references/citation-protocol.md
# Citation verification protocol
## Two-step verification
For every candidate reference:
**Step 1: existence.** Search the full title (or a distinctive fragment plus
the first author's surname).
- Found with matching title and authors: existence passes.
- Found but title or authors differ: flag and confirm whether it is the same
work.
- Not found: grade UNVERIFIABLE; it does not enter the survey. Never keep it
with a warning label attached.
**Step 2: quote accuracy.** Does the finding quoted in the text match the
paper's title and abstract? With metadata only, do not fabricate method
details or specific numbers.
## Five-grade verdict
| Grade | Meaning | Usable |
|---|---|---|
| VERIFIED | exists; quote matches | yes |
| MINOR | exists; small wording drift | yes, after correction |
| MAJOR | exists; quote misrepresents it | no, correct or delete |
| UNVERIFIABLE | existence cannot be confirmed | no |
| PAYWALL | exists; metadata insufficient for the claim | restricted to metadata-level statements |
## Grey zone means unused
An unconfirmable citation stays out. "There is probably a paper showing X"
with no retrieval hit: do not cite it. One fewer reference always beats one
invented reference.
In the report body, gaps left by unusable sources use the uniform
unconfirmed marker, for example "[unconfirmed: no work retrieved in this
direction]". This is a research report convention, not a paper convention:
survey reports are working materials addressed to the researcher, so an
explicit unconfirmed marker is honest signaling there. Paper prose (the
paper-writer and intro-drafter skills) forbids all such markers; the two
regimes are different products, and the difference is deliberate.
## Zero tolerance for invented citations
Detection signals:
- a title that matches the point being argued suspiciously perfectly;
- an author-title-year-venue combination with zero retrieval hits;
- one author appearing many times in the list but only once in retrieval.
On discovering an invented citation: delete it immediately, then re-examine
every judgment that depended on it.
references/hedge-calibration.md
# Hedge calibration
A survey's credibility depends on every sentence's strength matching its
evidence. Overclaiming is spotted instantly by expert readers; underclaiming
drains the survey of judgment.
## Core principle: claim strength never exceeds evidence strength
## The ladder
### Strong (multiple independent studies agree; meta-analysis; large-scale validation)
- English: demonstrate, establish, show, confirm; consistently, robustly
- Chinese: 表明、证实、确立；一致地、稳健地
### Medium (two or three studies; bounded conditions)
- English: suggest, indicate, support the view that, be consistent with,
point to
- Chinese: 提示、说明、支持……的看法、与……一致、指向
### Weak (one or two studies, or mechanism-only argument)
- English: may, might, could, appears to, seems likely
- Chinese: 可能、或许、似乎、看来
### No evidence
- English: to our knowledge, no study has examined; it remains an open
question
- Chinese: 据检索结果，尚无研究考察过；仍是一个开放问题
## Common failures
### Overclaiming
- Wrong: "X proves Y" (prove belongs to mathematics)
- Wrong: "this is the best approach" (unqualified superlative)
- Wrong: "the first / the only / the optimal" without thorough verification
- Right: bound the claim: "among methods evaluated on dataset D, this
achieves the highest score on metric M"
### Underclaiming
- Wrong: "it may possibly suggest that there could perhaps be..." (stacked
hedges)
- Right: "the evidence suggests a moderate positive effect"
### Correlation written as causation
- Wrong: "X causes Y" without causal-inference evidence
- Right: "X is associated with Y"
## Do not stack jargon for authority
Authority comes from precise judgment matched to evidence, not from term
density.
- Wrong: "The paradigmatic shift in the methodological landscape
necessitates..."
- Right: "Methods have changed substantially since 2023, and earlier
conclusions may no longer hold."
references/output-structure.md
# Report structure and formatting
## Table of contents
1. The survey skeleton
2. Title
3. Abstract
4. Introduction
5. Methodology section
6. Taxonomy section
7. Branch sections
8. Synthesis discussion
9. Open problems
10. Conclusion
11. References
12. Formatting discipline
13. Length
The deliverable reads like a real survey paper: explicit research questions,
a systematic method, a structured taxonomy, cross-branch synthesis, and
conclusions that the evidence has earned.
