econ-writing-workflow

name

econ-writing-workflow

description

Use as the main entry point for economics writing workflows in English or Chinese, including paper and research-proposal planning, scope and claim convergence, full paper drafting from research questions and result packages, major revision, document-level translation or compression, exhaustive semantic QA and final audits, abstract/introduction/results revision, Chinese top-journal adaptation, English and Chinese diction cleanup, AI/translationese removal, table and figure design, notes/captions, formatting checks, and routing to empirical workflow support when data cleaning, regressions, or variable construction are involved.

Economics Writing Workflow

Role

Use this skill as the user's main entry point for economics writing work. It does not replace the specialized skills; it routes the task to the right writing module, keeps the workflow ordered, and checks that the final output matches the user's target.

This skill is not a full empirical pipeline. For data cleaning, variable construction, regression code, estimation, diagnostics, or reproducibility tasks, pair with empirical-econ-workflow.

Project-Level Skill Pinning

For multi-step economics writing, table, figure, journal-submission, or paper-formatting workflows, check whether the nearest project-level AGENTS.md or CLAUDE.md already contains an economics-writing skill routing rule.

If no such rule exists, ask the user once whether to add a persistent local rule. Do not edit AGENTS.md or CLAUDE.md without explicit user approval, and do not ask for tiny one-off requests.

Recommended rule text:

## Economics Writing Skill Routing

For economics paper writing, journal adaptation, tables, figures, captions, notes, Word/PDF formatting, and submission checks, first route through `econ-writing-workflow`.

Use `cn-top-econ-writing` for Chinese top-journal prose and journal-specific submission style. Use `econ-table-figure-design` for tables, figures, maps, captions, notes, and export quality. Use `empirical-econ-workflow` when data cleaning, variable construction, regressions, sample audits, or estimation code are involved.

Keep the relevant skill active throughout the task unless the user explicitly opts out. Do not treat skill use as a one-time setup step: at the start of each substantive phase, and whenever the work crosses from prose into tables, figures, maps, journal formatting, data facts, or empirical code, reclassify the current subtask and load the relevant specialized skill or reference file.

During manuscript drafting or revision, separate paper-facing text from author memos and task logs before writing each paragraph, table note, figure note, appendix note, or footnote. Only reader-facing manuscript text may enter the paper. Author workflow notes, submission-positioning discussion, reasons for moving/deleting material, and author-agent internal decisions must stay in the response, task README, revision memo, or checklist.

Before drafting claim-bearing prose, determine whether the author has explicitly fixed the intended meaning for the requested scope. If not, inspect the available context, ask the smallest grouped set of material questions, restate the intended meaning for confirmation, and wait. Do not draft the affected prose or use a `TODO` to bypass unresolved author intent. A `TODO` may represent a missing factual detail only when it does not change the already confirmed meaning.

Before drafting each paper-facing paragraph, apply draft-time argument clarity: ground abstract concepts in observable/model objects, state the comparison for relative or causal claims, keep design language separate from findings, distinguish mechanisms from plain heterogeneity, position close literature by the margin advanced, and keep terminology stable.

Before drafting literature-dependent claims, check whether the project has local reference papers, literature notes, `.bib` files, or source ledgers. Ask whether the literature is sufficient and whether the user wants help finding publicly available papers. If a Chinese or paywalled source cannot be obtained locally, tell the user and continue only with unaffected content. A non-semantic citation or verification detail may remain as an author-facing `TODO`; if the missing source affects literature positioning, theory, mechanism, variable meaning, identification, contribution, policy implication, or claim strength, return `evidence_conflict` and leave the affected manuscript text unwritten. When the paper uses literature-based data, theories, variables, mechanisms, or empirical choices, read the relevant source material and align with it instead of inventing a new framing.

Preserve evidence boundaries without turning them into defensive disclaimers. Prefer an accurate claim type, calibrated verb, and the minimum scope qualifier needed for the reader. Before adding even the first or only stand-alone negative caveat, state affirmatively what the result, estimate, threshold, scenario, or model object represents and what interpretation it supports. This is explanatory wording, not favorable spin. Admit a separate negative sentence only when it blocks a concrete material misreading not already excluded by the calibrated wording; bind it to the exact claim or number and use it as a last resort. Repeat it only when the identification, sample, period, geography, extrapolation target, or evidence grade changes, or when a stand-alone abstract, caption, note, journal rule, or referee request requires self-contained wording. Do not impose a mechanical per-section caveat quota.

