nsfc-mianshang-review
Skill4 starsChanged 3 days agoMITView on GitHub
Use this skill as the workflow entry point for NSFC General Program proposal review. It coordinates PDF/TXT preparation, cache-first analysis, and staged use of the bundled review skills. For full-review tasks, follow this exact path before reading detailed references: Detailed rules are below. Use references/output-templates.txt for templates and references/scoring-rubric.txt when consistent ratings, A/B/C grades, attribute mismatch, or ranking support is needed. Use this skill when: Do not use this skill to: For a single explicitly specified proposal, continue through workflow…
In this repository10 other files it gives its agents.
The file
Show its rules and text
We split this skill into the rules it gives an agent
83 rules. About 0 are common, 42 are distinctive, and 41 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈42
Distinctive · estimate
This author's own choices. Worth a look.
41
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Distinctive: line 13
Use this skill as the workflow entry point for NSFC General Program proposal review.
- Distinctive: line 28
Before writing each stage file, attempt the explicit supporting skill call when the runtime exposes a Skill mechanism.
- Very specific: line 29
If a supporting skill cannot be called, continue with fallback, but record the exact provenance and use completed_with_protocol_violation …
- Very specific: line 31
Keep Kill-mode judgment in every full workflow by default.
- Very specific: line 32
When assigning A/B/C priority or comparing proposals, use references/scoring-rubric.txt: exact matching rubric dimensions, 0-5 raw …
- Distinctive: line 39
Use this skill when:
- Distinctive: line 49
Do not use this skill to:
- Very specific: line 63
If the runtime asks for tool permission, request the smallest useful action scope, preferably limited to reading proposal-derived TXT …
- Distinctive: line 67
If the user provides a PDF, local PDF-to-TXT extraction is the default entry behavior of this workflow.
- Distinctive: line 68
If the input is a PDF, read it only during the PDF-to-TXT conversion step.
- Distinctive: line 69
After TXT extraction succeeds, never read the PDF again.
- Distinctive: line 70
If the PDF is encrypted or text extraction is blocked, do not treat this as a normal failure.
- Distinctive: line 71
Only require human intervention for PDF preparation when one of the following is true:
- Very specific: line 76
Prefer .txt as the primary working copy.
- Distinctive: line 77
Run all generated workflow artifacts under the user-specified review directory.
- Distinctive: line 78
All code execution for this workflow must stay inside the user-provided review directory or the inferred proposal-local review directory.
- Distinctive: line 79
Do not write workflow artifacts to system folders, user profile cache locations, global skill directories, or agent home directories by …
- Distinctive: line 80
If a safe local review workspace cannot be created without writing to a protected or unrelated directory, stop and ask the user for a …
- Distinctive: line 82
Keep the proposal folder clean.
- Distinctive: line 112
Use two evidence chains:
- Very specific: line 136
Use references/scoring-rubric.txt for 0-5 scoring details, weights, cap rules, and A/B/C mapping.
- Very specific: line 141
Use the exact dimension names and weights from references/scoring-rubric.txt.
- Distinctive: line 142
Do not rename, merge, split, replace, or omit rubric dimensions.
- Very specific: line 146
Do not report the final weighted score only as a 0-5 average such as 2.4/5.
- Distinctive: line 150
Use this structure under the user-specified review directory:
- Distinctive: line 179
Create TXT artifacts first with:
- Distinctive: line 185
If the user explicitly provides a review directory, add:
- Very specific: line 191
If --workdir is omitted, the script creates nsfc-review-<proposal-stem> in the parent directory of the folder containing the proposal.
- Distinctive: line 199
Run this first on the proposal folder:
- Very specific: line 241
Only proposals with status: READY_FOR_REVIEW should enter model-based review.
- Distinctive: line 247
After batch preparation, stop the current batch session and tell the user to open a new agent window/session for review.
- Distinctive: line 249
Use one generated task card per new session:
- Very specific: line 270
Update _batch/review_status.tsv after every proposal:
- Distinctive: line 280
Do not let one failed proposal affect another proposal.
- Very specific: line 284
After model-based review, run review-completion QC on existing nsfc-review-* folders:
- Distinctive: line 294
Every generated stage file must start with a provenance header.
- Very specific: line 316
Use this closed vocabulary for skill_call_result:
- Very specific: line 329
Use success only after the supporting skill was actually invoked.
- Very specific: line 330
Use not_installed only after the runtime indicates the named skill is unavailable.
- Very specific: line 331
Use permission_blocked only when the runtime or user denies the call.
- Very specific: line 332
Use runtime_not_supported only when the environment has no skill-to-skill call mechanism.
- Very specific: line 333
Use call_failed only after a real call attempt fails.
- Very specific: line 334
Use not_attempted only when no call attempt was made.
- Very specific: line 373
Do not write status: completed until the file contains substantive analysis and no 待补充 placeholders.
