cap is a maximum, not a target. The stopping check (Step 3f) still governs early termination. A simple question that finds comprehensive answers in 1 round should stop at 1 — the cap of 2 means it can go to 2 if needed, not that it must.

Search budget (adapted from upstream research-input.ts presets + estimateMaxSearches): announce the budget in the plan — "Search budget: ≤N queries", N = per-round breadth × max rounds (script: estimate-budget --breadth B --rounds R). When round 2+ runs deep-dive chains (§3a), the upper bound is tree-mode: estimate-budget --breadth B --rounds R --halving, upstream's breadth-halving recursion. The budget is a ceiling the loop must not exceed; prefer the cheapest round that answers well, per the §3f stopping check.

Language anchoring: Detect the output language from the query and normalize to a BCP 47 tag (e.g. en, zh-Hans, zh-Hant, ja). Use it consistently throughout — executive summary, analysis, contradictions, gaps, conclusion. Source titles, URLs, proper nouns, and code may stay in their original language; search queries may use any language that helps evidence gathering. If the user explicitly switches language mid-research, update the anchor and use the new language for all subsequent output.

Report category detection:

  • "vs", "compare", "better than" → comparison
  • "best", "top", "recommend", "buy" → product
  • "how to", "guide", "steps" → how-to
  • "is it true", "verify", "fact check" → fact-check
  • "what is", "explain", "overview" → explainer
  • default → factual

Step 2 — Date Grounding

Inject the current date before any search. This is mandatory — LLMs default to training-cutoff years, producing stale queries.

Today's date is {current date as "DD Month YYYY"}. When a search query needs a year or refers to "latest"/"current"/"this year", use {current year} or relative wording — never a year inferred from training data.

Step 3 — Iterative Loop

Round structure

Each round follows: Query → Search → Extract → Synthesize → Check Stop

Run max N rounds where N = complexity cap from Step 1 (simple=2, moderate=3, complex=5). Most topics converge in 2-3 rounds. The cap is a ceiling, not a target.

Query rules, fetch budgets, quote anchoring, the synthesis template, the evidence store, and stop conditions are in references/step3-iterative-loop.md.

When you hit Read
Query generation, search and fetch, extraction, anchoring, synthesis, or the stopping check references/step3-iterative-loop.md
Evidence-store commands and the three delivery gates references/validation-gates.md
The evidence.json intermediate references/structured-evidence-format.md
Source families, shelf lives, the basis of money figures, user-supplied reports, date forms references/evidence-hygiene.md
A "compare N things on M points" question, or a repeated research job references/comparison-mode-and-recipes.md
The default search or reader fails, or the plan needs a country, language or date filter references/search-and-fetch-fallbacks.md
  1. 3a — Query Generation (gap-driven) — Breadth comes from the Step 1 preset (simple=2, moderate=3, complex=4). Round 1 targets plan sub-questions; later rounds target gaps and recorded follow-ups.
  2. 3b — Search and Fetch — Use the donsetch MCP search/fetch tools when the user asked for them; otherwise web_search / web_extract. Never abort over a provider preference. Optional, paid: when the plan needs the same public facts for several organisations (size, mandate, sector exposure, recent commitments), one Exa Agent API run can collect them as cited rows (about US$0.10 a run, ceiling US$1.10; quote first; organisations only, never people). Its output enters the evidence store like any other source — each fact keeps the page it cites and gets its own label. Procedure: entity-research/references/organisation-evidence-exa.md.
  3. 3b.1 — Evidence Persistence (mandatory for 5+ source reports) — From round 1, persist with init-run, register-source, add-claim, add-evidence, and check-quote on scripts/research_validation.py.
  4. 3c — Quality Filter and Extraction — Discard low-quality results before extracting.
  5. Verbatim snippet anchoring (mandatory for claims you store) — Every stored claim carries a verbatim contiguous excerpt from that source's page text.
  6. Evidence basis — the four-label discipline — Tag every material fact with the basis you are asserting it on.
  7. 3d — Synthesis — Fold this round's extractions into the cumulative research state.
  8. 3e — Structured Evidence (recommended for reports with 5+ sources) — Write the evidence.json intermediate before the final report. Format: references/structured-evidence-format.md.
  9. 3f — Stopping Check — After synthesis, decide whether the report is comprehensive enough. Otherwise open another round, up to the cap.

Step 4 — Final Report

Produce the final report. Minimum 800 words (scale with topic complexity).