## 1. The survey skeleton
Numbering rule: all body sections are numbered consecutively (Introduction
through Conclusion); only the Abstract and References go unnumbered. Branch
count floats with the topic (typically two to four), so the section numbers
of the discussion, open-problems, and conclusion sections float too.
```
# [Topic]: [angle / core judgment]
## Abstract
[problem, scope, method, core findings, significance; one paragraph]
## 1. Introduction
[why this survey is needed; the research questions RQ1..RQn; how the
report is organized]
## 2. Methodology
[search tools and perspectives; keywords; time range; inclusion and
exclusion; final corpus size; why this taxonomy]
[non-CS topics add an evidence-type table: type, meaning, credibility]
## 3. Taxonomy
[the classification axes, branches, and empty cells (= gaps), shown
before the detail]
## 4. [Branch one]
[thematic prose with in-sentence comparison; a comparison table when
three or more works are comparable; caption states the conclusion]
## 5. [Branch two] ...
## N. Synthesis discussion
[cross-branch analysis: trends, tensions, shared findings]
## N+1. Open problems and future directions
[specific, literature-backed gaps]
## N+2. Conclusion
[answers RQ1..RQn, one by one; then the survey's own contribution]
## References
```
## 2. Title
Title = topic + angle, never topic alone.
- Weak: "Large language models in medicine"
- Strong: "Large language models in medical imaging: the gap between
general capability and clinical usability"
## 3. Abstract
One paragraph, roughly 150-250 words. Problem (why this investigation),
scope (what it covers), method (how), core findings (the two or three that
matter), significance (why the reader should care). The abstract gives
direction; it does not enumerate every conclusion. Its job is to let the
reader decide whether to read on.
## 4. Introduction
Answers three questions:
1. **Why is this survey needed?** Not "the field matters" (that needs no
survey), but "what existing understanding or existing surveys lack, and
what this one adds".
2. **What exactly will it answer?** Two or three research questions, each
concrete and answerable. "How is CRISPR doing clinically" is not an RQ;
"which diseases have phase-III or approved CRISPR therapies" is.
3. **How is it organized?** One paragraph of structure preview.
## 5. Methodology section
Search perspectives (list them), main keyword groups, time range, inclusion
and exclusion criteria, final corpus size, and the reasoning behind the
taxonomy design. For non-CS topics, include the evidence-type table so the
reader can weigh different kinds of evidence.
## 6. Taxonomy section
The survey's skeleton, shown before the detail. A good taxonomy has MECE
axes, visible empty cells (each one a gap and a finding), identified
straddlers (works that challenge the axes), and branches that align with the
RQs so each branch answers part of the question set.
## 7. Branch sections
Internal structure of each branch:
1. One opening sentence placing the branch in the framework.
2. Thematic prose, comparing works within sentences: consensus, divergence,
and the conditions that explain divergence.
3. A comparison table when three or more works are comparable (columns are
dimensions; missing values read "not reported"; caption states the
conclusion).
4. An optional one-line takeaway for long branches.
## 8. Synthesis discussion
Not a recap of the branches. Cross-branch analysis only: trends all branches
share; tensions between one branch's conclusions and another's premises;
methodological problems the core papers have in common (for example, none
independently replicated); the key cross-branch comparison table if none
appeared earlier.
## 9. Open problems
Specific and literature-backed, never "more research is needed":
- Weak: "this direction needs more study"
- Strong: "the [X by Y] cell of the taxonomy is empty: no work applies X
under Y, although the branch-two evidence suggests it would matter
because..."
Absence claims need evidence too: "no work retrieved" is only writable after
the search actually looked.
## 10. Conclusion
Answers the Introduction's research questions one by one; that closes the
loop the Introduction opened. No new evidence here. The final paragraph
states what the survey itself contributes: the new framing, the exposed
tension, or the mapped gap, not "we wrote a survey".
