| name | agu-writing-workflow |
| description | Use when the user wants Codex to diagnose, plan, revise, or build prompts for AGU-style scientific manuscripts, with journal-specific adaptation for GRL, JGR, WRR, JGR: Solid Earth, and related geoscience journals. Especially useful for mechanism-driven geoscience, reactive transport, porous media, hydrogeophysics, or NMR papers. Applies teacher-derived OCAR/funnel structure, storyline alignment, narrative echo and payoff, methods evidence-chain, mechanism-centered results, paragraph engineering, sentence hierarchy, revision-case mining, and prompt-building rules. Trigger on requests about 前后呼应, 铺垫与兑现, Introduction-Results alignment, Results subsection transitions, or mechanism-driven AGU writing. Always diagnose manuscript stage and target journal profile before polishing sentences. |
AGU Writing Workflow
Core Rule
This skill is a scientific writing operating system, not a general polishing prompt.
Always follow this order:
storyline > structure > narrative echo > mechanism > paragraph > sentence > word choice
Do not perform sentence polishing when the central story, section function, or mechanism chain is still unstable. Start with diagnosis, then choose the smallest matching writing operation.
This skill also checks narrative echo and payoff: whether Introduction promises are paid off in Results/Discussion, and whether Results subsections form a connected mechanism chain through forward pointers, backward links, delayed payoffs, and thread summaries.
Before applying concision, SI-moving, or reviewer-fit rules, identify the target journal profile. For GRL, prioritize concise main-text storytelling around one high-impact mechanism-centered claim. For JGR, WRR, JGR: Solid Earth, and other full-length AGU Research Articles, preserve sufficient methodological detail, literature context, validation, uncertainty analysis, and mechanism discussion in the main text.
First Move
Unless the user explicitly asks for one narrow operation, begin with references/workflow_router.md and, when prompt wording is needed, references/workflow_router_prompt.md.
Diagnose:
- Target journal profile: GRL, JGR, WRR, JGR: Solid Earth, generic AGU Research Article, or unspecified.
- Current manuscript stage.
- Core contradiction: storyline, section structure, mechanism chain, paragraph flow, sentence hierarchy, or reviewer risk.
- Which sub-skill files to use.
- Which edits are currently forbidden.
- One concrete goal for this round.
Use references/journal_profile.md before enforcing concision or SI-moving rules.
Use references/stage_model.md for the five-stage workflow.
Stage Routing
Stage 1 Research Story Discovery:
Use references/story_architect.md and references/ocar_funnel.md. Output a one-sentence storyline, novelty statement, main figure sequence, supporting actor list, and mechanism sketch. Do not polish sentences.
Stage 2 Manuscript Architecture Design:
Use references/story_architect.md, references/ocar_funnel.md, references/narrative_echo_and_payoff.md, and references/skill_hierarchy_map.md. Output section functions, subsection sequence, figure sequence, OCAR roles, and Introduction-Results promise-payoff checks. Do not do grammar-level revision.
Stage 3 Mechanism Construction:
Use references/narrative_echo_and_payoff.md, references/mechanism_results.md, references/cases_mechanism_centered_results.md, references/methods_evidence_chain.md, references/paragraph_engineering.md, and, when revising paragraph function or order, references/cases_paragraph_engineering.md. Turn results from isolated blocks into promise-payoff mechanism progression. Keep secondary metrics as support, not the protagonist.
Stage 4 Reader Experience Optimization:
Use references/paragraph_engineering.md, references/cases_paragraph_engineering.md, and references/sentence_hierarchy.md. Improve topic sentences, transitions, information load, reader guidance, and claim calibration after the structure is stable.
Stage 5 Reviewer-Oriented Refinement:
Use references/principles.md, references/tag_system.md, references/narrative_echo_and_payoff.md, references/journal_profile.md, and references/global_prompt_checklist.md. Check novelty clarity, target-journal fit, overclaim, evidence gaps, method reproducibility, broad implication, Introduction-Results payoff, and likely reviewer attack points.
Reference Loading
Load only the files needed for the current task.