- Very specific: line 374
Use execution_mode: explicit_skill_call only when the runtime actually exposed and invoked the supporting skill.
- Distinctive: line 375
Before writing a stage file, you must attempt an explicit supporting skill call when the runtime exposes a Skill mechanism.
- Very specific: line 376
Use attempted_skill_call: yes only after an actual call attempt.
- Very specific: line 377
Use execution_mode: fallback_agent_role when the current agent completed the role because explicit skill invocation was unavailable or …
- Very specific: line 378
If the supporting skill is installed or visible but no explicit call was attempted, do not write status: completed.
- Very specific: line 379
Do not pair attempted_skill_call: no with skill_call_result: runtime_not_supported.
- Very specific: line 380
If provenance is unknown, write execution_mode: fallback_agent_role and state the uncertainty in fallback_reason.
- Very specific: line 387
If the proposal is a PDF, the workflow should automatically begin by running scripts/extract_nsfc_text.py with an available Python runtime …
- Very specific: line 389
If the proposal is already TXT, copy or reference it as extracted/00_full_text.txt inside the review directory.
- Very specific: line 391
If no workdir is provided, infer one from the proposal path by placing nsfc-review-<proposal-stem> in the parent directory of the folder …
- Very specific: line 393
If extraction reports ENCRYPTED_PDF, stop and ask the user for the next input format.
- Distinctive: line 395
Do not redirect extraction outputs into system folders, agent cache locations, or global skill directories.
- Distinctive: line 399
Start each review with a short routing note:
- Distinctive: line 421
Treat explicit skill calls as protocol-required, not optional polish.
- Very specific: line 423
When fallback is used after a real attempt, set attempted_skill_call: yes, record skill_call_result, and write a concrete fallback_reason.
- Very specific: line 424
When no explicit attempt was made even though the skill was visible or likely installed, set attempted_skill_call: no, skill_call_result: …
- Very specific: line 426
Do not skip literature checks merely because literature-review is unavailable.
- Very specific: line 430
Before final judgment, create review/nsfc_review_cache.txt.
- Distinctive: line 492
Do aims answer the scientific question rather than list techniques?
- Distinctive: line 494
Do expected outcomes overreach the evidence?
- Distinctive: line 504
Verify the most important novelty claims.
- Distinctive: line 506
Check whether the proposed gap is real and current.
- Distinctive: line 507
Avoid broad literature review unless explicitly requested.
- Very specific: line 536
default Kill-mode decision section with the exact heading KILL-MODE DECISION.
- Very specific: line 559
Create review/06_submitted_review_comment.txt from review/05_final_review.txt.
- Distinctive: line 600
If the workflow did not inspect critical non-text figures directly, explicitly include a manual-review reminder covering:
- Distinctive: line 606
When the user is doing internal triage, pre-screening, kill-mode review, or asks for a stricter funding recommendation, apply a …
- Distinctive: line 608
Do not let general significance, available platforms, or decent preliminary basis outweigh structural flaws.
- Distinctive: line 609
Treat the following as potentially fatal unless the proposal itself clearly neutralizes them:
- Very specific: line 615
If two or more structural flaws remain after considering the proposal's own evidence, default to not recommended for support or an …
- Very specific: line 616
Use scientific-critical-thinking and reviewer-2-simulator as the primary drivers of fundability.
- Very specific: line 617
Use scientific-writing only to clarify and polish the judgment.
- Distinctive: line 668
Use this section to name the structural reasons the proposal should be rejected or deprioritized.
- Distinctive: line 672
Use the workflow as a state machine:
- Very specific: line 688
Use references/scoring-rubric.txt only when a consistent rating standard is needed.
- Very specific: line 689
Use references/output-templates.txt only when drafting file outputs or final review text.
- Distinctive: line 693
Do not generate a complete grant application for the applicant.
- Distinctive: line 694
Do not fabricate policy requirements, citations, project details, or reviewer criteria.
- Distinctive: line 695
If generative AI use is discussed, remind the user that NSFC 2026 requires human verification and truthful disclosure/labeling of …
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
NSFC General Program Review
Use this skill as the workflow entry point for NSFC General Program proposal review. It coordinates PDF/TXT preparation, cache-first analysis, and staged use of the bundled review skills.
Execution Fast Path
For full-review tasks, follow this exact path before reading detailed references:
- Convert the proposal PDF/TXT into
extracted/00_full_text.txt,extracted/01_section_index.txt, andextracted/extraction_report.txtby runningscripts/extract_nsfc_text.py --init-review-files. - Build or update
review/nsfc_review_cache.txtfrom TXT only. - Produce one file per supporting skill, in order:
01_scientific_critique.txtusingscientific-critical-thinking02_literature_checks.txtusingliterature-review03_reviewer2_stress_test.txtusingreviewer-2-simulator04_peer_review_draft.txtusingpeer-review05_final_review.txtusingscientific-writing06_submitted_review_comment.txtusingscientific-writingas a form-comment adapter
- Before writing each stage file, attempt the explicit supporting skill call when the runtime exposes a Skill mechanism.
