Econ Writing Workflow Multiagent

Economics Writing Workflow Multiagent

Role

This is an experimental coordination skill for large economics writing projects. It does not replace or rewrite the stable econ-writing-workflow; it sits beside it as a beta workflow.

Use this skill only as the controller for multi-agent or staged-agent work. It should coordinate specialized skills, maintain a shared paper state, manage context budget, and integrate outputs. It should not duplicate the substantive rules already maintained in:

• econ-write
• cn-top-econ-writing
• econ-table-figure-design
• empirical-econ-workflow

Use native sub-agents for independent semantic acceptance review and explicit handoff templates for every delegated task. Staged same-agent roles may support planning or drafting, but they do not satisfy the independent-review gate. If required native reviewers are unavailable, return audit_incomplete; never simulate independent reviewers sequentially and report a pass.

When To Use

Use this beta workflow for:

• full paper drafting from a research question, result package, tables, figures, variable notes, and design notes;
• complete research proposals, document-level translations or compressions, and their final semantic acceptance audits;
• long revision projects with many sections, tables, figures, appendix items, or referee comments;
• projects where English and Chinese writing modules both matter;
• tasks where a single-agent context is likely to lose track of contributions, variables, magnitudes, caveats, or table/figure placement;
• comparisons between the stable workflow and a multi-agent-aware workflow.

Do not use it for:

• one paragraph of polishing;
• a single abstract rewrite;
• one table note or one figure caption;
• a short, unambiguously meaning-preserving translation;
• pure data cleaning, regressions, estimation, or code execution without a writing integration task.

For small tasks, use the stable econ-writing-workflow or the relevant child skill directly.

Core Rule

Before splitting work across agents or roles, create or update a shared paper state. The paper state is the single source of truth. No agent may invent or silently alter:

• research question;
• contributions;
• data, sample, period, or variable definitions;
• identification or model assumptions;
• coefficient values, magnitudes, mechanisms, robustness results, or caveats;
• literature claims or citations.

Evidence-boundary accuracy is not a license to write defensive disclaimers. Default to calibrated verbs, the correct claim type, and only the scope qualifiers needed for the sentence. Before admitting 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. Retain a separate negative sentence only when it blocks a concrete material misreading not already excluded by that wording; bind it to the exact claim or number and use it as a last resort. An agent may not self-authorize that last resort: unless the exact negative proposition is already frozen in the author-intent contract, return needs_author and stop the affected wording for author adjudication. Repeat it only when the method, sample, period, geography, extrapolation target, or evidence level changes, or when a standalone abstract, caption, or note must remain intelligible, or a journal/referee explicitly requires it. Do not impose a per-section count. Removing a first, unique, or repeated no-information caveat is not loss of evidence discipline, but removing a concrete identification assumption or changing the effective claim remains a blocking semantic change.

Missing non-semantic information may remain as a concrete TODO. Missing meaning or evidence needed to support a claim returns clarification_required or evidence_conflict and leaves the affected prose unwritten.

Multiagent state must point to the stable workflow's one authoritative author_intent_contract; it must not copy it into a parallel authority. Only a complete frozen-current + ready contract authorizes claim-bearing drafting. Explicit author confirmation of the complete teach-back normally freezes the contract in the same event. Partial, proposed, unconfirmed, held, or materially unresolved intent never authorizes drafting, and there is no legacy compatibility path around this gate.

Context compression governs handoffs and duplicated input, not the reader-facing manuscript. Preserve the accepted artifact budget and substantive main-text coverage unless the user, project rules, or verified format requirements authorize a shorter output.

For an accepted mature manuscript, default to patch_existing or reorder_existing_blocks. Do not use clean-slate section redrafting unless the author explicitly authorizes full_redraft for the recorded scope. A controller's judgment that the existing section is weak or unusable is a reason to request that authorization, not a substitute for it.

Routing

Load the relevant reference file before starting each phase:

• Before any multi-agent task, load references/controller_startup_checklist.md.
• Before drafting or semantic acceptance, load references/exhaustive_semantic_qa_protocol.md; also load the stable workflow's authoritative author-intent reference rather than inventing a multiagent intent schema.
• For full papers/proposals, major revisions or restructures, substantive literature changes, and final audits with references, load the stable workflow's econ-writing-workflow/references/literature-grounding/02_literature_coverage_and_citation_integrity.md.
• For shared facts and the paper_state schema, load references/paper_state_protocol.md.
• For full drafting, mature-manuscript revision, restructuring, shortening, or appendix relocation, load references/artifact_conservation_and_depth_gates.md before setting the paper spine or assigning roles.
• For role definitions and when to use true sub-agents versus staged roles, load references/agent_roles.md.
• For structured delegation and return formats, load references/handoff_templates.md.
• For deciding whether to split work, load references/context_budget_rules.md.
• For merging outputs and resolving disagreements, load references/integration_and_conflict_resolution.md.
• For controller-led section agents, section cards, and section-level integration, load references/section_agent_protocol.md.
• For controller-mediated collaboration between section agents and functional agents, load references/cross_agent_collaboration_protocol.md.
• For complete result-package-to-paper workflows, load references/full_paper_multiagent_workflow.md.

