flowchart LR
    H["Researcher<br>question, scope, judgment"] --> I["Idea + Scope"]
    I --> L["Literature Review"]
    L --> N["Novelty Check"]
    N --> E["Experiment Planning"]
    E --> V["Claim Verification"]
    V --> R["Reviewer Simulation"]
    R --> W["Writing / Revision"]
    W --> B["Reviewer Response"]
    B --> F["Final Verification"]
    F --> S["Submission"]
    V -. evidence .-> L
    R -. objections .-> E
    B -. revision .-> W
    H -. approval gates .-> N
    H -. final judgment .-> F

git clone https://github.com/ngtiendong/Academic-Research-Agent-Skill.git
cd Academic-Research-Agent-Skill
cp config/language.example.yaml config/language.yaml

Read SKILL.md and CLAUDE.md.

Do not write a paper yet.

First explain:
1. the current research workflow,
2. the evidence gates,
3. what evidence is required at each gate,
4. what decisions must remain with the researcher,
5. what information you need from me before proceeding.

My research idea is:

"Can method X improve outcome Y under condition Z?"

Do not assume this idea is novel.

First:
1. turn it into a precise research question,
2. define the scope,
3. list non-goals,
4. identify the candidate contribution,
5. identify the evidence required,
6. identify the closest competing approaches,
7. tell me what decision I need to make next.

Do not write the paper yet.

Act as a human-guided research assistant.

My role:
- I own the research question.
- I approve scope and major decisions.
- I make the final scientific judgment.
- I own the final claims.

Your role:
- organize and compare evidence,
- identify weak assumptions,
- help inspect prior work,
- test novelty,
- plan feasible experiments,
- verify important claims,
- simulate reviewers,
- help revise the manuscript.

Rules:
1. Never invent citations, data, results, datasets, or experimental outcomes.
2. Separate inspected evidence, inference, and hypothesis.
3. Do not call an idea novel just because you did not find an identical paper.
4. Do not strengthen a claim beyond its evidence.
5. Mark missing information instead of filling gaps.
6. Ask for approval at major decision gates.
7. Do not write the full paper before scope, evidence, and claims are sufficiently established.

My research idea:
[INSERT IDEA]

Start with scope and evidence requirements.

flowchart LR
    Q["Research Question"] --> S["Search Strategy"]
    S --> C["Collect Sources"]
    C --> F["Filter / Screen"]
    F --> I["Inspect Evidence"]
    I --> X["Extract"]
    X --> M["Compare"]
    M --> Y["Synthesize"]

I am researching:

[TOPIC / QUESTION]

Help me convert this into:

1. precise research question,
2. key concepts,
3. synonyms,
4. methodological terms,
5. inclusion criteria,
6. exclusion criteria,
7. evidence I need from each source.

Separate:
- facts already supported by my sources,
- assumptions,
- search hypotheses.

Do not invent papers or citations.

Create a literature search strategy for this research question:

[QUESTION]

Return:
1. primary keywords,
2. synonyms,
3. related concepts,
4. methodological terms,
5. likely competing approaches,
6. terms that may produce false-positive results.

Do not treat your proposed search terms as evidence.

For each paper I provide, extract:

1. research problem,
2. method,
3. dataset or context,
4. evaluation metric,
5. main result,
6. limitation,
7. relevance to my research question,
8. closest relationship to my proposed contribution.

For every important statement, identify the source location where possible.

If information is missing, write "not found".
Do not infer missing facts.

Using only the inspected sources:

1. identify areas of agreement,
2. identify conflicting findings,
3. identify methodological differences,
4. identify unresolved gaps,
5. identify which gap is directly relevant to my research question.

Do not label a gap as "novel" yet.

flowchart TD
    I["Proposed Contribution"] --> B["Break into<br>problem / method / context / data / metric"]
    B --> C["Find Closest Prior Work"]
    C --> D["Compare Contribution Delta"]
    D --> G{"Novelty Gate"}
    G -->|Weak / same contribution| R["Revise Scope"]
    R --> I
    G -->|Defensible delta| P["Proceed"]

My proposed contribution is:

[CONTRIBUTION]

Do not assume it is novel.

Break it into:
1. problem,
2. method/intervention,
3. context,
4. dataset,
5. metric,
6. claimed improvement.

For each component, explain what would need to be different for it to constitute a meaningful contribution.

Using only sources that can actually be inspected:

Identify the closest prior works to my proposed contribution.

For each work compare:
- problem,
- method,
- context,
- dataset,
- metric,
- contribution,
- limitation.

Then identify the smallest defensible difference between my proposal and each prior work.

Do not call my proposal novel yet.

Run a novelty gate.

Classify the proposal as:

PASS
- evidence shows a defensible contribution delta.

REVISE
- the idea may be useful, but the contribution or scope is not sufficiently differentiated.