Do not let the model or a reviewer self-authorize that last resort. Unless the exact negative proposition is already confirmed and frozen by the author, withhold it from manuscript prose and return `clarification_required` with the affirmative alternative and proposed contrast.

When writing regression results, main findings, mechanisms, heterogeneity, robustness, abstracts, introductions, or conclusions that mention coefficient size, translate estimates into economic magnitude using at least one appropriate benchmark. Prefer natural units or real policy changes when available; use means, standard deviations, or percentile spreads only when they fit the variable and identification source. Never invent descriptive statistics.

When producing tables, figures, maps, or edited images, load `econ-table-figure-design/references/labels/01_reader_facing_labels.md` and apply its reader-facing label gate before treating the task as complete. Verify the final rendered artifact itself: text, legends, colorbars, inset maps, axes, notes, and substantive visual content must be intelligible and must not overlap, obscure, or cover useful information.

After adding the rule, include a clear marker heading so future agents can detect it and avoid asking again.

Paper Plan And Artifact Lifecycle

For multi-step paper work or a major paper section, first locate the project's single authoritative current-plan store and current-plan document. If none exists, ask the author whether to establish it before substantive production. Do not create parallel current-plan files.

Read the authoritative current plan before each substantive phase. Keep only the latest accepted objective and claim sequence, admitted evidence and paper role, main-text or appendix placement, reader-facing terminology, unresolved decisions, current non-goals, and authorization boundaries. When the author approves, rejects, narrows, or replaces a plan, update the current plan in place during the same task. Keep historical process in task logs, manifests, or audit packages rather than competing current-plan documents.

Develop exploratory and candidate tables, figures, maps, measures, specifications, model outputs, and other research objects outside the centralized current-artifact store. Track two dimensions separately:

• artifact_state: exploratory, candidate, author-confirmed, frozen-current, rejected, or superseded;
• manuscript_state: not-integrated or manuscript-integrated.

Never infer a later state or approval from an earlier one. An object may remain frozen-current while also being manuscript-integrated; a revision that replaces it may become superseded while an older manuscript still cites that revision.

Before moving to a new substantive phase, update the authoritative current plan and check whether an author-confirmed object still lacks an explicit freeze decision. If so, ask whether to freeze it as the sole current revision. Do not repeatedly ask after the project record contains the decision, and do not freeze a merely plausible, validated, or polished candidate.

For every freeze, replacement, rejection, or manuscript integration, load and follow references/artifact-lifecycle/01_artifact_state_and_freeze_protocol.md. That protocol requires a machine-readable registry, a complete freeze manifest, SHA-256 coverage for code, inputs, outputs, and validation records, and a verified immutable audit bundle before any current revision is replaced. Do not report a freeze complete while a required manifest, checksum, approval, or recovery artifact is missing.

Keep the centralized current-artifact store current-state only. Build and validate replacement candidates outside it. Before updating stable current paths, preserve and verify the old frozen revision in the superseded audit store; then switch the replacement into the stable paths and update the registry. Do not accumulate competing current names such as v2, final2, new, or latest, and do not treat current-state cleanup as permission to delete audit history or original inputs.

Treat substantive approval, freeze approval, replacement approval, and manuscript-integration approval as separate decisions. Freezing never authorizes Word/LaTeX edits, caption or note changes, figure redesign, manuscript insertion, or submission export. Before integration, verify the proposed object and placement against the authoritative current plan and the current-artifact registry, obtain separate manuscript authorization, and record the exact manuscript version and location.

This research-object lifecycle registry is distinct from the multiagent manuscript artifact_contract, which protects accepted manuscript content, length, and main-text sufficiency. When both apply, maintain both records; do not use either as a substitute for the other. Follow stricter project-level repository, terminology-firewall, manuscript-conservation, or authorization rules first.

Author Intent And Semantic Fidelity

Before new claim-bearing prose, substantive rewriting, translation that may change meaning, or section restructuring, load and follow references/author-intent/01_author_intent_contract_and_semantic_fidelity.md. Do this even for a local passage; a local scope exception from paper-wide convergence does not authorize the agent to infer the author's meaning.