## 11. References
Strictly uniform format, numbered in order of first appearance:
```
[1] Author1, Author2, et al., "Title," Venue, Year.
[2] Author1, Author2, "Title," arXiv:XXXX.XXXXX, Year.
```
More than three authors: et al. Titles in quotes. No URLs, no DOIs. No
unverified entries.
## 12. Formatting discipline
1. Section headings are theme names, not mechanical labels.
2. No emoji.
3. Bold only for core judgments and first occurrences of key terms.
4. No boilerplate openings ("With the rapid development of AI..."); open
with a concrete fact.
5. Citations as bracketed numbers.
6. Mixed-language text keeps technical terms in English with a half-width
space between English and Chinese.
7. Evidence strength is carried by wording (the hedge ladder), never by
explicit labels.
## 13. Length
No hard cap. Long enough to answer the RQs, and nothing that does not
advance them: the self-check is that deleting a paragraph must hurt an RQ's
answer, or the paragraph was padding.
references/quality-gates.md
# The six internal quality gates
Internal only; never shown to the reader. Each gate has detection questions,
a three-level verdict (CLEAR / SUSPECTED / INSUFFICIENT), and a failure
route.
## Angle (CRITICAL)
**Failure mode**: the output is a topic listing, not a survey with a
judgment.
Detection:
1. Can the survey's core judgment be stated in one sentence? Naming only the
topic means the angle never stood up.
2. Compared with a raw search-result list, what judgment does this add? If
the answer is "nicer formatting", it is not a survey.
3. How does the angle differ from existing surveys of the same area?
Finding a strong angle: the best angles come from unexpected convergence or
tension across independent sources: several unrelated benchmarks pointing at
one conclusion; a systematic gap between what theory worries about and what
practice struggles with; different disciplines reaching one bottleneck by
different routes. If every paper says roughly the same thing, the angle is
"consensus has formed, but remains contested under condition X".
**Route**: back to the brief; re-derive the angle.
## Coverage (MAJOR)
**Failure mode**: key works missing; conclusions built on partial evidence.
Detection:
1. Three or more works per sub-direction?
2. Does every taxonomy cell hold literature, and is each empty cell a true
gap rather than a search miss?
3. Are the high-frequency works cited by the found papers themselves
present?
4. Is any method school or research group entirely absent?
5. Do discarded search results point at an uncovered sub-direction?
**Route**: targeted supplementary search (back to the scouting phase).
## Citation (CRITICAL)
**Failure mode**: references that do not exist or are misquoted.
Detection:
1. Did every reference pass the verification protocol?
2. Is every specific finding quoted in the text confirmable from the paper's
title or abstract?
3. Does any title match the argued point suspiciously perfectly?
4. For any author cited more than three times: was each entry independently
verified?
**Route**: back to verification; invented citations are deleted on
discovery, along with every judgment that leaned on them.
## Taxonomy (MAJOR)
**Failure mode**: papers listed one by one instead of organized by theme.
Detection:
1. Are section titles theme names or paper names?
2. Are the axes mutually exclusive and collectively exhaustive?
3. Are empty cells real gaps or artifacts of the axis choice?
4. Were straddling papers identified and discussed?
**Route**: redesign the axes or regroup.
## Calibration (MAJOR)
**Failure mode**: judgment strength out of proportion to evidence strength.
Detection:
1. Does every "demonstrates / establishes" rest on multiple independent
studies? A single study earns "suggests".
2. Any unqualified superlatives ("X is the best approach")?
3. The reverse failure too: strong convergent evidence hedged into mush
("may possibly suggest")?
**Route**: re-calibrate wording against the hedge ladder.
## Weaving (MAJOR)
**Failure mode**: papers summarized in isolation; no in-sentence comparison.
Detection:
1. Does every thematic paragraph have at least one sentence citing and
comparing two or more works?
2. Any "A did X. B did Y. C did Z." runs?
3. Are contradictions presented with condition analysis, or waved off as
"this is debated"?
4. Do the comparison tables reveal meaningful differences (columns as
dimensions, captions stating conclusions)?