- If a supporting skill cannot be called, continue with fallback, but record the exact provenance and use
completed_with_protocol_violationwhen a visible/installed skill was not attempted. - Identify the proposal's research attribute as
free_exploration,goal_oriented, orunclear, and include an attribute-match judgment in04_peer_review_draft.txt,05_final_review.txt, and06_submitted_review_comment.txt. - Keep Kill-mode judgment in every full workflow by default; the user does not need to request it separately.
04_peer_review_draft.txt,05_final_review.txt, and06_submitted_review_comment.txtmust contain the exact headingKILL-MODE DECISION. - When assigning A/B/C priority or comparing proposals, use
references/scoring-rubric.txt: exact matching rubric dimensions, 0-5 raw dimension scores, weighted points, total weighted score out of 100, cap rules, and ethics/compliance gate. - End with
06_submitted_review_comment.txt, a concise form-ready comment file matching NSFC review-table fields, preserving a briefKILL-MODE DECISION, and ending with a manual-review reminder.
Detailed rules are below. Use references/output-templates.txt for templates and references/scoring-rubric.txt when consistent ratings, A/B/C grades, attribute mismatch, or ranking support is needed.
When to Use This Skill
Use this skill when:
- reviewing NSFC General Program / 面上项目 proposals;
- running internal pre-submission triage or kill-mode review;
- extracting PDF proposal text into staged review artifacts;
- evaluating novelty, scientific logic, feasibility, risk, and funding priority;
- producing polished Chinese review comments from staged evidence files.
When Not to Use This Skill
Do not use this skill to:
- generate a full grant application for an applicant;
- review unrelated journal manuscripts when NSFC proposal structure is irrelevant;
- perform budget-only, attachment-only, or administrative-only checks;
- inspect confidential non-text figures unless the user explicitly provides safe access;
- bypass human expert review for final funding, ethics, biosafety, or confidentiality decisions.
Non-Interactive Execution Policy
For a single explicitly specified proposal, continue through workflow stages automatically unless extraction fails, a required password is missing, the output directory is unsafe, or a safety/compliance issue appears. Do not stop at normal single-proposal stage boundaries to ask the user to say go.
For batch workflows, this non-interactive policy applies only to extraction/preparation QC and review-task-card generation. Do not perform model-based review for multiple proposals in the same agent window/session. Each queued proposal should be reviewed in a fresh agent window/session using its generated task card.
If the runtime asks for tool permission, request the smallest useful action scope, preferably limited to reading proposal-derived TXT files, running scripts/extract_nsfc_text.py, and writing artifacts under the review directory.
Non-Negotiable File Rules
- If the user provides a PDF, local PDF-to-TXT extraction is the default entry behavior of this workflow. Do not require manual step-by-step user preparation unless extraction is blocked.
- If the input is a PDF, read it only during the PDF-to-TXT conversion step.
- After TXT extraction succeeds, never read the PDF again. All later review work must use TXT files, cache files, and short copied evidence snippets.
- If the PDF is encrypted or text extraction is blocked, do not treat this as a normal failure. Ask the user which format they want to provide next: password, decrypted PDF, exported TXT, or OCR text.
- Only require human intervention for PDF preparation when one of the following is true:
- the PDF is encrypted and no password is available;
- the PDF is image-based and OCR text is not available;
- the extracted text is too incomplete or too corrupted to support review;
- a safe local review directory cannot be created.
- Prefer
.txtas the primary working copy. Markdown is optional and only for human-facing summaries. - Run all generated workflow artifacts under the user-specified review directory. Do not write to system, global agent, or package/cache directories.
- All code execution for this workflow must stay inside the user-provided review directory or the inferred proposal-local review directory.
- Do not write workflow artifacts to system folders, user profile cache locations, global skill directories, or agent home directories by default.
- If a safe local review workspace cannot be created without writing to a protected or unrelated directory, stop and ask the user for a review directory.
- The user should only need to provide the proposal path and, when needed, the PDF password. If the user does not provide a review directory, create a clean proposal-specific review folder in the parent directory of the proposal folder. Example: for
D:\AI projects\claude_code\2026国自然评审\项目书\proposal.pdf, useD:\AI projects\claude_code\2026国自然评审\nsfc-review-proposal\. - Keep the proposal folder clean. Do not write extracted or review files into the proposal folder unless the user explicitly asks.
- Chinese source text is expected. It does not reduce workflow quality. Preserve scientific terms accurately and produce final review text in polished Chinese by default.