Output order is overview-first: the reader gets the answer and the at-a-glance comparison before the detailed reasoning. Do not bury the comparison table after the per-topic analysis.

Carry each material fact's evidence-basis tag inline (§3c), lead on [VERIFIED] / [SOURCED], keep [REASONED] / [ESTIMATED] claims framed as indicative, and paste the Evidence key legend below the Sources table so the tags decode.

Findings, not process, in the body. The report reads as research on the subject: no claim IDs, store or gate mechanics, rebuilt stores, rewordings or tool names in the sections. Process facts a reader needs to weigh the evidence go in one plain sentence under Gaps ("figures rest on search summaries, not full page text"); everything else goes in the stats block and <run_dir>.

Synthesis prompt structure: When writing the final report, explicitly consult: (1) the original question, (2) the research plan (sub-questions, success criteria), (3) the cumulative research state from all rounds, (4) the evidence.json if produced, and (5) the numbered source list. Do not rely on the cumulative state alone if it has been compressed across many rounds — go back to the per-round findings to verify key claims. This prevents detail loss when the research state was aggressively summarized due to token budget pressure.

Structure

# [Report Title]

> **Research date:** [date] · **Rounds:** [N] · **Sources:** [N] · **Category:** [type]

## Executive Summary
[2-3 paragraph overview of key findings — the answer up front. If using structured evidence, note claim count and refute count here so the reader knows the evidence base.]

## [Comparison Table | Overview | Key Findings at a Glance]
[For comparison reports: a markdown table comparing options/entities across criteria. For factual/explainer reports: a numbered list of key findings with claim references. This section gives the reader the complete picture in one screen — the detailed analysis below is the supporting reasoning, not the main event. If using structured evidence, cite claim IDs like [c1], [c3] in table cells or list items.]

## [## Section per sub-question]
[Detailed analysis with inline citations — this is the supporting reasoning for the overview above. Each section traces back to the overview claims.]

### [### Subsections as needed]
[...]

## Contradictions
[If sources disagree, present both sides with attribution. Do not silently arbitrate. If no contradictions found, state "No direct contradictions between sources." This section is REQUIRED — its absence is a quality signal that counter-evidence wasn't searched.]

## Gaps
[What couldn't be determined from available sources. Each gap should note: what's unknown, why it matters, and whether it could be resolved with more research. If using structured evidence, reference the writing_context items. If no gaps, state "No significant gaps identified." **Source-set bias check (mandatory):** the first gap line must state the bias your whole source set shares — one dominant language, era, school, jurisdiction, or industry — adapted from Munger inversion. A homogeneous source set is itself a gap, even when every individual source is primary-tier.]

## Conclusion
[Synthesis of findings, implications, remaining uncertainties — ties back to the executive summary and overview. The conclusion confirms or qualifies the overview, it doesn't introduce new analysis.]

---
## Sources

**Quality distribution:** [N] primary · [N] secondary · [N] tertiary — [healthy/acceptable/weak]

| # | Title | URL | Quality | Accessed |
|---|-------|-----|---------|----------|
| 1 | [title] | [url] | primary/secondary/tertiary | [date] |

**Citation format:** Inline citations use `[N]` matching the source table above. On first mention of a source, include `[N]` — subsequent mentions may use `[N]` alone. For claims from the same source, cite the number once per paragraph unless the source is paginated.

**Evidence key** — `[VERIFIED]` corroborated across ≥2 independent, cited, dated sources · `[SOURCED]` from one named source, not independently corroborated · `[REASONED]` analytical judgement / inference · `[ESTIMATED]` calculation or stated assumption.

Category-specific overview sections

Category Overview section format
comparison Markdown table comparing options across criteria, with a verdict row or column. The detailed sections below provide the reasoning per option.
product Ranked list with pros/cons, price range, and a "top pick" callout. Detailed sections cover each product.
how-to Numbered overview of the steps. Detailed sections cover prerequisites, execution, and troubleshooting per step.
fact-check Evidence for/against the claim in a two-column table, with a preliminary verdict. Detailed sections assess source credibility and reasoning.
explainer Numbered key findings or a "progressive depth" overview (simple → deep). Detailed sections go deeper per concept. Glossary if technical.
factual Numbered key findings with claim references. Detailed sections provide the supporting evidence per finding.

Fallback report

If LLM synthesis fails (timeout, error, garbled output), compile raw findings into a basic report:

  • List all findings grouped by sub-question
  • Include source URLs and keep each finding's evidence-basis tag
  • Add note: "This is a raw findings compilation; synthesis was not completed."
  • Never output "No information could be gathered" if any sources were fetched — always compile what exists.