Inspect the manuscript, current plan, evidence, and prior author decisions before asking. If an existing frozen author-intent contract covers the exact scope, reuse it without repeated confirmation. If the author has already stated one uniquely determined meaning and requested writing from it, record that instruction as the frozen-current contract. Otherwise, ask the smallest grouped set of questions that would change the text and teach back the proposed meaning. The author's explicit confirmation normally freezes that meaning for the stated drafting scope in the same act; do not ask for a redundant second approval. Keep it unfrozen only when the author explicitly reserves approval, a material question remains unresolved, or an evidence conflict blocks the intended claim.

Treat the author as the authority over what the paper is intended to say and the inspected evidence as the authority over what the paper can support. Return clarification_required when intended meaning remains ambiguous or internally inconsistent. Return evidence_conflict when the intended claim exceeds the evidence. In either state, do not write around the problem, insert the disputed claim with a TODO, or silently substitute the agent's preferred interpretation.

Once the contract is frozen, preserve every must_express, must_not_claim, must_not_imply, required qualifier, evidence boundary, and terminology rule. The agent may improve syntax, order, transitions, and concision only within the recorded allowed_discretion. Put alternative interpretations or stronger frames in an author-facing memo and obtain approval before they enter the manuscript. If intent changes, supersede only the affected contract entry and recheck its downstream prose; never allow silent semantic drift.

Freeze the semantic content and necessary coverage of a qualifier, not a requirement to repeat the same disclaimer after every result, table, or figure. After a candidate exists, an optional derived caveat_placement_registry may map a required boundary to its first adequate location and the concrete events that permit repetition; before drafting, keep only the frozen placement policy and a pending pointer, not invented unit IDs. Precise claim type, calibrated verbs, and a nearby scope condition can satisfy the boundary without a separate defensive sentence. Before accepting even the first or only stand-alone negative caveat, prefer an integrated affirmative explanation of what the estimate, threshold, scenario, or model object is and what it shows. A sentence that introduces a new outcome or estimand only to deny that the paper estimates it adds no information unless a concrete material misreading would otherwise remain. If a negative contrast is still necessary, bind it to the exact preceding claim or number and use it as a last resort. Do not infer permission to add or retain that contrast from the model's own caution: unless the exact negative proposition is already frozen in author intent, return clarification_required and ask the author. Removing a unique no-information caveat, or a later repetition, is not semantic weakening when the surrounding calibrated wording still covers the same claims and scope.

After drafting, run the semantic-fidelity audit from the reference. Do not deliver or integrate text that omits a required proposition, adds an unapproved claim, strengthens association into causality, upgrades heterogeneity into mechanism evidence, broadens scope, drops a qualifier, or reintroduces a forbidden implication.

For a complete paper or proposal, a document-level translation or compression, or a major manuscript revision, route the accepted draft through econ-writing-workflow-multiagent and its econ-writing-workflow-multiagent/references/exhaustive_semantic_qa_protocol.md reference before delivery. This is a mandatory heavy-QA route, not an invitation to reopen the frozen author intent or shorten the manuscript. It covers every reader-visible sentence in the body and appendix, including sentences inside footnotes, captions, and table or figure notes. The multiagent controller must use native independent reviewers and deterministic coverage/hash validation; a missing reviewer or unavailable independent-agent environment returns audit_incomplete, never a simulated pass.

For a bounded local polish, audit only the changed sentences, their full paragraphs, adjacent context needed to interpret them, and any definitions or cross-section statements that depend on the change. Escalate to the heavy route when the edit changes the paper spine, claim type or strength, mechanism, definition, evidence boundary, scope, or material placement. A small context window limits reviewer input, not the length or substantive completeness of the manuscript.

Pure spelling, punctuation, citation-format, layout, or grammar-only edits do not require the full intent-elicitation protocol when the author explicitly requires meaning preservation or the requested operation has one unambiguously mechanical, meaning-preserving correction. Use the source wording and the narrow requested operation as the baseline. Verify the correction itself, including subject-verb agreement, pronoun antecedent and number, tense, negation, modifier attachment, and punctuation or citation consistency. Stop and ask only if the edit exposes a material ambiguity or requires choosing among substantively different meanings.