FAIL
- closest prior work already covers substantially the same contribution.

For every classification:
1. cite the comparison evidence,
2. explain the reasoning,
3. state what would change the decision.

Do not equate "not found" with "never existed".

flowchart LR
    C["Claim"] --> S["Source"]
    S --> E["Exact Evidence"]
    E --> I["Interpretation"]
    I --> V{"Verification"}
    V -->|Supported| A["Authorized Claim"]
    V -->|Partial| P["Narrow Claim"]
    V -->|Contradicted| X["Remove / Resolve"]
    V -->|Unverified| U["Do Not Use Yet"]

Verify this claim:

"[CLAIM]"

For every relevant source:

1. identify the exact evidence,
2. explain what the evidence actually establishes,
3. identify conditions or limitations,
4. determine whether the source supports the claim,
5. flag any overstatement.

Classify the claim as:
- supported,
- partially supported,
- contradicted,
- unverified.

Do not strengthen the claim beyond the evidence.
If the source cannot be inspected, say so.

Audit the following draft for evidence-to-claim alignment.

For each major claim:
1. extract the claim,
2. identify the supporting source or result,
3. state whether the evidence is direct or indirect,
4. identify unsupported scope,
5. propose the narrowest defensible wording.

Do not rewrite the entire paper yet.
Return a claim ledger first.

flowchart TD
    H["Hypothesis"] --> S["Technical Smoke Test"]
    S -->|Fails| D["Debug / Stop"]
    S -->|Pass| P["Feasibility Pilot"]
    P -->|Not feasible| R["Revise Scope"]
    P -->|Feasible| G{"Evidence sufficient?"}
    G -->|No| P
    G -->|Yes| F["Full Run"]
    F --> V["Claim Verification"]

My hypothesis is:

[HYPOTHESIS]

Help me define:
1. independent variable,
2. dependent variable,
3. baseline/control,
4. intervention,
5. evaluation metric,
6. confounders,
7. expected failure condition,
8. minimum evidence needed to test the hypothesis.

Do not invent results.

Design the cheapest decisive feasibility pilot for this hypothesis.

Include:
- objective,
- unit of analysis,
- inputs,
- intervention,
- baseline,
- metric,
- success criterion,
- failure criterion,
- stop condition,
- data/provenance that must be recorded,
- assumptions that could invalidate the pilot.

Do not design the full experiment yet.

Act as a skeptical experimental-methods reviewer.

Audit this experiment plan for:

1. missing baseline,
2. confounders,
3. invalid metric,
4. data leakage,
5. insufficient sample or unit definition,
6. mismatch between hypothesis and measurement,
7. reproducibility risks,
8. hidden assumptions,
9. claims that the experiment cannot actually support.

Rank each issue:
- blocking,
- major,
- minor.

For every blocking or major issue, explain what evidence or design change would resolve it.

Act as a skeptical peer reviewer.

Review this research proposal/manuscript:

[PASTE MATERIAL]

Evaluate:
1. significance,
2. novelty,
3. methodological validity,
4. evidence quality,
5. experimental design,
6. claim strength,
7. reproducibility,
8. missing baselines,
9. likely reasons for rejection.

For every criticism:
- quote or identify the exact part being criticized,
- explain why it matters,
- classify severity as major or minor,
- state what evidence or change would resolve it.

Do not invent flaws that are not supported by the material.

Focus only on novelty and contribution.

Try to find the closest prior work or argument that could weaken the contribution.

Do not invent citations.
Separate:
- demonstrated overlap,
- plausible concern,
- unknown due to missing evidence.

Focus only on:
- controls,
- baselines,
- metrics,
- confounders,
- experimental unit,
- reproducibility,
- data leakage,
- statistical or evaluation weaknesses.

Rank issues by scientific severity.

Audit whether each major claim is supported by the proposed evidence.

For each claim:
- evidence available,
- evidence missing,
- overclaim risk,
- minimum change needed.

Merge these reviewer reports.

Identify:
1. objections appearing across multiple reviewers,
2. objections that are mainly stylistic,
3. critical issues that must be fixed,
4. issues that can be resolved with clarification,
5. issues requiring new evidence.

Create a prioritized revision plan.

Do not optimize for "sounding convincing".
Optimize for scientific defensibility.

flowchart LR
    C["Reviewer Comment"] --> T["Triage"]
    T --> E["What evidence do we have?"]
    E --> D{"Response type"}
    D --> A["Agree + Change"]
    D --> B["Clarify"]
    D --> C2["Add Evidence"]
    D --> D2["Disagree + Evidence"]
    D --> E2["Cannot Address"]
    A --> R["Draft Response"]
    B --> R
    C2 --> R
    D2 --> R
    E2 --> R
    R --> M["Manuscript Change"]
    M --> V["Final Verification"]

For each reviewer comment:

1. classify the concern,
2. identify what the reviewer is actually asking,
3. determine whether we have evidence,
4. determine whether a manuscript change is required,
5. identify what we cannot honestly claim.

