Research Funding Proposals

Build a current, evidence-linked scientific argument around applicant-authored material. Historical applications and templates may inform non-expressive structure, but they are never current format authority and must not be copied into a public or applicant package.

The primary scope is NSFC and provincial funds. Guangdong is the bundled provincial example, not a rule set for every province. Other provinces use custom adaptation through the existing --program other option; there is no --program custom option or built-in preset for every province. Identify the actual province, funder, year, and project type, then use that program's current official guides, attachments, system template, and institutional requirements. Do not inherit Guangdong or NSFC eligibility, budgets, deadlines, codes, section formats, or AI-authoring permissions. Other funders remain supported through the generic workflow; the skill name stays writing-funding-proposals.

Start with policy and authority

Before substantive work, identify the target year, funder, program, project type, institutional instructions, current official sources, actual system template, confidentiality class, and requested scope. Read references/authority-and-policy.md and a matching bundled profile when available; for custom adaptation, use the actual program's current sources without assuming a provincial profile exists.

The authoring-policy mode controls what this skill may produce:

• EVIDENCE_AND_AUDIT_ONLY: organize sources and claims, compare topics, build an argument map, prepare section briefs, audit applicant-authored text, and return exact revision tasks. Do not produce insertion-ready application prose.
• ASSISTED_DRAFTING_ALLOWED: use only after a directly checked current funder and institutional source permits the requested assistance. Preserve the applicant's authorship, factual responsibility, required declarations, and review.
• UNVERIFIED: apply EVIDENCE_AND_AUDIT_ONLY until resolved.

The checked 2026 NSFC profile is EVIDENCE_AND_AUDIT_ONLY: the official notice says applications must not be directly generated with generative AI and generated research or reference information must be verified. Do not bypass that rule by calling generated prose a template, rewrite, translation, or polished final draft.

A bundled VERIFIED_PROFILE supports working-mode routing only. Final mode requires a new VERIFIED authoring-policy record with non-empty registered source IDs, a parseable checked date, authoring_policy scope, coverage for both funder and institution, and an explicit completed live refresh. Bind each policy source to one controlled authority_kind (funder or institution), project program and year, scope, checked date, and live-refresh ID; the binding must agree with the registered source row. Final validation rejects funder/institution bindings that reuse the same (issuer, url_or_path, clause_locator) identity; a joint document needs distinct clause locators. It defaults --as-of to today and requires every refresh date to equal it. Final authority, template, and financial-budget requirement sources must separately match the selected program/year, their controlled source-register scope/type, and as_of; a local template source path must equal the registered artifact path. Use an explicit historical --as-of only to revalidate an archived snapshot, never to claim that old evidence is current. Read references/authority-and-policy.md before recording this contract; it records local evidence, not an inferred official permission.

Create or resume a workspace

Resolve SKILL_DIR as the directory containing this SKILL.md. Every bundled script, reference, profile, and template is below that directory; never resolve through the package parent or assume the user's current working directory.

For a new project:

python3 "$SKILL_DIR/scripts/init_proposal_workspace.py" <project> \
  --project-id <id> --program <nsfc|guangdong|other> \
  --year <year> --project-type <type>

The initializer creates human-facing records for authority, evidence, topic selection, the scientific argument, section briefs, figures, conditional financial budget, commitments, reviews, and final-format evidence. It does not bundle an official form and refuses to overwrite a non-empty directory.

For an independently runnable archived fixture, see examples/README.md. It creates a synthetic workspace outside the installed skill, uses a fixed historical snapshot, and demonstrates only the local record-completeness check; it is never real policy, applicant, scientific, eligibility, or funding evidence.

On resume, read project.json, actual artifacts, the newest source entries and review findings, then run:

python3 "$SKILL_DIR/scripts/validate_proposal_workspace.py" <project> --mode working

Use observed files and validator output over stale status prose. Do not mark a stage complete because a neighboring artifact exists.

Full-project workflow

Read references/full-project-workflow.md. The stages are scientific dependencies, not a quota of forms:

Stage	Decision or artifact
S0	Current authority, funder-plus-institution authoring policy, material classification, and official-source register
S1	Claim ledger and applicant-role evidence
S2	Candidate portfolio with data and program feasibility
S3	Dated novelty search, scoop verdict, and strongest falsifier
S4	One frozen contribution lane and route or code fit
S5	Q/C/V/O/CL/F scientific argument graph
S6	Section briefs, artifact-specific commitments, conditional financial budget, paragraph jobs, page budget, risks, and annual outputs
S7	Applicant-authored root sections and evidence-preserving review
S8	Minimal sufficient editable figure and table portfolio
S9	Integration into the current official template without changing scientific scope
S10	Fatal, science, and readability review passes
S11	Actual build, PDF inspection, and final-size visual review
S12	Local candidate record-completeness check and explicit external-submission boundary

Continue through every safe dependent stage. Pause only for a genuine claim or route fork, missing load-bearing evidence, prohibited authoring action, confidential-data transfer, destructive change, or external submission.