Then route substantive work to child skills:

• English writing and revision: econ-write.
• English diction cleanup: econ-write with econ-write/references/english-diction/.
• Chinese top-journal writing: cn-top-econ-writing.
• Chinese diction cleanup: cn-top-econ-writing with cn-top-econ-writing/references/chinese-diction/.
• Chinese argument logic: cn-top-econ-writing with cn-top-econ-writing/references/argument-logic/.
• Tables, figures, notes, captions, titles, row and column names, axes, legends, palettes, and main-text versus appendix placement: econ-table-figure-design. Apply its mandatory reader-facing-label reference and two-level cold-reader/full-recoverability review; do not create a parallel multiagent label schema.
• Data cleaning, variable construction, regression estimation, and reproducibility: empirical-econ-workflow.

Default Workflow

For complex paper tasks, proceed in this order:

1. Run the controller startup checklist.
2. Classify the request and decide whether multi-agent coordination is justified.
3. Locate the stable workflow's authoritative author-intent contract. Clarify, teach back, and obtain confirmation when needed; proceed only when its exact scope is frozen-current + ready.
4. Build or update paper_state, content obligations, definition registry, optional frozen caveat-placement policy and pending_candidate registry pointer, four-way ledgers, literature authority pointers and status when triggered, and, when applicable, the artifact contract and baseline measurements. Materialize the derived caveat registry only after candidate unit IDs exist.
5. Audit inputs and keep non-semantic missing facts as TODO; stop semantic or evidence gaps with clarification_required or evidence_conflict.
6. Settle the paper spine and create section cards with intent IDs, rewrite modes, and artifact budgets when section-level work is needed.
7. Assign narrow drafting or functional roles using the handoff template, then integrate outputs against the frozen state and ledgers rather than memory.
8. For every complete paper/proposal, document-level translation/compression, or major revision, generate the deterministic QA manifest and run the four isolated semantic-review roles from the exhaustive-QA protocol.
9. When literature audit is triggered, run deterministic citation integrity and the independent Literature Coverage and Citation Integrity Role outside the core four-role plus conservation-role assignment schema, using its separate hash-bound literature-audit-assignment/1.0 record. Accept neither artifact by inspection alone: run scripts/validate_literature_audit.py over the live authorities, deterministic report, assignment, role result, and applicable core assignment registry.
10. Resolve conflicts by authority rather than majority vote; apply only bounded minimum patches, then re-review changed and dependent units.
11. Run diction only after semantic acceptance, then run post-diction semantic re-review.
12. Run final deterministic citation closure when references are present, revalidate the final-hash literature assignment/result, then run conservation and independent main-text sufficiency checks, followed by final consistency.
13. Deliver only when every applicable gate is pass; otherwise return the exact blocking status, unresolved items, and required user decision.

Output Check

Before finalizing, confirm:

• the stable workflow's author-intent authority was reused rather than copied into a competing multiagent contract;
• claim-bearing drafting began only after the exact requested scope was frozen-current + ready;
• the controller startup checklist was run before delegation;
• every delegated or staged pass used paper_state as the authority;
• every section agent worked from a controller-approved section card when section-level drafting was used;
• every section-functional collaboration went through the controller and updated the section card when needed;
• disagreements were resolved explicitly by authority; unresolved semantic, evidence, definition, or scope conflicts retain a blocking status, while only missing non-semantic facts may remain as user-facing TODO;
• table/figure decisions were integrated into the argument spine;
• every code-generated paper-facing or final table/figure has a current hash-bound label audit, and the Table/Figure Role completed the basic-identity, full-recoverability, language, and rendered-artifact passes required by econ-table-figure-design;
• every triggered literature role has a current gate-local literature-audit-validation/1.0 pass from scripts/validate_literature_audit.py; a plausible role memo or a deterministic citation pass alone was not treated as acceptance, and this local pass was not presented as whole-manuscript delivery authorization;
• prose edits did not delete central contributions, mechanisms, magnitudes, concrete necessary evidence boundaries, or design features; a first, unique, or repeated no-information defensive sentence was rewritten as affirmative interpretation or removed without being treated as a semantic loss, while every retained negative sentence remained bound to a concrete claim and material misreading and had exact frozen author-intent authority; an otherwise recommended keep ended in needs_author, not an agent-generated pass;
• any mature-draft revision followed its recorded rewrite mode and mapped every substantive source block to a conservation-ledger disposition;
• context or handoff limits were not used as manuscript-length instructions;
• the QA manifest captures every reader-visible body, appendix, footnote, caption, and note object; every sentence/heading review target is covered in exhaustive mode, or the declared changed and dependency set is covered in bounded-change mode;
• every required native semantic role returned a valid current-hash result, every high-risk unit met the contract's explicit independent-review minimum, and no uncertain, missing, stale, invalid, or unresolved finding was treated as a voteable pass;
• reviewers returned structured findings without editing the manuscript, and controller patches plus diction changes received the required semantic re-review;
• every triggered literature audit has a current separate assignment and is current-hash bound; applicable coverage clusters, admitted citekeys, the text-to-evidence ledger, and the final visible bibliography close under the literature contract;
• the integrated manuscript satisfies the artifact contract, cumulative compression permission, and main-text self-containment gate;
• deterministic audit status is pass; if it first returned approval_required, any scoped approval was recorded and the audit was rerun to pass;
• no fabricated data, citations, results, or policy implications entered the output.