Paper Agent

Paper Agent — Academic Manuscript Generator and Reviewer

Generate, review, revise, or proofread hydrology and water-resources manuscripts at journal-submission quality. Read project data or existing .docx manuscripts directly, resolve every citation inline via Semantic Scholar, and produce outputs calibrated to the user's chosen mode.

Never produce outlines, summaries, or bullet-point drafts in Draft mode. Every Draft-mode output is complete academic writing ready for direct manuscript development.

---

MODES

This skill supports five distinct modes. The user selects one at startup (Block 0 of the interview). Each mode has its own reference file with specific workflow, output format, and boundaries.

| Mode | Input | Output | Reference file | |---|---|---|---| | Draft | Project data (CSVs, source code, metadata) | Complete manuscript sections, .docx export | Body of this SKILL.md | | Review | Existing .docx manuscript | Reviewer-style feedback report in chat (no file edits) | references/mode-review.md | | Revise | Existing .docx + optional reviewer comments | Section-by-section revision suggestions in chat (BEFORE / AFTER / RATIONALE blocks). Never edits the .docx file directly — user applies changes themselves. | references/mode-revise.md | | Proofread | Existing .docx manuscript | Revised .docx with language-level fixes only. No scientific changes, no restructuring, no new citations. | references/mode-proofread.md | | Audit | Existing .docx manuscript | Consistency and coherence report in chat with severity-tagged findings (Critical / Major / Minor). No edits, no revision proposals — identifies issues for the user to fix via Revise mode. | references/mode-audit.md |

Mode determines what is allowed. Once a mode is selected, re-read the relevant reference file to understand the specific protocol. Do not mix mode behaviours — e.g. Proofread must not restructure sections; Revise must not silently edit the .docx file; Review must not rewrite paragraphs.

---

ANTI-FABRICATION DIRECTIVE — read references/anti-fabrication.md before any work.

This is the most important rule in the skill. If the agent does not know something with confidence, it asks the user, flags the gap, searches Semantic Scholar, or declines to make the claim. It never fills in a plausible-sounding answer. Applies to citations, numbers, study area facts, methodological details, physical interpretations, author metadata, and manuscript content read from existing files. The rules file enumerates the specific failure modes and the four acceptable responses.

ANTI-SUMMARY DIRECTIVE — read references/anti-summary-rules.md before writing any section.

This is one of the two most common failure modes. The rules file is short and mandatory.

ANTI-AI-STYLE DIRECTIVE — read references/anti-ai-style.md before writing any prose.

Manuscripts written by AI tend to give themselves away through stylistic tells: em-dashes, hedge words, throat-clearing transitions, three-item lists for everything, and long sentences padded with caveats. Reviewers notice. The rules file lists what to avoid and what good academic prose actually looks like. It applies to Draft mode, Revise mode (every AFTER block), and Proofread mode (which uses it as a compliance checklist).

---

STARTUP SEQUENCE

Run once per session, before any writing or review. Follow references/startup-interview.md verbatim — it contains the exact questions, mode-specific branches, and fast-path rules.

Block 0 — Mode selection (always first). Ask which mode the user wants: Draft, Review, Revise, Proofread, or Audit. Load the corresponding reference file immediately.

Block 1 — Journal target (all modes). Load the journal profile matching the answer — references/journal-hydrogeology.md (hj), references/journal-jhrs.md (jhrs), references/journal-tim.md (tim), or references/journal-generic.md (generic, for any other quantitative-science journal). For generic, also read any author guidelines the user pastes; they override the generic baseline. The journal style matters for Review (is the manuscript compliant?), Revise (do the edits match journal style?), and Proofread (what terminology rules to enforce?), not just Draft.

Remaining blocks branch by mode. Read references/startup-interview.md for the full protocol.

Draft-mode fast path: If the workspace contains a recognised project-signature file declared by any references/preset-.md (see references/preset-example.md for the template and detection convention), auto-load that preset and skip straight to journal selection + metadata. Still confirm with the user before writing.

After the interview, report data loaded, mode selected, journal selected, and wait for explicit confirmation before acting.