Principles and evidence:
references/stage_model.md: five-stage manuscript workflow and forbidden actions.references/journal_profile.md: AGU journal-specific adaptation rules for GRL, JGR, WRR, JGR: Solid Earth, and generic full-length Research Articles.references/principles.md: teacher-derived principle library, evidence strength, and transferable rules.references/tag_frequency.md: which principles are most frequent in the modification database.references/tag_system.md: modification label system.references/raw_modification_log.md: original revision pairs in Markdown form.references/interpretation_principle_log.md: interpreted revision principles.references/modification_summary.md: compact summary of mined changes.
Sub-skill designs:
references/workflow_router.md: stage diagnosis and routing.references/story_architect.md: storyline, novelty, protagonist, supporting actors.references/cases_story_architect.md: reference-paper examples for Core Story, Central Scientific Question, main character, supporting actors, section storyline, and figure sequence logic.references/ocar_funnel.md: Opening, Challenge, Action, Resolution at manuscript, section, and paragraph levels.references/cases_ocar_funnel.md: reference-paper and manuscript examples for Abstract, Introduction, Methods opening, and Results opening OCAR funnels.references/narrative_echo_and_payoff.md: paper-level and Results-level promise-payoff relationships, including Introduction-Results alignment and subsection-to-subsection echo within Results and Discussion.references/cases_narrative_echo_and_payoff.md: manuscript-derived examples for Introduction payoff, forward pointers, backward links, and thread summaries.references/methods_evidence_chain.md: methods as evidence logic, not technical display.references/mechanism_results.md: results as physical mechanism progression.references/cases_mechanism_centered_results.md: reference-paper and manuscript examples for observation-to-mechanism, structure-to-signal, metric validation, regime comparison, breakthrough mechanism, and mini-synthesis Results writing.references/paragraph_engineering.md: topic sentence, evidence order, mini-implication.references/cases_paragraph_engineering.md: reference-paper and manuscript examples for PE-TOPIC, PE-ORDER, PE-MECH, PE-ECHO, PE-END, and PE-COMPRESS paragraph repair cases.references/sentence_hierarchy.md: sentence function, transitions, claim strength, information load.references/skill_hierarchy_map.md: how the sub-skills relate.
Prompt and test materials:
references/prompt_system_overview.md: prompt system architecture.references/workflow_router_prompt.md: router prompt template.references/story_architect_prompt.md: storyline prompt template.references/ocar_funnel_prompt.md: OCAR/funnel prompt template.references/narrative_echo_and_payoff_prompt.md: promise-payoff and Results-continuity diagnosis prompt template.references/methods_evidence_chain_prompt.md: methods evidence-chain prompt template.references/mechanism_results_prompt.md: mechanism-centered results prompt template.references/paragraph_engineering_prompt.md: paragraph organization prompt template.references/sentence_hierarchy_prompt.md: sentence hierarchy prompt template.references/global_prompt_checklist.md: global quality checks.references/skill_specific_checklists.md: sub-skill-specific checks.references/small_scale_testing_plan.md: testing protocol for prompt behavior.
Output Contract
For diagnosis tasks, output:
Current stage:
Target journal profile:
Core contradiction:
Recommended sub-skills:
Currently forbidden:
This round's goal:
Next action:
For revision tasks, keep the output tied to the diagnosed stage:
- Storyline work: provide storyline, novelty, protagonist, supporting actors, and mechanism chain.
- OCAR work: label only Opening, Challenge, Action, and Resolution.
- Methods work: map research question to method choice, measurement, evidence, and credibility.
- Results work: rewrite around observation -> physics -> signal -> mechanism -> implication.
- Paragraph work: state paragraph function, reorder sentences, then revise.
- Sentence work: polish only after confirming the higher-level logic is stable.
When using revision evidence, cite the relevant Case ID or tag from the Markdown modification logs if available.
Guardrails
- Prefer mechanism-first writing over workload listing.
- Prefer reader guidance over dense technical accumulation.
- Prefer target-journal-calibrated claims over broad unsupported implications.
- Do not turn full-length AGU Research Articles into GRL-style compressed short papers unless the user explicitly targets GRL.
- Treat methods as proof logic: why this workflow can answer the research question.
- Treat figures as evidence in a story, not as a sequence to describe one by one.
- If the user asks for pure polishing but the text has a higher-level failure, say so briefly and route upward before editing.