**Route**: rewrite the flagged sections against the synthesis framework.
references/search-strategy.md
# Search strategy
## Keyword construction
Keywords are constructed systematically, not free-associated. Patterns:
**Pattern 1: core method x application domain**
- "retrieval augmented generation" + "question answering"
- "CRISPR base editing" + "clinical trial"
**Pattern 2: core mechanism x target task**
- "chain of thought" + "scientific reasoning"
**Pattern 3: field + survey / benchmark / meta-analysis**
- "LLM agent survey 2025 2026"
- "social media adolescent mental health meta-analysis"
**Pattern 4: key researcher x direction** (when the brief names specific
researchers)
- "Haidt social media adolescent"
- "Orben screen time well-being"
**Pattern 5: citation network of known papers**
- extract key terms from already-found core papers; search what they cite
and what cites them
Try at least one round per pattern. In multi-perspective search, each
perspective uses one or two patterns; three to five perspectives at about
two rounds each sets the floor, with no ceiling: coverage decides.
## Blind-spot discovery
After the second round, run a blind-spot check:
1. **Coverage**: does every sub-direction of the brief have three or more
works?
2. **School check**: is any method school entirely absent?
3. **Citation-chain check**: are the high-frequency works cited by the found
papers themselves in the list?
4. **Discard check**: do the "found but not planned for use" results point
to an uncovered sub-direction? Do not throw them away; they are the
blind-spot clues.
5. **Recency check**: anything important from the last year? All-old results
usually mean the keywords missed the current wave.
## Adjusting by discipline
Literature distributes differently across fields:
- **CS / AI**: arXiv preprints dominate and move fast; the last six months
can be decisive.
- **Biomedicine**: clinical journals matter; search "clinical trial" plus
the disease name; note trial phase.
- **Social science**: search effect sizes and designs (RCT, DID, IV)
alongside the topic.
- **Cross-disciplinary topics**: search each discipline separately, then
merge; a single discipline's keywords will not surface the others.
references/self-adversarial.md
# Self-adversarial review
Run after synthesis, before presentation. The aim is not perfection; it is
catching the blind spots that feel fine while writing and get asked about
the moment someone else reads.
## Three retrospective questions
### 1. Starting over, would the brief change?
Return to the frozen research brief. Would what you now know change the
angle?
- No: CLEAR.
- Somewhat, core intact: add one line to the limitations.
- Substantially: frame-lock detected; seriously consider redoing the brief.
### 2. Any unexamined premises?
Is the survey built on assumptions never made explicit? Typical ones:
assuming higher accuracy means better; assuming English-dataset results
speak for other languages; assuming the current trend continues.
Found: add the qualification in the limitations or the affected section.
### 3. Are the conclusions evidence-driven or selection-driven?
For each core conclusion:
- Does counter-evidence exist, and was it presented?
- If only supporting works are cited: is that because none oppose, or
because none were searched for?
- State the inverse of the claim: does anything support the inverse?
Selective citation found: search for the opposing side, present it in the
contested section.
## Perspective-omission check
- **Discipline**: did a cross-disciplinary topic get written from one
discipline only?
- **Geography**: does the literature over-concentrate in a few regions'
research groups?
- **Method**: did one method family (say, deep learning) crowd out the
traditional approaches?
- **Criticism**: is the survey uniformly positive? Where are the failures,
limits, and critics?
Omissions found: supplement in the open-problems or limitations sections.
## Concession rules
When this review finds a problem with a conclusion:
- Counter-evidence overturns it: rewrite the claim; downgrade to a
conditional.
- Counter-evidence weakens it: keep the claim, add the qualification.
- A different-but-compatible perspective: present both in the contested
section.
- A vague unease with no locatable counter-evidence: change nothing. Do not
soften an evidence-backed judgment to perform humility.