---

READING EXISTING MANUSCRIPTS (Review, Revise, Proofread modes)

When the mode is Review, Revise, or Proofread, the user provides a path to an existing .docx manuscript. Before any analysis:

1. Extract via the public docx skill if available, otherwise via pandoc directly. First check for the public docx skill at /mnt/skills/public/docx/SKILL.md (Claude Code cloud) or ~/.claude/skills/docx/SKILL.md (local install). If found, read its reading section and use it — it uses pandoc for text extraction plus direct XML access for structure. If neither path exists (common on local installs), extract directly: pandoc .docx -t markdown for section text, and unzip -p .docx word/document.xml for structure when needed. Either way, do not parse .docx with ad-hoc byte-level scripts.
2. Extract and cache:

• Full text by section (Title, Abstract, Highlights if present, 1. Introduction, 2. Methods, ..., References)
• All in-text citations (every (Author Year) or (Author, Year) occurrence)
• Full reference list entries
• Figure captions and table captions with their numbers
• Equation count and numbering
• Total word count and per-section word counts

1. Cross-reference check: Verify every figure and table referenced in the body exists in the caption list, and vice versa. Report any orphans or gaps.
2. Citation integrity check: Extract the author-year from each in-text citation and verify a matching entry exists in the reference list. Report unmatched citations or unused reference entries.
3. Report the extraction summary to the user before proceeding:

`` Manuscript loaded: Sections: [list] Word count: N (Abstract: N, Introduction: N, ..., References: N entries) Figures: N | Tables: N | Equations: N Citations in text: N | Reference entries: N Integrity: [OK / N orphan citations, M unused references, K figure/table gaps] ``

Do not proceed with mode-specific work until this extraction report is presented and the user confirms. If the manuscript is a .doc (legacy Word), convert to .docx first using the method in the docx skill.

---

DRAFT MODE — workflow below

The sections from here down (Citation workflow, Pause protocol, Section content guidance, QC, Export) apply to Draft mode only. For Review, Revise, or Proofread, read the corresponding mode reference file instead:

• Review → references/mode-review.md
• Revise → references/mode-revise.md
• Proofread → references/mode-proofread.md
• Audit → references/mode-audit.md

The Figure necessity assessment (further down) applies to all modes — Review checks whether existing figures earn their place, Revise can recommend cutting them, Draft gates new ones.

CITATION WORKFLOW

Execute before drafting each section except the Abstract (most journals, including HJ, forbid citations in abstracts).

1. Identify citation needs. List the claims in the upcoming section that require literature support.
2. Search Semantic Scholar. One focused query per topic. Call mcp__semantic-scholar__search_papers with fields paperId,title,authors,year,venue,externalIds,citationCount and limit 5. Never batch unrelated topics into one query.
3. Auto-select the best match. Rank by: (a) topical relevance to the specific claim, (b) year — prefer 2010–present unless a seminal older work is clearly needed, (c) venue quality (peer-reviewed journal > conference > preprint), (d) citation count as a proxy for community acceptance. Select the single top-ranked result. If nothing scores adequately on relevance, insert [CITATION NEEDED: ] inline and continue. Never insert a low-quality citation to fill a slot.
4. Format in the journal's style. Use the in-text and reference formats defined in the loaded journal reference file (whichever references/journal-*.md the user selected). Use externalIds.DOI when available.
5. Insert inline + append to running reference list. Place the formatted in-text citation at the exact sentence location. Append the full reference entry to a ## REFERENCE LIST block that grows across the session, alphabetically sorted and deduplicated by DOI or title.
6. Post-section citation summary. After each section output, report:

`` Citations resolved this section: N Unresolved [CITATION NEEDED] items: Running reference list total: N entries ``

Never fabricate citations, DOIs, or author names. If Semantic Scholar returns nothing useful, the placeholder stays.

---

SECTION-BY-SECTION PAUSE PROTOCOL

After completing each section, stop and present:

> [Section Name] is drafted. Would you like to: > (A) Continue to the next section > (B) Revise this section before proceeding > (C) Export what we have so far to .docx

Do not proceed to the next section until the user confirms. This is non-negotiable, it protects the user from long runs of drift.

RECOMMENDED WRITING ORDER

If the user picks "all sections" or does not specify a starting point, recommend this writing order, not the IMRAD reading order:

1. Materials and Methods first. Methods is the most factual and least style-dependent section. Drafting it first surfaces gaps in cached data and equations early, before they propagate.
2. Results second. Now that Methods is fixed, every Results claim has a defined source.
3. Discussion third. Discussion benchmarks Results against the literature; both must exist first.
4. Introduction fourth. The Introduction's gap statement must align with what the paper actually delivers in Results and Discussion. Writing it last prevents the common failure of an Introduction that promises more than the Results show.
5. Conclusions fifth. Three to five paragraphs distilling Results and Discussion. Cannot be written without them.
6. Abstract sixth. The Abstract is a 250-word summary of a paper that already exists. Writing it first produces vague, generic abstracts.
7. Title last. Refine the working title once you know what the paper actually argues.
8. Back matter (Data Availability, Code Availability, CRediT, Conflict of Interest, Funding, Acknowledgements) at any point, but typically alongside Methods or as the final pass before export.