Step 5 — Research Stats

After the report, output a compact stats block:

---
📊 Research stats: [duration] · [N] rounds · [N] queries · [N] page fetches · [N] sources cited · provider: [donsetch/built-in]

Step 5.5 — Validation Gates (mandatory before delivery)

Deterministic gates — scripts/research_validation.py, stdlib-only, checks structure not judgment. Every report with 5+ sources passes three gates before delivery: validate-report (structure), verify-citations (inline [N] ↔ Sources rows), and verify-claims --strict (stored evidence must support each factual claim; a claim stored with basis: verified needs sources on ≥2 different sites; at least one polarity: refute claim, or --refute-none "<what you searched>"). Loop: validate → fix → re-run all three, max 3 cycles; still failing → stop and report the remaining problems to the user honestly. Never skip the gates, never deliver with a red gate. Quick 2-3 source reports (no store): verify-claims warns instead of failing — record it in the stats block. Full contract, flags, and the failure loop: references/validation-gates.md.

Optional Jev pre-screen. On instances with TypeSafe Jev configured, add --jev-prescreen to verify-claims for an advisory semantic pre-check that flags gate false-reds/false-greens before the deterministic run. It degrades to a clean SKIPPED line when Jev is not set up — never blocks, never changes the gate's exit code. See the contract in references/validation-gates.md.

Then run the plan's Done-when tests yourself — the gates check structure, these check the answer. Read the report against each test and fix what fails (a stale figure, an amount without its basis); a test that still fails is reported to the user as outstanding, plainly. Record the result in the stats block ("done-when: 3/4 met — [the outstanding one]").

Snippet anchoring prerequisite. verify-claims scores the stored snippets — it cannot tell a well-copied quote from a well-paraphrased one. That is what check-quote (§3c) is for: a red gate whose snippets never passed check-quote means anchoring was skipped, which is itself a defect. Anchor first, store second, gate third.

Step 6 — Follow-up Report Revision (optional)

For "verify X in the earlier report" / "update the pricing section" requests on a delivered report. Adapted from upstream's evidence panel + report-revision.ts: a follow-up researches narrowly and patches only the passages citing the finding — the rest of the report stays byte-identical.

  1. Map. python3 <skill_dir>/scripts/research_validation.py map-citations --report <report.md> [--citation N] returns the top-level blocks citing each [N] (code fences excluded; exit 1 = no block cites N). This defines exactly what a follow-up may touch — nothing else.
  2. Bounded follow-up research. One round, ≤2 query directions derived from the target finding and the user's instruction. Same snippet-first triage and quality filter (§3b/§3c); new page text joins the existing run dir; register new sources before citing them.
  3. Patch cited blocks only. Revise the blocks from step 1 (tables and lists are updated as whole blocks). Leave unrelated paragraphs untouched. If the evidence contradicts the existing claim, state the correction; if it is insufficient, say so instead of inventing a fix.
  4. Sources. Append new entries to the Sources table with the next free numbers; never renumber existing citations — old inline [N]s must keep resolving.
  5. History. Save the revision as <name>-rev.md next to the original (-rev2, -rev3, … for further rounds). The original file stays untouched; on failure or cancellation, nothing is saved.
  6. Re-run all three gates (Step 5.5) on the revised report before delivery. A red gate after revision means the revision did not land cleanly — fix, don't ship.

Follow-on Investment Analysis (optional)

When the report covers an investment question and the user provides deal parameters, build a quantitative pro-forma using execute_code. Always model the as-is use as a baseline. See references/real-estate-investment-analysis.md for the template.

See references/structured-evidence-format.md for the evidence.json schema and writing-context vs claims distinction.

Vault Integration (optional)

For larger research tasks (10+ sources) or when the user may want to follow up with grounded Q&A:

  1. Create a vault per notebooklm-mode at <project_folder>/research-<topic>/
  2. Save each source as a numbered file in sources/ with verbatim extracts
  3. Use ingest_source.py to write + index atomically
  4. After the report, tell the user: "Sources saved to vault at [path]. You can ask follow-up questions grounded in these sources — say 'notebooklm mode' to query the vault."

This is optional — the skill works fully without a vault for one-off reports.

Persistent citation database (absorbed source-tracker)

For research that spans sessions, keep a portable SQLite citation database instead of scattering URLs. The scripts live in scripts/source_db.py and scripts/url_health.py (formerly the standalone source-tracker skill). Full workflow doc: references/citation-database.md (schema details in citation-schema.md, export formats in citation-export-formats.md).