## Output
No separate report. The output is the corrections themselves: revised
claims, added qualifications, supplemented contested sections. If all three
questions come back clean, proceed to presentation without manufacturing
self-criticism.
references/synthesis-framework.md
# Synthesis framework
## Table of contents
1. The three-step flow (extraction, clustering, serial synthesis)
2. Weaving: in-sentence cross-comparison
3. Comparison-table design
4. Presenting contradictions
5. The per-section skeleton
## The three-step flow
### Step 1: extraction
From each paper, extract four things that bear on the research brief:
- **Core finding**: one or two sentences.
- **Method signature**: what method, and how it differs from the others.
- **Conditions**: under what conditions the finding holds; known limits.
- **Relations**: whom it supports, challenges, or extends.
Irrelevant papers are dropped whole at this step.
### Step 2: thematic clustering and MECE taxonomy
Cluster by **research question or method school**, never by paper. Mark each
outline node with:
- **Presentation**: paragraph (dense synthesis of several works) or table
(side-by-side comparison of three or more comparable works).
- **Assigned works**: which papers this node may cite.
Taxonomy rules:
- Axes mutually exclusive, collectively exhaustive.
- An empty cell is a gap, and a gap is a finding ("no work applies X to Y"
is a result worth reporting).
- Papers that straddle categories get discussed explicitly; they challenge
the axis and are often the most insightful.
- If one axis produces too many straddlers, change the axis.
### Step 3: serial section synthesis
Write section by section in outline order. Each section's input is the
brief, the works assigned to that section, and the sections already
written. Serial, not parallel: parallel drafting produces duplication,
style drift, and missing transitions.
Constraint inherited from step 2: each section cites only its assigned
works. Writing from "whatever I remember of the pool" is how citations
drift.
## Weaving: in-sentence cross-comparison
This is the measurable difference between a survey and a search-result
summary. Human-written surveys compare several works inside one sentence;
naive generation summarizes papers one at a time in isolation.
**Anti-pattern (serial listing)**:
> Smith et al. (2024) proposed method A for task X and achieved 85 percent
> accuracy. Lee et al. (2025) proposed method B for task X and achieved 87
> percent.
**Pattern (in-sentence comparison)**:
> While Smith et al. [1] and Lee et al. [2] both target task X, they take
> fundamentally different routes: Smith et al. lean on retrieval
> augmentation, reaching 85 percent, whereas Lee et al. rely on in-context
> learning, reaching 87 percent at three times the inference cost. Wang et
> al. [3] further show the retrieval-based strategy generalizes better to
> the related task Y, suggesting broader applicability despite the lower
> headline number.
**Operating rule**: every thematic paragraph contains at least one sentence
that cites and compares two or more works. A paragraph citing a single work
reads as an abstract, not a survey.
## Comparison-table design
- Columns are comparison dimensions (method trait, strength, limitation,
metric), not paper attributes.
- Rows are the works.
- **The caption states the conclusion** ("retrieval methods lead on recall
but cost 3-5x the latency of end-to-end methods").
- Missing information reads "not reported"; cells are never left blank and
never filled with invented numbers.
- Drop sparse columns (a dimension only one or two works report does not
deserve a column).
## Presenting contradictions
Contradictions between works are among the most valuable content a survey
has. **Do not average them away.**
When A finds X effective and B finds X ineffective:
1. Present both findings plainly.
2. Analyze the condition differences (dataset, metric, setting, sample
size).
3. If a judgment is possible, give the lean ("under condition Z, B's
evidence is stronger").
4. If not, say so: "current evidence does not decide this; a test under W
would".
Never: write "this is debated" and move on; average the two; cite only the
side that fits the narrative.
## The per-section skeleton (run before writing every section)
1. **Claim**: what do these works, taken together, tell us? Not "what does
this section cover".
2. **Strongest supporting evidence**: which works, and does independent
convergence exist?
3. **Counter-evidence or qualifications**: which works contradict or bound
the claim?
4. **Condition differences**: what explains the disagreement?
5. **Cross-section links**: what does this section support or challenge
elsewhere?
The skeleton is never printed; it drives the prose. Self-check afterwards:
delete the citations mentally; if what remains is still an argued judgment
rather than a list of facts, the section is sound.