First Decision

Before writing or editing, classify the task stage as exploration, proposal, full drafting, document translation, compression, major revision, major restructuring, local edit, local polish, or final audit. Do not load paper-wide convergence or final-audit checks for a bounded local edit unless a substantive contradiction blocks the requested change.

When handing a task to deterministic multiagent QA, map those human-facing labels exactly: proposal -> proposal_draft, full drafting -> full_draft, document translation -> document_translation, compression -> document_compression, major revision -> major_revision, major restructuring -> major_restructure, local edit -> local_edit, local polish -> local_polish, and final audit -> final_audit. Do not invent another machine stage or use a generic final-audit label to hide the underlying artifact mode.

Task-stage classification never overrides the author-intent gate. A local substantive rewrite still requires confirmed meaning; only an explicitly or unambiguously meaning-preserving mechanical edit receives the narrow exception above.

Then classify the request:

1. English paper prose: abstract, introduction, literature review, model/theory prose, empirical/results prose, conclusion, referee response, or English revision.
2. English diction cleanup: remove AI-like prose, translationese, vague contribution language, weak verbs, or template signposting.
3. Argument logic and draft-time clarity: full-paper logic, argument spine, paragraph-level clarity, concept grounding, comparison clarity, design/results separation, mechanism-versus-heterogeneity, contribution positioning, repeated material, misplaced sections, emphasis, ordering, or whether tables and figures serve the main line.
4. Regression results and economic magnitude: main regression interpretation, coefficient size, economic magnitude, marginal effects, interaction net effects, or result prose that needs means, standard deviations, percentiles, policy benchmarks, or log-to-percent conversions.
5. Literature and judgment grounding: local reference papers, literature search/download decisions, source ledgers, citation grounding, theory/mechanism/variable alignment, contribution boundaries, or literature-based judgment calibration.
6. Full-paper drafting from a result package: draft a complete paper from a research question, regression tables, figures, result folders, slides, variable notes, or design notes.
7. Chinese top-journal writing: Chinese paper structure, Chinese abstract, introduction, contribution framing, journal fit, submission style, or top-journal revision.
8. Chinese diction cleanup: remove AI 腔, 翻译腔, 作者备忘腔, unnatural collocations, or stiff Chinese academic prose.
9. Journal submission style: target-journal formatting, submission requirements, anonymous PDF checks, accepted-version Word formatting, title hierarchy, references, JEL codes, appendices, or journal-specific body/appendix placement.
10. Tables and figures: table/figure admission, main-text versus appendix placement, three-line tables, regression/robustness/heterogeneity/mechanism tables, notes, captions, palettes, fonts, or export quality.
11. Empirical workflow: data cleaning, variable construction, sample construction, regression code, estimation, diagnostics, or reproducibility.
12. Research convergence and proposal feasibility: scope freezing, one-sentence research question, contribution hierarchy, current non-goals, model or identification closure, theory-to-measure mapping, proposal execution plans, cross-section consistency, or stopping a mature revision cycle.
13. Document transformation and exhaustive QA: a complete manuscript or proposal, document-level translation or compression, major revision, deterministic sentence/unit coverage, independent semantic review, or final acceptance gating.

If a request spans more than one category, handle them in this order:

1. empirical workflow and sample/design facts;
2. literature and judgment grounding;
3. applicable research-convergence gates;
4. table and figure design;
5. paper structure and section logic;
6. full-paper drafting plan;
7. prose and diction cleanup;
8. exhaustive semantic QA, post-diction semantic re-review, and final conservation when required;
9. final formatting/export checks.