Scope
Review these components:
- Basic project identity: title, field, keywords, research attribute, duration, collaborators.
- Title.
- Chinese abstract.
- Project rationale: scientific value, research gap, significance, domestic/international status, key scientific question.
- Research content: objectives, scientific questions, tasks, technical route, distinctive features, innovation, annual plan.
- Research basis: prior work, preliminary data, feasibility analysis, risks and mitigation, and working conditions directly relevant to feasibility.
- Related ongoing/completed projects only when they affect overlap, continuity, feasibility, or independence.
- Applicant/team background only when it affects ability to complete the proposed work.
- Representative papers only when they support preliminary basis, novelty, or feasibility.
- Non-text visual materials only through explicit human review. This includes mechanism schematics, technical workflow diagrams, and experimental result figures that are not faithfully captured by text extraction.
Exclude low-value administrative review by default: budget details, attachment directory mechanics, unit-consistency declarations, recommendation-letter logistics, routine certificates, and generic form-filling checks.
Also treat figure-heavy, non-text proposal content as a manual-review boundary by default. If key claims depend on images, diagrams, gels, microscopy panels, model schematics, or workflow figures that were not reliably converted into text, the final review must explicitly say those materials still require human inspection.
Research Attribute And Scoring
For every full review, record the applicant-selected research attribute and judge whether the proposal content matches it:
free_exploration: curiosity/frontier-driven basic research.goal_oriented: basic research driven by economic/social/national/clinical need.unclear: attribute is missing or cannot be confidently extracted.
Use two evidence chains:
free_exploration:
academic frontier or unknown phenomenon
-> basic scientific question
-> original hypothesis/mechanism/principle
-> generalizable new knowledge
goal_oriented:
economic/social/national/clinical need
-> key bottleneck
-> underlying unknown basic scientific problem
-> mechanistic research in this proposal
-> generalizable new knowledgeAttribute mismatch levels:
none: selected attribute and proposal logic are aligned.mild: wording is imperfect, but the scientific logic supports the selected attribute.moderate: label and main logic are visibly misaligned, but a scientific question still exists; maximum grade B.severe: the selected attribute is central to the proposal logic but the chain fails; treat as a decision-level defect and usually grade C.
Use references/scoring-rubric.txt for 0-5 scoring details, weights, cap rules, and A/B/C mapping. Do not treat the weighted score as automatic: cap rules, ethics/compliance gates, and Kill-mode judgment can override the numeric total.
Scoring output is low-freedom and must be reproducible:
- Select exactly one rubric table:
free_explorationorgoal_oriented. - Use the exact dimension names and weights from
references/scoring-rubric.txt. - Do not rename, merge, split, replace, or omit rubric dimensions.
- For each dimension, output
raw_score / 5,weight, andweighted_points = raw_score / 5 * weight. - Output
Total weighted score: <x>/100. - Output
Grade before cap rules,Cap rules triggered,Final grade after cap rules, andFunding opinion. - Do not report the final weighted score only as a 0-5 average such as
2.4/5; this is not sufficient for ranking or cross-proposal comparison.
Workflow Artifacts
Use this structure under the user-specified review directory:
review_dir/
extracted/
00_full_text.txt
01_section_index.txt
extraction_report.txt
review/
nsfc_review_cache.txt
01_scientific_critique.txt
02_literature_checks.txt
03_reviewer2_stress_test.txt
04_peer_review_draft.txt
05_final_review.txt
06_submitted_review_comment.txtFor batch jobs, also use:
work_root/
_batch/
manifest.tsv
extraction_qc_report.txt
review_queue.tsv
review_status.tsvCreate TXT artifacts first with:
<python> scripts/extract_nsfc_text.py --proposal <proposal.pdf-or-txt> [--password <pdf-password>] --init-review-filesIf the user explicitly provides a review directory, add:
--workdir <review_dir>If --workdir is omitted, the script creates nsfc-review-<proposal-stem> in the parent directory of the folder containing the proposal. --init-review-files pre-creates the review-stage skeleton files under review/ so the workflow can continue without a separate file-creation prompt.
Batch Workflow
For multiple proposals, split the job into deterministic preparation and one-proposal-per-session model review. Batch mode prepares files and task cards only; it must not perform model-based review for multiple proposals in the same session.
Batch Step 1: Extract And QC
Run this first on the proposal folder:
<python> scripts/batch_extract_and_qc.py --proposal-dir <proposal_dir>Optional arguments:
--work-root <work_root>
--batch-dir <work_root>/_batch
--password <shared_password>
--password-file <passwords.tsv>
--limit <N>passwords.tsv should be UTF-8 tab-separated with columns:
file password
proposal_a.pdf password-a
proposal_b.pdf password-bThis step must not perform scientific review. It only creates or updates nsfc-review-* folders, initializes review skeletons, and writes:
_batch/manifest.tsv_batch/extraction_qc_report.txt_batch/review_queue.tsv_batch/review_status.tsv_batch/review_task_index.tsv_batch/review_prompts.txt_batch/review_tasks/<index>_<proposal>.md
QC checks include:
00_full_text.txtexists and has sufficient text volume;01_section_index.txtcontains title, Chinese abstract, project rationale, research content, and research basis;- extraction report does not show encrypted/missing-dependency/unsupported-input errors;
- review skeleton files exist;
- review skeleton files contain required provenance headers.