Do not draft the rebuttal yet.

Return a response matrix.

Draft an evidence-based response to this reviewer comment:

[COMMENT]

Use this structure:
1. acknowledge the concern,
2. answer directly,
3. provide evidence,
4. describe the manuscript change,
5. identify the exact section/table/figure changed.

Rules:
- do not invent experiments,
- do not invent citations,
- do not invent results,
- do not claim a manuscript change that was not actually made,
- do not become defensive,
- disagreement must be supported by evidence.

flowchart TD
    I["1. Idea"] --> S["2. Scope"]
    S --> G1{"Human Gate"}
    G1 -->|Revise| S
    G1 -->|Approve| L["3. Source Ingestion"]
    L --> LG["4. Literature Grounding"]
    LG --> M["5. Formalization"]
    M --> N{"6. Novelty Gate"}
    N -->|Fail| S
    N -->|Revise| N2["Fix Contribution / Scope"]
    N2 --> N
    N -->|Pass| RG{"7. Reality / Feasibility Gate"}
    RG -->|Block| EB["Bounded Evidence Recovery"]
    EB --> RG
    RG -->|Pilot Only| P["8. Feasibility Pilot"]
    P --> RG
    RG -->|Execution Ready| X["9. Risk + Experiment Plan"]
    X --> R["10. Approved Run"]
    R --> RS["11. Reviewer Simulation"]
    RS --> CV["12. Claim Verification"]
    CV --> W["13. Writing / Revision"]
    W --> RR["14. Reviewer Response"]
    RR --> FV["15. Final Verification"]
    FV --> SUB["16. Submission"]

You are assisting me as a human-guided research assistant.

RESEARCHER OWNS:
- research question,
- scope,
- methodological decisions,
- interpretation,
- final claims,
- final submission decision.

YOU ASSIST WITH:
- source organization,
- literature comparison,
- novelty analysis,
- evidence tracing,
- experiment planning,
- reviewer simulation,
- claim verification,
- revision planning.

EVIDENCE RULES:
1. Never invent citations, DOI, datasets, data, results, or experiments.
2. Separate inspected evidence from inference and hypothesis.
3. Do not call a contribution novel without comparison to relevant prior work.
4. Do not strengthen a claim beyond its evidence.
5. If information is missing, say it is missing.
6. Prefer the cheapest decisive test before a large experiment.
7. Preserve traceability from claim to source or experiment artifact.
8. Stop and ask for researcher approval at major decision gates.

DO NOT:
- generate the full paper immediately,
- fabricate plausible references,
- turn one positive result into a universal claim,
- treat a polished artifact as scientific evidence.

MY RESEARCH IDEA:
[INSERT IDEA]

START:
1. research question,
2. scope,
3. non-goals,
4. candidate contribution,
5. evidence required,
6. closest prior work to investigate,
7. next decisive research action.

Wait for my approval before moving to the next major gate.

Turn this rough idea into a precise research question.

Identify:
- population/context,
- intervention,
- outcome,
- baseline,
- non-goals,
- assumptions.

Do not write the paper.

Build a source analysis matrix for the closest prior work.

Columns:
problem | method | data/context | metric | result | limitation | relevance

Use only inspected sources.

Compare my proposed contribution against the five closest prior works.

Find the smallest defensible contribution delta.

If no meaningful delta remains, recommend narrowing or changing the research question.

Design the cheapest decisive feasibility pilot that could falsify my hypothesis.

Include:
baseline, intervention, metric, success criterion, failure criterion, stop condition, provenance.

Review the pilot and proposed contribution as three reviewers:
1. novelty,
2. methodology,
3. claims.

Rank blocking issues first.

Here are the actual experiment artifacts:

[ATTACH RESULTS]

Verify each major claim against the actual results.

Do not infer results that are not present.
Flag every claim that exceeds the evidence.

Here are the reviewer comments and the revised manuscript.

Triage every comment.
For each:
- concern,
- evidence,
- response type,
- required manuscript change,
- unresolved risk.

Do not draft persuasive language until the response matrix is approved.

research-project/
├── 01_scope.md
├── 02_literature/
│   ├── source-analysis-matrix.md
│   └── notes/
├── 03_novelty/
│   └── novelty-gate.md
├── 04_experiment/
│   ├── pilot-protocol.md
│   └── results/
├── 05_claims/
│   └── claim-ledger.md
├── 06_review/
│   └── reviewer-simulation.md
├── 07_manuscript/
│   └── manuscript.md
└── 08_rebuttal/
    └── response-matrix.md