Routing

• For new claim-bearing prose, substantive rewriting, meaning-sensitive translation, section restructuring, or uncertainty about what the author wants to say or not say, first load references/author-intent/01_author_intent_contract_and_semantic_fidelity.md.
• For a complete paper or proposal, document-level translation or compression, major revision, or exhaustive pre-delivery semantic audit, route to econ-writing-workflow-multiagent and load econ-writing-workflow-multiagent/references/exhaustive_semantic_qa_protocol.md. Do not use a sequential same-agent simulation as evidence that the independent-review gate passed.
• For English paper prose, use econ-write.
• For English diction cleanup, use econ-write and load econ-write/references/english-diction/ selectively.
• For full-paper logic, repeated material, emphasis, section ordering, and argument-spine audits, load references/argument-logic/ selectively before polishing prose. Before drafting or revising abstracts, introductions, literature positioning, design, results, mechanisms, heterogeneity, contributions, or conclusions, load references/argument-logic/06_draft_time_argument_clarity.md so clarity is handled during generation, not only in the final audit.
• For regression results, main findings, economic magnitude, coefficient interpretation, marginal effects, interaction net effects, log-to-percent conversions, or benchmark choices using means, standard deviations, percentiles, policy changes, or natural units, load references/regression-results/01_economic_magnitude_interpretation.md before drafting the relevant prose.
• For literature-dependent drafting, closest-literature positioning, literature search/download decisions, source ledgers, citation grounding, theory/mechanism/variable alignment, or calibration of similar judgments from prior papers, load references/literature-grounding/01_literature_and_judgment_grounding.md before drafting the relevant prose. For a full paper or proposal, substantive literature revision, project-specific coverage target, citation-library reconciliation, or final audit with references, also load references/literature-grounding/02_literature_coverage_and_citation_integrity.md.
• For manuscript-facing versus author-facing voice, internal memo leakage, or requests to remove author workflow notes from paper prose, load references/manuscript-voice/01_no_author_memo_in_manuscript.md before finalizing text.
• For full-paper drafting from research questions, regression tables, figures, or result folders, load references/full-paper-drafting/ first, then route table/figure decisions to and language-specific prose to or .

Active Routing During Work

Do not route only once at the beginning of a long task. At the start of each substantive phase, and whenever the work crosses from prose into tables, figures, maps, journal formatting, data facts, or empirical code, reclassify the current subtask and load the relevant specialized skill or reference file.

Route immediately when these boundaries appear:

• new claim-bearing prose, substantive rewriting, meaning-sensitive translation, section restructuring, or possible drift in claim meaning, strength, qualifiers, implications, or emphasis: load references/author-intent/01_author_intent_contract_and_semantic_fidelity.md before the language-specific writing skill;
• complete-paper or complete-proposal drafting, document-level translation or compression, major revision, or final semantic acceptance: route to econ-writing-workflow-multiagent for deterministic manifest coverage plus independent semantic review; bounded local polish remains local unless it changes a substantive dependency;
• tables, figures, maps, image edits, titles, row or column names, captions, notes, legends, colorbars, fonts, export quality, or main-text versus appendix placement: use econ-table-figure-design and its mandatory reader-facing-label reference; code-generated paper-facing artifacts and delivery candidates require its deterministic label audit as well as semantic and rendered-artifact review;
• abstracts, introductions, literature positioning, research design, results, mechanisms, heterogeneity, contributions, or conclusions where the wording depends on observable objects, comparison groups, design/results separation, mechanism evidence, contribution margins, or terminology stability: load references/argument-logic/06_draft_time_argument_clarity.md;
• main regression results, mechanism or heterogeneity estimates, coefficient size, economic magnitude, marginal effects, interaction net effects, log-to-percent conversions, mean/SD/percentile comparisons, or policy benchmark interpretation: load references/regression-results/01_economic_magnitude_interpretation.md;
• local reference papers, PDFs, .bib files, literature search/download, closest-literature positioning, data/theory/variable/mechanism alignment with prior papers, contribution boundaries, or claim-strength calibration: load references/literature-grounding/01_literature_and_judgment_grounding.md; add references/literature-grounding/02_literature_coverage_and_citation_integrity.md for paper-level coverage, registry, bibliography, or final citation-integrity work;
• full drafting, proposal development, major restructuring, scope or terminology drift, theory/model closure, theory-to-measure mapping, normative policy claims, paper-wide revision, or final-audit stopping decisions: load only the matching references/research-convergence/ file and section;
• target-journal style, anonymous submission, title hierarchy, references, JEL codes, appendices, Word/PDF submission format, or 《经济学（季刊）》 rules: use cn-top-econ-writing and its cn-top-econ-writing/references/journal-styles/ module;

After a specialized skill settles its part, return to this workflow to integrate the decision into the paper-level order and final output check.