SKILL.md
---
name: deep-research
description: >-
Runs a deep, survey-grade literature investigation on a research topic:
freezes research questions, searches from multiple adversarial
perspectives, verifies every citation, synthesizes evidence into a MECE
taxonomy with in-sentence cross-comparison, and answers the research
questions in an evidence-first report. Use when the user asks for a
literature review, a survey of a field, research landscape mapping, or a
deep dive into what is known about a topic.
license: CC-BY-NC-SA-4.0
---
# Deep Research
## Overview
The role is a senior researcher who has written strong surveys and served as
a section editor. Someone brings a research topic; the job is to map the
terrain, lay out the key works, surface the tensions and the gaps, and end
with a judgment the evidence has earned.
This is not a search engine and not a bullet-point answer. The deliverable
reads like a real survey paper: questions defined first, evidence developed
systematically, conclusions emerging from the evidence rather than announced
up front and back-filled.
## When to use this skill
- "Survey this direction for me." "What is the state of X?"
- "What are the key works, debates, and open problems in this area?"
- "I need a literature review before committing to this topic."
## When NOT to use this skill
- Judging whether one specific idea is worth pursuing: `idea-evaluator`.
- Drafting or polishing paper prose: `paper-writer`, `intro-drafter`,
`paper-polish`.
- A quick lookup of one or two papers: plain search suffices; this skill is
for depth.
## Capability check
- **Literature search**: a scholarly search tool if available; otherwise web
search over scholarly indexes; otherwise shell access to public APIs
(Crossref, Semantic Scholar, arXiv, DBLP). With no retrieval at all, this
skill does not run: say so rather than writing a survey from memory.
- **Sub-agents**: when the environment can spawn parallel sub-agents, each
search perspective runs as its own agent with a clean context; without
them, run the perspectives serially yourself. The method does not change,
only the execution.
## Step 0: freeze the research brief
Before any searching, pin down what is being investigated.
**The brief must contain research questions**: not a topic, but two or three
specific, answerable questions. They are the survey's promise; the
conclusion must answer each one.
If the user gives only a topic ("survey CRISPR for me"), convert it with one
or two clarifying questions:
- "CRISPR clinical progress" becomes RQ1: which diseases have phase-III or
approved CRISPR therapies? RQ2: where is the safety boundary for in-vivo
editing? RQ3: what limits accessibility?
- "social media and adolescent mental health" becomes RQ1: does causal
(not correlational) evidence support harm? RQ2: what is the effect size?
RQ3: who is most affected?
Brief = topic + research questions + angle + intended reader. Every later
step answers to this brief; drift is not allowed.
The strongest angles come from unexpected convergence or tension across
independent sources. The Angle gate in references/quality-gates.md gives the
tests.
## Phase 1: scout (search and verify)
Generate three to five **search perspectives** around the brief, each with
its own question and its own keywords:
- the mainstream school (who advances this direction, and with what);
- the critics (who doubts it, with what counter-evidence);
- adjacent fields (what other disciplines say about the same question);
- methodology (who challenges how the mainstream measures things:
meta-analyses, systematic reviews);
- application and policy (what deployment or regulation evidence exists).
Each perspective searches independently, wide then narrow: a first round on
core keywords, a second round narrowed by what the first returned (method
names, dataset names, terms of art). Keyword construction patterns and the
discipline adjustments: references/search-strategy.md.
When sub-agents are available, run one per perspective in parallel; each
returns candidate works with one-line findings. Then merge across
perspectives and check blind spots: perspectives that returned almost
nothing (a real gap, or bad keywords?), sub-directions no perspective
covered, high-frequency works cited by the found papers but missing from
the pool, and discarded results that point somewhere uncovered.
**Every candidate reference is verified before use.** The two-step protocol,
five-grade verdict, and the grey-zone rule (unconfirmable means unused):
references/citation-protocol.md.
Corpus size has no quota; coverage of every sub-direction (three or more
works each) is the bar. Over-collecting is fine; the synthesis phase
filters.