  • Add every source at the moment it passes the quality filter, before synthesis: python scripts/source_db.py add --url URL --topic TOPIC --title TITLE --type web --session-id RUN_ID. Never add an invented URL — only URLs observed through a real source channel.
  • Database path: set SOURCE_TRACKER_DB or pass --db-path; for multi-session work use an explicit path (for example <project>/.hermes/research/sources.db). The default ./sources.db depends on the current working directory.
  • Before final delivery: python scripts/source_db.py dedup (merges http/https and www. URL variants; the survivor keeps the earliest accessed_at), then python scripts/source_db.py export --topic TOPIC --format markdown --output bibliography.md (also bibtex, csv, json).
  • Link health before final citations: python scripts/url_health.py --stale-days 30 --batch-size 50 — it HEAD-checks stale URLs and marks HTTP >= 400 as dead (verified = 0). Dead links are recorded, not errors.
  • URL normalization: canonical storage lowercases scheme/host, strips fragments and non-root trailing slashes, preserves query strings; dedup comparison additionally treats http/https and www. variants as one source.

Subagent Mode (optional)

For genuinely parallel research across distinct sub-topics (e.g. researching 3 unrelated companies):

delegate_task(
  goal="Research [sub-topic] as part of a deep-research project. Run web searches, extract content, and return findings with source URLs and titles.",
  context="You are a research subagent. Topic: [sub-topic]. Sub-questions: [list]. Use web_search and web_extract. Return structured findings with citations.",
  toolsets=["web"]
)

Rules:

  • Only use subagents when sub-topics are genuinely independent
  • Collect results from all subagents, then synthesize yourself
  • Never delegate the synthesis step — the orchestrator must see all findings
  • Subagent summaries are self-reports — verify source URLs exist before citing

Concurrency and Rate Limits

  • Full-page fetches are on-demand and budgeted (§3b): ≤2/round, run-wide cap by complexity — snippets carry the triage
  • If web_extract fails on a URL, note it and move on — don't retry endlessly
  • If web_search returns no results, apply the §3a refinement protocol including the identical-rewrite stop

External provider head-to-head (optional benchmarking)

When the user wants this stack benchmarked against an external research provider (e.g. Perplexity), read references/research-provider-benchmark.md first. Standing rules: same raw question to both sides, no date or scope injection; verify the provider's central claims independently before scoring; report cost, breadth, and complementary positioning rather than a single winner.