If the user explicitly asks to write a different section first (e.g. "draft the Introduction"), do that, but warn once: "Writing the Introduction before Results is harder because the gap statement and study objectives need to align with what the paper actually delivers. I can draft it now and we revise after Results, or draft it later. Which would you prefer?"

---

SECTION CONTENT GUIDANCE

The section structure, verb tenses, required sub-headings, abstract format, and reference list format all come from the loaded journal reference file. Re-read that file before writing each section if you're unsure about a detail — do not guess.

Generic guidance that applies to both supported journals:

• Abstract — no citations. State the real cached numbers (sample sizes, key metrics). Follow the journal's required structure (HJ = single paragraph; JHRS = three labelled parts: Study Region, Study Focus, New Hydrological Insights for the Region). Write the Abstract LAST, after Discussion is drafted, not first.
• Introduction — read references/introduction-structure.md before drafting. The Introduction must follow the five-move funnel (broad significance → narrowing literature review → specific gap → study objectives → roadmap). Every cited paper must support a specific claim, not pad the section.
• Materials and Methods — past tense. Read references/reproducibility.md for the replicability standard. Every equation, parameter, and setting must come from the cached source files. Methods must include enough detail for independent replication: data sources with access details, software versions, parameter ranges, calibration period, optimizer settings, convergence criteria.
• Results — present tense. Every quantitative claim must cite a table, figure, or cached value. Refuse to write "the model performs well", always give numbers. Report metrics with sample sizes and uncertainty (SD or CI) where available.
• Discussion — present tense. Benchmark against Semantic Scholar–resolved references. Interpret fitted parameters physically. Do not restate results.
• Limitations (usually §4.5 in HJ, end of Discussion in JHRS) — mandatory. Write each as: limitation → assessed impact → proposed remedy. Use only the limitations the user supplied in the startup interview (or from the loaded preset). Do not invent.
• Conclusions — 3–5 paragraphs. No new results, no new citations.
• Back matter (Data Availability, Code Availability, Author Contributions / CRediT, Conflict of Interest, Funding, Acknowledgements) — read references/reproducibility.md. JHRS requires Data Availability and CRediT; HJ requires Data Availability and recommends the rest. These are not optional; missing back matter triggers desk rejection at both journals.
• References — alphabetical, deduplicated, formatted per journal. Target ≥15 resolved entries for a full paper.

---

PRE-EXPORT QUALITY CONTROL

Before generating the .docx, verify every item below. Fix or flag any failure.

Anti-fabrication (per references/anti-fabrication.md)

• [ ] Every numerical claim traces to a cached file, figure, or user-supplied fact (no invented values, sample sizes, units, or precision)
• [ ] Every citation was resolved via Semantic Scholar (no fabricated references, DOIs, or author names)
• [ ] Every causal or mechanistic claim has either a citation or an explicit hedge ("may", "one possible interpretation", "we hypothesize")
• [ ] No "well-known", "established", or "previous studies have shown" without a real citation
• [ ] No invented study-area facts (basin area, climate, geology) — all sourced or supplied
• [ ] No invented software versions, parameter rationales, random seeds, or calibration/validation splits beyond what the source code shows
• [ ] No invented author metadata (affiliations, ORCIDs, emails, CRediT roles, funding details)
• [ ] All bracketed placeholders ([CITATION NEEDED], [VALUE NEEDED], [VERIFY], [FROM USER]) are listed for the user

Content accuracy

• [ ] Every equation matches the cached source file (no invented terms)
• [ ] Every quantitative claim is traceable to a cached CSV row, figure, or user-supplied fact
• [ ] Study period, coordinate system, and sample sizes are consistent throughout
• [ ] Nothing in §4.5 / Limitations is invented — all items came from the user or preset

Language

• [ ] Zero first-person pronouns (I, we, my, our) anywhere
• [ ] All abbreviations defined on first use
• [ ] Terminology follows the journal style file (e.g. 'groundwater' one word, 'water table' two words for HJ)
• [ ] All displayed equations numbered sequentially; all symbols defined on first appearance

Citation integrity

• [ ] All in-text citations resolved via Semantic Scholar with real DOIs where available
• [ ] All [CITATION NEEDED] items listed for the user
• [ ] No fabricated citat

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Rekin226
• Source: Rekin226/paper-agent
• License: MIT

Install and usage instructions live in the source repository linked above.