Only proposals with status: READY_FOR_REVIEW should enter model-based review. Each such proposal receives a task card under _batch/review_tasks/.
This is extraction/preparation QC only. It does not certify that model-based review has completed.
Batch Step 2: New Session Review Tasks
After batch preparation, stop the current batch session and tell the user to open a new agent window/session for review.
Use one generated task card per new session:
读取并执行:<work_root>/_batch/review_tasks/001_<proposal>.mdThe task card contains the review directory, allowed input files, required stage files, scoring constraints, completion QC command, and a stop-after-this-project instruction.
The first task should be treated as the pilot review. After it finishes, the human check should confirm:
- extracted full text is complete enough;
- section index is aligned;
- cache is decision-relevant, not a full-text summary;
- stage files are substantive and have
status: completed; - provenance headers correctly report
explicit_skill_callorfallback_agent_role; - final review is polished Chinese and contains manual-review boundaries.
Batch Step 3: Repeat One Task Per New Session
For each remaining queued proposal, open another fresh agent window/session and execute the next task card. Do not ask one agent session to review several proposals in sequence.
Update _batch/review_status.tsv after every proposal:
NOT_STARTED
IN_PROGRESS
COMPLETED
FAILED
NEEDS_HUMAN_REVIEWDo not let one failed proposal affect another proposal. Record the failure reason and start the next task only in a fresh session.
Batch Step 4: Review Completion QC
After model-based review, run review-completion QC on existing nsfc-review-* folders:
<python> scripts/review_completion_qc.py --work-root <work_root>This checks whether 01 through 06 review files are completed, whether supporting skill provenance is explicit/fallback/protocol-violating, and whether the final review contains Kill-mode judgment, manual-review reminders, a funding recommendation, and a form-ready submitted review comment. This step is separate from extraction QC so batch preparation remains fast and deterministic.
Required Provenance Headers
Every generated stage file must start with a provenance header. Keep the header when editing the file, and update it before marking the stage complete.
For 01_scientific_critique.txt, 02_literature_checks.txt, and 03_reviewer2_stress_test.txt, use:
NSFC_REVIEW_STAGE_PROVENANCE
workflow_skill: nsfc-mianshang-review
workflow_version: 0.5.6
stage: <scientific_critique | literature_checks | reviewer2_stress_test>
supporting_skill: <scientific-critical-thinking | literature-review | reviewer-2-simulator>
attempted_skill_call: <yes | no>
skill_call_result: <success | unavailable | failed | permission_blocked | runtime_not_supported | not_attempted>
execution_mode: <explicit_skill_call | fallback_agent_role | unavailable_not_attempted>
status: <completed | completed_with_protocol_violation | partial | blocked | pending>
input_files:
- extracted/00_full_text.txt
- review/nsfc_review_cache.txt
output_file: review/<stage_file_name>.txt
generated_at: <YYYY-MM-DD HH:MM local time, or unknown>
fallback_reason: <none | supporting skill unavailable | skill call failed | runtime does not support skill-to-skill routing | permission blocked | skill was installed but not called | other>Use this closed vocabulary for skill_call_result:
success
not_installed
permission_blocked
runtime_not_supported
call_failed
not_attemptedRules for these values:
- Use
successonly after the supporting skill was actually invoked. - Use
not_installedonly after the runtime indicates the named skill is unavailable. - Use
permission_blockedonly when the runtime or user denies the call. - Use
runtime_not_supportedonly when the environment has no skill-to-skill call mechanism. - Use
call_failedonly after a real call attempt fails. - Use
not_attemptedonly when no call attempt was made.not_attemptedplus substantive output must be treated as a protocol issue unless the runtime truly lacks any Skill mechanism.
For 04_peer_review_draft.txt and 05_final_review.txt, use the same single-supporting-skill stage header pattern as above.
04_peer_review_draft.txt must use:
stage: peer_review_draft
supporting_skill: peer-review
output_file: review/04_peer_review_draft.txt
input_files:
- review/nsfc_review_cache.txt
- review/01_scientific_critique.txt
- review/02_literature_checks.txt
- review/03_reviewer2_stress_test.txt05_final_review.txt must use:
stage: final_scientific_writing
supporting_skill: scientific-writing
output_file: review/05_final_review.txt
input_files:
- review/04_peer_review_draft.txt06_submitted_review_comment.txt must use:
stage: submitted_review_comment
supporting_skill: scientific-writing
output_file: review/06_submitted_review_comment.txt
input_files:
- review/05_final_review.txtRules:
- Do not write
status: completeduntil the file contains substantive analysis and no待补充placeholders. - Use
execution_mode: explicit_skill_callonly when the runtime actually exposed and invoked the supporting skill. - Before writing a stage file, you must attempt an explicit supporting skill call when the runtime exposes a Skill mechanism.