Pitfalls

  • Overclaiming basis — when torn between two labels, pick the weaker one. A source's "about half" stays ~50% [SOURCED].
  • Conflating claims with context — subject facts are claims; scope, methodology, and availability limits are writing_context.
  • Subagent self-reports — verify source URLs exist before citing. The orchestrator synthesizes.
  • Clarification: one round — if the reply is still vague, proceed with the best interpretation of intent.
  • Complexity misclassification — complexity is breadth of synthesis, not how impressive the topic sounds. A narrow factual topic with six sub-questions is moderate.
  • Token budget fills before the cap — on long extractions, summarize aggressively when context is tight rather than waiting for the round ceiling.
  • Word-boundary matching — "port" matching "transport" or "support" is a false hit; filter on topical relevance, not substring overlap.
  • Keyless search noise on jurisdiction- or regulator-scoped questions — when a keyless web_search returns weak:true/degraded and is dominated by unrelated verticals, re-run with the authority or domain named in the query ("Singapore Monetary Authority", site:mas.gov.sg) before concluding anything is absent. Absence from a noisy keyless result set is not evidence of absence — the same query re-scoped routinely returns the primary documents that were missing.
  • Mid-research re-scope — when the user narrows the question after Round 1, do not restart the loop on the new scope: run one targeted pass, cite only sources that answer the narrowed scope, and state in the report what changed versus the broader framing.
  • Store claims are quote-shaped — the deterministic scorer supports a claim only when ONE stored snippet covers every figure, year AND entity the claim text carries; anything else caps at partial−0.01 (strict red for factual). The law, validated live three times in one session: multi-fact claims split per snippet, claim text byte-aligns to its quote, and prose paraphrase stays in the report body. Instances of the same law: CJK claims extract no tokens, so even an identical snippet scores ~0.33 — byte-identical claim text (entity axis 1.0) or anchor via check-quote (tokenization-free) instead; glommed capitalized runs ("Scrape URL FIRECRAWL" → one entity scrape url) never set-match a claim that says "Firecrawl" alone; alphanumeric brand prefixes ("x402" → figure 402) cap unless the snippet carries the brand text. Diagnose with the script's _figures/_entities/_tokens helpers before rewriting.
  • Anchored to the wrong substrate — check-quote --source-id reads pages/<source_id>.txt; verify the id resolves to the URL the sentence actually came from before treating a mismatch as fabrication — the same sentence on another page's file gives a false red.
  • Unanchored snippets — storing a snippet (or evidence) that never passed check-quote is how a smooth paraphrase becomes "evidence": the scorer sees familiar tokens and rates it supported. Verbatim excerpt, checked against the saved page text (<run_dir>/pages/<source_id>.txt), repair-once on mismatch — never massage the snippet toward the claim, never splice across paragraphs. No contiguous excerpt exists? The claim is not evidence-backed; store it honestly and let the gate flag it.
  • Store CLI ordering and field pairs — the subcommand must precede --dir (research_validation.py add-claim --dir <d> ..., never --dir <d> add-claim). add-claim takes claim+kind+polarity+basis+snippet+source_id; add-evidence takes evidence_snippet+source_id — swapping the pairs stores the snippet in source_id and surfaces later as unregistered-source_id warnings. basis takes ONLY the four labels — a source TIER name (primary/secondary) is rejected (rc 2), and kind takes no refute value: refute is a polarity. When looping store calls in a script, check every response's JSON — a rejected call exits 2 silently while the loop's "stored" message lies.
  • Citation density in the body — every Sources row must be cited at least once (verify-citations fails on uncited rows); required sections (## Executive Summary, ## Contradictions, ## Gaps, ## Conclusion) must be ##-level, not #.
  • One story counted twice — a press release and three articles reprinting it are one source family; two sites is not two sources. [VERIFIED] needs two independent families (references/evidence-hygiene.md §1).
  • A stale figure presented as current — give every time-sensitive fact its as-of period and flag it past its shelf life; undated figures are flagged as undated (evidence-hygiene.md §2).
  • An offer headline reported as the deal value — the announced and completion values of one deal are different bases. State the basis of every money figure and reconcile before choosing (evidence-hygiene.md §3).
  • A fixed budget spread across a grid — nine queries over eight items leaves each a single snippet and no context. Size a comparison run at one query per item plus two.
  • A stale headline from a search summary — leading a comparison row with a full-year figure when the same run found the later interim one. Start each item from its own latest disclosure; the newest figure in the run wins.
  • Searching before the comparison outline is agreed — a field redefined after the searches means re-running every item.
  • Process noise in the report — claim IDs, gate failures and store rebuilds belong in the stats block and the run folder, not in the findings.
  • Fallback if synthesis fails — compile raw findings with citations and tags; empty output is the failure mode.
  • Evidence key — paste the four-label legend under the Sources table so [VERIFIED] / [SOURCED] / [REASONED] / [ESTIMATED] decode.

Related Work

Odysseus — PewDiePie's self-hosted AI workspace — includes a "Deep Research" feature with multi-step web research and source reading, conceptually similar to this skill's Think → Search → Extract → Synthesize → Stop loop. This skill is a pure-prompt workflow (no UI, no server) designed to run inside any agent's tool loop.

AnotiaWang/deep-research-web-ui (MIT) — Nuxt UI wrapping dzhng/deep-research with a hardened engine, and the main vendor source for this skill. Vendored: verbatim-quote evidence anchoring, v1.8.2 (shared/utils/search-learning.ts → check-quote, §3c); structured search mini-plans with frozen constraints, planning discipline and provider guidance, v1.9.0 (shared/utils/search-plan.ts → §3a); search-budget presets + breadth-halving estimate, v1.9.0 (shared/utils/research-input.ts → Step 1 + estimate-budget); bounded deep-dive threading on follow-up directions, v1.9.0 (lib/core/deep-research.ts → §3a); six-way round-outcome taxonomy with per-reason bounded retry, v1.9.1 (shared/utils/search-assessment.ts → §3a); snippet-first on-demand reading with fetch budgets, v1.9.0 (lib/core/read-source.ts → §3b); term-scored verbatim passage selection, v1.9.0 (lib/core/source-context.ts → §3d); surgical report revision with citation→block mapping, v1.9.0 (shared/utils/report-revision.ts → Step 6 + map-citations). Evaluated, not vendored: full tree recursion, streaming-JSON structured-output repair, UI/deployment layers, and dzhng's "the source is irrelevant" prompt (contradicts this skill's provenance discipline). Full two-sided map: references/anotiawang-deep-research-concepts.md.