Reference router

Task	Read
intake, annual refresh, policy, or format	references/authority-and-policy.md, any matching bundled profile, and the actual program's current sources
evidence, topic, novelty, route, or application code	references/evidence-topic-and-route.md
argument map, outline, section brief, or prose review	references/argument-and-sections.md and references/plain-chinese-prose.md
work-package commitment, annual output, risk, or fallback	references/commitment-calibration.md
technical-route figure, table, reviewer audit, PDF, or handoff	references/figures-review-and-delivery.md

For an AI-for-biology or computational-biology proposal, do not select an information- science code merely because AI or omics appears in the title. Compare the central new knowledge, natural reviewer community, claimed output, validation endpoint, and strongest mismatch for plausible alternative codes using current official labels.

Calibrate scientific commitments

Read references/commitment-calibration.md before turning the scientific graph into research contents, annual tasks, outputs, or risk statements. Classify each work package as exactly one of:

• deliverable mainline: a dependency-backed result the proposal commits to deliver and can assess with named acceptance evidence;
• cautious exploration: a potentially valuable hypothesis with a cheap falsifier, finite decision point, and useful negative outcome, not a guaranteed primary result;
• boundary confirmation: an early scope or feasibility check with an explicit go or no-go rule before downstream work depends on it.

Do not promote every idea to deliverable mainline merely to sound ambitious. Do not hide a required feasibility gate as cautious exploration, and do not present boundary confirmation as the final scientific contribution. Keep the class consistent across the argument map, section briefs, timeline, annual outputs, risks, and figures.

Scientific integrity rules

• Register every opened authority or scientific source with its scope and verification state. A search result, filename, remembered citation, or unopened PDF is not evidence.
• Give every load-bearing scientific, feasibility, applicant-foundation, and compliance claim a stable CL* record. Preserve numerical values, uncertainty, population scope, citation scope, and applicant role.
• Never infer a grant number, applicant contribution, data licence, access right, preliminary result, eligibility fact, or ethics status from a neighboring proposal, author position, publication count, or generic platform description.
• Compare topic candidates before prose expansion. Record data_feasibility, linkable variables, independent unit, strongest cheap baseline, discriminating validation, failure criterion, IDENTICAL/ADJACENT/UNCERTAIN, and FRAMING/DATA/TASK.
• Let the smallest dependency-complete Q/C/V/O/CL/F graph determine research-content and figure count. A funder heading does not require three symmetric aims.
• Separate completed work, transferable capability, planned work, and unavailable evidence. A polished sentence cannot repair an evidential gap.
• Keep restricted material local and out of fixtures, shared skills, examples, and public packages.

Figures and prose

This skill independently completes the proposal workflow within the verified authoring-policy mode; a prohibition on generated prose is not a missing skill. It first locks policy, evidence and the scientific argument. When current funder and institutional rules permit the requested AI-assisted figure, use the bundled standalone figures route: default to style-matched GPT Image 2.5 concepts followed by native editable PPTX using available tools directly, then proposal exports and visual QA. build-scientific-visualizations, when installed, adds optional design helpers; its absence does not stop that permitted default. Read figures, review, and delivery before production for policy, confidentiality, source-data, and capability boundaries. Its source ledger, semantic arrows, editability, and rendered QA apply; journal column widths do not. Derive the proposal width from the actual official template. A schematic is not proof of feasibility, and a decorative icon is not scientific evidence.

Run the prose audit on applicant-authored text:

python3 "$SKILL_DIR/scripts/audit_chinese_prose.py" <files>

The audit flags reviewable surface cues. It does not establish scientific accuracy, policy compliance, quality, or authorship. Inspect every finding in context and retain necessary official names, direct quotations, qualifiers, and citations.

Final-record completeness and external boundary

An official guide that mentions a form does not verify the actual form. Final mode requires the target-year system template, a completed funder-plus-institution authoring policy refresh, opened official sources, applicant review, frozen route, complete scientific graph, required conditional budget evidence, auditable commitments, three review passes, a successful local build, and visual inspection of the rendered artifact:

python3 "$SKILL_DIR/scripts/validate_proposal_workspace.py" <project> --mode final

The report declares validation_scope: record_completeness. It checks recorded statuses and local paths, not file format, rendering, or visual quality; the applicant still performs and records the real build and page-by-page inspection. It cannot determine eligibility, scientific truth, authorship, policy compliance, funding, or submission. Uploading, saving in a portal, or submitting is an external action that requires the applicant's explicit authorization and direct platform evidence.