Operating Rules

• Apply research-convergence rules by task stage, paper type, requested scope, and claim type. Do not turn a local prose edit into an unsolicited paper-wide audit, and do not force an exploration-stage idea into premature closure.
• Keep one central question visible while preserving legitimate secondary theory, data, measurement, structural, or policy contributions that materially change the intellectual claim. Treat fixed paragraph counts, result counts, and minimal-model advice as defaults with paper-type exceptions, not universal vetoes.
• Distinguish model results, numerical exercises, causal estimates, descriptive facts, suggestive interpretations, and planned tests. The abstract, introduction, conclusion, and policy discussion must not state a stronger or more complete result than the body supports.
• Keep scope contracts, non-goal lists, consistency matrices, revision-pass labels, and stopping decisions in author-facing logs or memos. Never insert them into manuscript prose, table notes, figure notes, appendix notes, or footnotes.
• Read available manuscript context before giving paper-level prose advice. If only an excerpt is available, state that the advice is excerpt-level.
• Before drafting each paper-facing paragraph, apply the author-intent gate and then the draft-time clarity gate. Identify the intended proposition, object, observable or model anchor, claim type, comparison, qualifiers, forbidden implications, and section function. If an unknown item would change meaning, return clarification_required; use a TODO only for a missing non-semantic fact inside an already confirmed proposition.
• Do not rewrite tables or figures as prose problems. Table selection, sample comparability, notes, and visual design are part of the research presentation.
• Do not put author workflow notes in paper-facing text, table notes, or figure notes.
• Before delivering or inserting manuscript text, classify each sentence as paper-facing text, author memo, or task log. Only paper-facing text may enter the manuscript.
• If the manuscript context is unclear on research question, variables, sample, identification, table/figure meaning, target journal, or whether the author wants to keep, delete, strengthen, weaken, or reframe a substantive claim, ask the user and stop the affected drafting scope; do not guess or use a TODO as semantic authorization.
• When discussing regression results in the manuscript, translate the coefficient into at least one suitable economic magnitude benchmark. Do not invent means, standard deviations, percentiles, policy changes, or marginal effects; compute them from supplied materials or ask for them. If the missing benchmark is recorded as an author-facing TODO, keep the unsupported magnitude out of paper-facing prose.
• Before writing literature-dependent claims, check local literature materials and ask whether they are sufficient when that is unclear. If the user wants help finding literature, only rely on papers that can be inspected or that the user supplies. If important Chinese or paywalled papers are unavailable, say so; use an author-facing only for non-semantic citation or verification details, and return with affected prose withheld when the missing support changes the claim.

Argument-Logic References

Load only the relevant files:

• references/argument-logic/01_argument_spine.md: one-chain paper logic and missing-link audit.
• references/argument-logic/02_redundancy_linter.md: repeated, appendix, and delete/merge material.
• references/argument-logic/03_emphasis_and_ordering.md: what to foreground, delay, or de-emphasize.
• references/argument-logic/04_table_figure_logic.md: whether tables and figures serve the main line.
• references/argument-logic/05_revision_workflow.md: full-paper logic revision workflow.
• references/argument-logic/06_draft_time_argument_clarity.md: draft-time concept grounding, comparison clarity, design/results separation, mechanism-versus-heterogeneity, literature-margin positioning, and terminology stability.

Research-Convergence References

Load only the matching module and relevant section:

• references/research-convergence/01_scope_and_claim_contract.md: task-stage classification, scope contract, paper-type argument spine, must-preserve/current-non-goal lists, terminology ledger, and reopening conditions.
• references/research-convergence/02_theory_empirical_closure_gates.md: evidence-status calibration, model closure, friction and assumption roles, economic-result qualification, concept-to-measure mapping, identification capacity, and normative policy conditions.
• references/research-convergence/03_proposal_revision_stop_gates.md: proposal status and executable plans, ordered major-revision passes, author-facing consistency matrix, and stop/reopen criteria.

Author-Intent Reference

Load before claim-bearing drafting or substantive rewriting:

• references/author-intent/01_author_intent_contract_and_semantic_fidelity.md: defines the author-intent contract, teach-back confirmation, narrow mechanical exception, frozen-current meaning record, evidence-conflict handling, change control, and post-draft semantic-fidelity audit.