Anjos2/recursive-research (MIT) — Claude Code prompt skill for long-cycle recursive research. Vendored (v1.9.2, guidance-level): saturation-based stop signal (§3f), coverage balancing across plan sub-questions (§3a), mandatory source-set bias check in the Gaps section (Step 4). Not vendored: cycle machinery, WDM weighted-source scoring, per-domain seed source lists, preventive context-pause thresholds, local-corpus mode. The ≤5% threshold is unvalidated upstream (repo last pushed 2026-04-22, no field hardening); it ships as a signal, not a rule. Map: references/recursive-research-concepts.md.

DocsGPT (arc53, MIT) — ResearchAgent Plan → Research → Synthesize with adaptive depth, clarification, token-budget tracking, and empty-search refinement. Mapping: references/docsgpt-concepts.md.

references/skills-monetization.md — condensed AI-skills monetization knowledge bank (marketplaces, what sells, pricing) from the 2026-07-01 research dossier.

Evals

evals/routing-fixtures.json holds lightweight contract fixtures — sample request → expected routing, required output fields, and forbidden output patterns. They are specs, not run against a live model; the repo-root tests/test_routing_fixtures.py validates they stay well-formed and route to real skills.

04 同仓库更多 Skills

moonlight-lupin
moonlight-lupin/agent-skills

website-scraping

Generic playbook for extracting structured data from any website — hotel prices, flight fares, e-commerce listings, real estate, jobs, competitor product catalogues, anything where the goal is to turn one or more URLs into clean records on disk. Use this skill whenever the user mentions scraping, harvesting, extracting, or pulling data from a website; names a specific site or competitor they want data from; asks how to handle JavaScript-rendered pages, Cloudflare blocks, bot detection, Turnstile challenges, or Playwright; wants to monitor prices over time; needs to automate a copy-paste research task; or says things like "just get the data from X" or "I need a script that grabs N from site Y". Covers recon, picking the lightest extraction tool that works, surviving anti-bot defences, and writing clean JSONL output with a run manifest. Scope is scraping only — landing data in a database, scheduling, and downstream pipeline work are explicitly out of scope and live in other tools.

★ 123
moonlight-lupin
moonlight-lupin/agent-skills

greet

Greets a named person warmly.

★ 123
moonlight-lupin
moonlight-lupin/agent-skills

skill-quality-review

Use when auditing or reviewing skill quality — batch checks across a library, deep per-skill grep methodology, usage stats, efficacy testing, or fix planning. Single consolidated home for the writing-for-agents review framework and the 3-surface review methodology.

★ 123
moonlight-lupin
moonlight-lupin/agent-skills

people-enrichment

Use when the user has a list of names or companies and wants their job, employer, work history, LinkedIn URLs or company profiles looked up, or wants to find people or companies matching criteria — trigger on phrases like "enrich these contacts", "find their LinkedIn", "look up these people", "who works at X", "find directors at Y", "enrich these companies", or "get firmographics for". Enrich and search People Data Labs (PDL) Person and Company data, writing results to .xlsx. Five operations: enrich named PEOPLE into a profile, employment history and LinkedIn URL; IDENTIFY an ambiguous person as several scored candidates; SEARCH people by criteria (company, title, location); enrich COMPANIES into firmographics (industry, size, employees, HQ, founded, LinkedIn); and SEARCH companies by criteria. Do not scrape LinkedIn directly — this uses a licensed data aggregator (PDL) instead, for legal and reliability reasons.

★ 123
moonlight-lupin
moonlight-lupin/agent-skills

notebooklm-mode

Source-grounded research pipeline inspired by NotebookLM. Collect verbatim source extracts into a temporary workspace, then answer questions and create deliverables grounded in those sources. Two grounding modes: strict (vault-only) and augmented (vault + labeled [background] knowledge). Trigger when the user asks to research a topic from sources, wants NotebookLM-style grounded answers, asks to build a resource vault, or says "notebooklm mode".

★ 123
moonlight-lupin
moonlight-lupin/agent-skills

news-monitoring

Use when a topic or news stream needs recurring monitoring — cover it with web search, multi-language sources, digest formatting, and automated delivery via Hermes cron jobs. Covers search strategy, source selection, Chinese-language platforms, and digest templates.

★ 123