## Phase 2: synthesize
With the verified corpus in hand: extract per-paper findings, cluster by
theme into a MECE taxonomy, assign works to branches, then write the
branches serially, each section citing only its assigned works. In-sentence
cross-comparison is mandatory; contradictions are presented with condition
analysis, never averaged away. The full method:
references/synthesis-framework.md.
After synthesis, run the self-adversarial review (three retrospective
questions, perspective-omission check, concession rules):
references/self-adversarial.md.
## The six gates and the iteration loop
Throughout, six internal gates guard quality (never shown to the reader):
| Gate | One line | Severity | Failure route |
|---|---|---|---|
| Angle | a judgment, or just a listing? | CRITICAL | back to Step 0 |
| Coverage | key works all found? | MAJOR | targeted re-scout |
| Citation | references real and honestly quoted? | CRITICAL | re-verify; delete inventions |
| Taxonomy | organized by theme, MECE? | MAJOR | redesign axes |
| Calibration | claim strength matches evidence? | MAJOR | re-calibrate wording |
| Weaving | in-sentence comparison present? | MAJOR | rewrite flagged sections |
Detection questions and routes: references/quality-gates.md.
The loop this creates is the point: a Coverage or Citation failure sends the
work back to scouting for a **targeted** supplement (new works append to the
corpus; the synthesis updates incrementally), not a restart. Iterate until
every gate is CLEAR and the self-adversarial pass stops finding corrections.
Depth is driven by the topic's complexity, not by a fixed round count.
## Presentation
Organize the final report as a survey paper: abstract; introduction with the
research questions; methodology; taxonomy; branches with comparison tables;
cross-branch synthesis discussion; open problems; a conclusion that answers
the RQs one by one; references. The skeleton, per-part guidance, and
formatting discipline: references/output-structure.md.
Wording strength follows the hedge ladder (claim never exceeds evidence):
references/hedge-calibration.md.
**Evidence-first narration**: the introduction poses the questions, the body
develops the evidence, the conclusion answers. The reader derives the
conclusions alongside the text instead of being told first and shown
supporting quotes after. This ordering is the survey's defense against its
own confirmation bias.
## Cross-discipline evidence standards
Switch what counts as strong evidence by field:
- **CS / AI**: benchmarks, ablations, reproducibility; note model versions
and evaluation conditions.
- **Biomedicine**: trial phase, patient counts, follow-up length, approval
status.
- **Social science**: causal versus correlational designs (RCT, natural
experiments, IV); effect sizes and heterogeneity.
- **Economics and finance**: identification strategy (IV, DID, RDD);
statistical versus economic significance.
- **Cross-disciplinary**: separate model predictions, observational
attribution, and field experiments; carry uncertainty ranges.
Non-CS reports include the evidence-type table from the output-structure
reference so readers can weigh evidence kinds.
## Unconfirmed markers: a deliberate difference from paper prose
A survey report is a working document addressed to the researcher, so an
explicit unconfirmed marker ("[unconfirmed: no work retrieved in this
direction]") is honest signaling and is allowed, under the uniform
convention in references/citation-protocol.md. Paper prose (paper-writer,
intro-drafter) forbids all bracketed markers. The two regimes differ on
purpose; do not import this convention into paper drafting.
## Output language
The report follows the language of the user's request: a question asked in
Chinese gets a Chinese survey, English gets English. (This deliberately
differs from the drafting skills, which default to English: a survey is read
by the researcher, not submitted to a venue.) In Chinese reports, keep
technical terms in English with a half-width space around them. An explicit
language request always wins.
## Delivery
A full survey is written to a workspace file (Markdown), with the structure,
tables, and reference list intact. Short literature answers can stay in the
conversation. Capability degradations (perspectives run serially, weaker
retrieval for some literature) are disclosed in a short note.
## Boundaries with sibling skills
- Wants a verdict on one idea: `idea-evaluator`.
- Wants prose written or polished: `paper-writer` / `paper-polish`.
- The evaluator or writer discovers the user actually needs the landscape
first: they route here.