- Use
attempted_skill_call: yesonly after an actual call attempt. Do not use it for mental simulation or role-playing. - Use
execution_mode: fallback_agent_rolewhen the current agent completed the role because explicit skill invocation was unavailable or failed. - If the supporting skill is installed or visible but no explicit call was attempted, do not write
status: completed; writestatus: completed_with_protocol_violation. - Do not pair
attempted_skill_call: nowithskill_call_result: runtime_not_supported; usenot_attemptedunless the runtime itself explicitly reports no Skill mechanism. - If provenance is unknown, write
execution_mode: fallback_agent_roleand state the uncertainty infallback_reason; do not imply a skill call occurred. - Downstream ranking depends on these headers. Missing or vague provenance should be treated as lower input-quality evidence.
Required Workflow
1. Prepare Text Working Copy
If the proposal is a PDF, the workflow should automatically begin by running scripts/extract_nsfc_text.py with an available Python runtime to create extracted/00_full_text.txt and extracted/extraction_report.txt.
If the proposal is already TXT, copy or reference it as extracted/00_full_text.txt inside the review directory.
If no workdir is provided, infer one from the proposal path by placing nsfc-review-<proposal-stem> in the parent directory of the folder containing the proposal.
If extraction reports ENCRYPTED_PDF, stop and ask the user for the next input format. Do not continue reviewing from the PDF. This is an action-required state, not a workflow failure.
Do not redirect extraction outputs into system folders, agent cache locations, or global skill directories. If the inferred path is unsafe, ask for a local review directory before running extraction.
2. Route Before Deep Review
Start each review with a short routing note:
- Task type: full review, section review, pre-submission self-check, or revision planning.
- Available artifacts: TXT, section index, cache, final review draft.
- Skills to use and why.
- Excluded low-value administrative materials.
Default skill routing:
- Title and Chinese abstract:
peer-reviewfor quick proposal positioning and overall framing. - Project rationale:
scientific-critical-thinkingplus targetedliterature-review. - Research content:
scientific-critical-thinking,reviewer-2-simulator, andpeer-review. - Research basis:
scientific-critical-thinkingandpeer-review. scientific-critical-thinking: scientific question, causal logic, design rigor, statistics, feasibility, risk, overclaims.literature-review: targeted novelty, gap, and field-context checks.reviewer-2-simulator: harsh pre-submission stress test and likely reviewer objections.peer-review: structured grant-style evaluation and final organization.scientific-writing: polished Chinese or English review comments, revision advice, or response text.
Supporting skill routing policy:
- At each stage, first attempt to explicitly load or call the named supporting skill if the runtime exposes a Skill mechanism.
- Treat explicit skill calls as protocol-required, not optional polish. Do not choose fallback merely because the current agent can complete the role itself.
- Fallback is allowed only after one of these conditions is true: the supporting skill is not installed, the runtime has no Skill mechanism, the call is rejected by permission policy, or the attempted call fails.
- When fallback is used after a real attempt, set
attempted_skill_call: yes, recordskill_call_result, and write a concretefallback_reason. - When no explicit attempt was made even though the skill was visible or likely installed, set
attempted_skill_call: no,skill_call_result: not_attempted,execution_mode: fallback_agent_role, andstatus: completed_with_protocol_violation. - The workflow should still produce substantive files after fallback, but provenance must make the protocol deviation visible.
- Do not skip literature checks merely because
literature-reviewis unavailable. Use available web search, PubMed/Entrez tools, or concise manual query synthesis for targeted novelty checks.
3. Build Or Update Review Cache
Before final judgment, create review/nsfc_review_cache.txt. Do not summarize the entire proposal. Extract only decision-relevant content.
Cache schema:
NSFC GENERAL PROGRAM REVIEW CACHE
BASIC INFORMATION
Title:
Field:
Research attribute: free exploration / goal-oriented / unclear
Duration:
Collaborators:
CORE SCIENTIFIC QUESTION
CENTRAL HYPOTHESIS OR MAIN LOGIC
SIGNIFICANCE CLAIMS
- Claim:
Evidence in application:
Concern:
RESEARCH OBJECTIVES AND CONTENT
- Objective/Aim:
Methods:
Expected result:
Risk:
CLAIMED INNOVATION
- Innovation claim:
Basis:
Concern:
TECHNICAL ROUTE
PRELIMINARY BASIS AND FEASIBILITY
- Evidence:
Supports:
Limitation:
RISKS AND ALTERNATIVES
APPLICANT/TEAM FIT
RELATED PROJECT OVERLAP OR CONTINUITY
KEY LITERATURE CLAIMS TO VERIFY
CANDIDATE MAJOR CONCERNS
CANDIDATE MINOR CONCERNSReuse this cache in later steps. Return to 00_full_text.txt only for missing details or short evidence quotes.