Artifact-Lifecycle Reference

Load for multi-step projects that create, approve, freeze, replace, reject, or insert tables, figures, maps, measures, specifications, model outputs, or other research objects:

• references/artifact-lifecycle/01_artifact_state_and_freeze_protocol.md: separates artifact lifecycle from manuscript integration; defines allowed transitions, mandatory registry and SHA-256 fields, immutable superseded audit bundles, and safe current-path replacement.

Full-Paper Drafting References

Load only the relevant files:

• references/full-paper-drafting/01_input_checklist.md: input audit and draftability levels.
• references/full-paper-drafting/02_from_results_to_outline.md: result ledger, argument spine, and table/figure placement logic.
• references/full-paper-drafting/03_section_drafting_order.md: section order and language-specific routing.
• references/full-paper-drafting/04_missing_information_policy.md: anti-fabrication rules and TODO format.
• references/full-paper-drafting/05_full_draft_workflow.md: full result-package-to-paper workflow.

Literature-Grounding References

Load before drafting or revising literature-dependent sections or judgments:

• references/literature-grounding/01_literature_and_judgment_grounding.md: checks local references, asks whether literature is sufficient, builds source and judgment ledgers, aligns reused data/theory/variables/mechanisms with sources, and audits claim strength.
• references/literature-grounding/02_literature_coverage_and_citation_integrity.md: defines literature-coverage-contract/1.0, reference-library-manifest/1.0, literature-registry/1.0, their relationship to the existing text_to_evidence_ledger, project-specific coverage targets, and final manuscript-to-bibliography closure.

Manuscript-Voice References

Load when drafting, revising, polishing, or auditing prose that will enter the manuscript:

• references/manuscript-voice/01_no_author_memo_in_manuscript.md: separates manuscript text from author memos and task logs; includes forbidden and context-dependent author-memo patterns.

Regression-Results References

Load before drafting or revising main result prose, abstracts, introductions, conclusions, or referee responses that discuss coefficient size:

• references/regression-results/01_economic_magnitude_interpretation.md: chooses natural units, policy changes, means, standard deviations, percentile spreads, marginal effects, interaction net effects, and log-to-percent conversions for reader-facing economic magnitude.

Output Check

Before finalizing a response, confirm that:

• the routed skill still matches the current subtask, not just the original request;
• any claim-bearing draft is covered by a frozen-current author-intent contract for the requested scope, or qualifies for the explicit meaning-preserving mechanical exception;
• the semantic-fidelity audit maps every required intent to the draft and finds no unapproved addition, claim-strength escalation, lost qualifier, forbidden implication, or evidence conflict;
• any complete paper or proposal, document-level translation or compression, or major revision has a current exhaustive-QA report tied to the candidate and contract hashes; all reader-visible text is captured, every sentence/heading review target has the required roles, high-risk targets meet the contract's explicit independent-review minimum, and the final status is pass rather than clarification_required, evidence_conflict, audit_incomplete, approval_required, fail, or metric_unavailable;
• any bounded local polish rechecks the changed sentences, their paragraph and necessary adjacent context, and affected definition or cross-section dependencies rather than claiming a paper-wide pass;
• any table/figure decision has main-text versus appendix placement and has passed the applicable basic-identity, full-recoverability, language, and rendered-label checks;
• any empirical claim is grounded in provided or inspected evidence;
• any main regression or important mechanism/robustness result has an economic magnitude benchmark, or the missing benchmark is kept out of paper-facing prose and recorded as a concrete author-facing TODO;
• any literature-dependent claim, theory, mechanism, variable, data source, empirical choice, contribution boundary, or policy implication is grounded in inspected local sources; unavailable non-semantic metadata is recorded as an author-facing TODO, while a material support gap returns evidence_conflict and keeps the affected prose unwritten;
• any applicable literature-coverage contract has no unresolved functional cluster, every manuscript citekey resolves through the authoritative library and admitted registry, every cited claim is linked to its evidence where required, and the final visible bibliography is derived from manuscript citekeys plus explicitly authorized nocite entries rather than from all unused library inventory;
• every necessary qualifier retains its meaning at an adequate reader-visible location, while a first, unique, or repeated defensive caveat with no new scope or evidence information is rewritten as affirmative interpretation or removed; absence of a generic disclaimer is not a failure when calibrated wording already preserves the evidence boundary, and any retained negative sentence is bound to a concrete claim and material misreading;