4. Scientific Critique
Explicitly use scientific-critical-thinking on the cache plus targeted TXT snippets when available. If unavailable, complete this stage in the current agent using the same role:
- Is the central question a real basic-science question?
- Does the research attribute match the proposal content?
- Do aims answer the scientific question rather than list techniques?
- Are controls, sample sizes, statistical plans, model systems, validation paths, and endpoints adequate?
- Do expected outcomes overreach the evidence?
- Are risk and alternative plans concrete enough?
Save concise findings to review/01_scientific_critique.txt when producing files.
The file must begin with NSFC_REVIEW_STAGE_PROVENANCE, with stage: scientific_critique, supporting_skill: scientific-critical-thinking, and the correct execution_mode.
5. Targeted Literature Checks
Explicitly use literature-review only for key claims when available. If unavailable, complete this stage in the current agent with targeted web/literature searches:
- Verify the most important novelty claims.
- Identify obvious missing work or competing explanations.
- Check whether the proposed gap is real and current.
- Avoid broad literature review unless explicitly requested.
Save results to review/02_literature_checks.txt when producing files.
The file must begin with NSFC_REVIEW_STAGE_PROVENANCE, with stage: literature_checks, supporting_skill: literature-review, and the correct execution_mode.
6. Stress Test
Explicitly use reviewer-2-simulator after the cache and scientific critique when available. If unavailable, complete this stage in the current agent as a strict reviewer-2 role:
- Produce the strongest score-lowering arguments.
- Focus on top issues that would matter to NSFC reviewers.
- Convert vague concerns into concrete fixable problems.
- Explicitly distinguish between revisable weaknesses and structural flaws that should materially lower fundability.
Save results to review/03_reviewer2_stress_test.txt when producing files.
The file must begin with NSFC_REVIEW_STAGE_PROVENANCE, with stage: reviewer2_stress_test, supporting_skill: reviewer-2-simulator, and the correct execution_mode.
7. Peer Review Draft
Explicitly use peer-review to structure the grant-style judgment and save it to review/04_peer_review_draft.txt. If unavailable, complete the equivalent role in the current agent and record fallback provenance.
The draft must synthesize nsfc_review_cache.txt, 01_scientific_critique.txt, 02_literature_checks.txt, and 03_reviewer2_stress_test.txt. Do not bypass these stage files when producing the judgment.
The draft must include:
- research attribute match judgment;
- weighted scoring summary using the exact matching rubric dimensions and weights;
- ethics/compliance gate;
- cap-rule checks;
- default Kill-mode decision section with the exact heading
KILL-MODE DECISION.
The user does not need to ask for Kill-mode separately.
8. Final Scientific Writing
Explicitly use scientific-writing to polish review/04_peer_review_draft.txt into review/05_final_review.txt. If unavailable, complete the equivalent role in the current agent and record fallback provenance.
The final review must be written in polished Chinese unless the user explicitly requests another language. Keep the tone professional, restrained, concrete, and suitable for NSFC-style review or pre-submission internal review.
scientific-writing may improve clarity, structure, tone, concision, and Chinese expression, but it must not change the evidence judgment, reverse the funding recommendation, or soften the Kill-mode decision produced by the peer-review draft.
scientific-writing must preserve the scoring table from 04_peer_review_draft.txt with exact rubric dimensions, weights, raw scores, weighted points, total /100, grade before cap rules, cap rules, final grade, and funding opinion. It may polish the wording in the evidence-basis column, but must not collapse the scoring table into a 0-5 average or replace it with generic grant-review dimensions.
05_final_review.txt must contain the exact heading KILL-MODE DECISION. Under that heading, include:
- decision-level defects;
- number of structural flaws;
- cap-rule effect on the grade;
- strict internal recommendation.
9. Submitted Review Comment
Create review/06_submitted_review_comment.txt from review/05_final_review.txt. This file is for copying into the official review form, so it must be concise, polished Chinese, and organized by the form fields.
Include these fields:
COMPREHENSIVE EVALUATION
- Choose one: A excellent / B good / C medium / D poor
FUNDING OPINION
- Choose one: A priority funding / B fundable / C not funded
KILL-MODE DECISION
- Brief internal triage summary. Preserve the strict recommendation from `05_final_review.txt`.
SCIENTIFIC EVALUATION DESCRIPTION
- Use the official definition for the selected research attribute:
- free_exploration: originality/frontier basic research from curiosity or academic inspiration, not aimed at current application needs.
- goal_oriented: basic research driven by economic/social development needs or national needs.
SPECIFIC REVIEW COMMENTS
1. Attribute-specific question 1.
2. Attribute-specific question 2.
3. Attribute-specific question 3.
MANUAL REVIEW REMINDERFor free_exploration, the three specific comments should answer:
- The innovation of the research idea and the reasoning.
- The value of the scientific question and contribution to relevant frontiers.
- The applicant's innovation potential, feasibility of the research plan, and suggestions for improvement.
For goal_oriented, the three specific comments should answer:
- Whether the project addresses a basic scientific question behind economic/social development needs or national needs, with reasoning.
- The innovation of the scientific question and the scientific value of expected results.
- The applicant's innovation potential, feasibility of the research plan, and suggestions for improvement.
End the file with a reminder that text extraction cannot fully assess non-text figures, technical route diagrams, mechanism models, experimental images, representative papers, CV details, ethics/biosafety/compliance materials, and other form attachments; these require human review before submission.
If the workflow did not inspect critical non-text figures directly, explicitly include a manual-review reminder covering:
- mechanism diagrams or model schematics,
- technical workflow diagrams,
- experimental result figures, microscopy panels, gels, blots, plots, or other image-based evidence.
When the user is doing internal triage, pre-screening, kill-mode review, or asks for a stricter funding recommendation, apply a conservative decision rule:
- Do not let general significance, available platforms, or decent preliminary basis outweigh structural flaws.
- Treat the following as potentially fatal unless the proposal itself clearly neutralizes them:
- the core novelty claim is weak, already substantially occupied by prior work, or mainly rhetorical;
- the main mechanistic chain is not designed to close with decisive experiments;
- the proposal is overloaded, diffuse, or tries to do too many partially connected things;
- a high-risk hypothesis is central to the application but has weak direct preliminary support;
- translational or exploratory extensions materially dilute the core basic-science question.
- If two or more structural flaws remain after considering the proposal's own evidence, default to
not recommended for supportor an equivalent negative internal decision, even if the topic is important and the team appears capable. - Use
scientific-critical-thinkingandreviewer-2-simulatoras the primary drivers of fundability;peer-reviewshould organize the decision, not soften it. - Use
scientific-writingonly to clarify and polish the judgment. It must not reverse a negative decision produced by the evidence review.
Recommended output:
OVERALL ASSESSMENT
RESEARCH ATTRIBUTE MATCH
STRENGTHS
MAJOR CONCERNS
MINOR CONCERNS
FEASIBILITY AND RISK ASSESSMENT
INNOVATION ASSESSMENT
SCORING SUMMARY AND CAP RULES
Use this required format:
```text
Rubric used: free_exploration / goal_oriented
| Dimension exactly from scoring-rubric.txt | Weight | Raw score /5 | Weighted points | Evidence basis |
| ... | ... | ... | ... | ... |
Total weighted score: <x>/100
Grade before cap rules: A / B / C
Cap rules triggered: <none or list>
Final grade after cap rules: A / B / C
Funding opinion: A priority funding / B fundable / C not fundedETHICS AND COMPLIANCE GATE
SUGGESTIONS FOR REVISION
FUNDING RECOMMENDATION OR PRIORITY
For each major concern, cite the relevant section or cache item. Use short quotes only when necessary.
For every full workflow run, add one short section before the funding recommendation. Use this exact heading in `04_peer_review_draft.txt`, `05_final_review.txt`, and `06_submitted_review_comment.txt`:
```text
KILL-MODE DECISIONUse this section to name the structural reasons the proposal should be rejected or deprioritized. Do not bury decision-level defects inside minor wording suggestions.
Automation Pattern
Use the workflow as a state machine:
PDF_OR_TXT_INPUTTXT_READYCACHE_READYSCIENTIFIC_CRITIQUE_READYLITERATURE_CHECKS_READYSTRESS_TEST_READYPEER_REVIEW_DRAFT_READYFINAL_REVIEW_READYSUBMITTED_REVIEW_COMMENT_READY
At the start of every turn, identify the latest completed state from files in the review directory, then continue from the next state. Do not restart from the PDF once TXT exists.
References
- Use
references/scoring-rubric.txtonly when a consistent rating standard is needed. - Use
references/output-templates.txtonly when drafting file outputs or final review text.
Safety And Integrity
- Do not generate a complete grant application for the applicant.
- Do not fabricate policy requirements, citations, project details, or reviewer criteria.
- If generative AI use is discussed, remind the user that NSFC 2026 requires human verification and truthful disclosure/labeling of AI-assisted content.
- Flag ethical, biosafety, sensitive-information, or confidentiality risks only when they affect scientific review or compliance.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

