mumdark/nature-writing-studio

nature-writing

Codex / Claude Code skill that rewrites scientific text into Nature Portfolio English. Distilled from 750 Nature family papers + editorial style guidelines. Knowledge base: 60 writing rules, 122 phrases, 132 AI-tell blacklist, 22 domain registers, 12 section skeletons.

What this skill does

Given any user-supplied scientific text - claim list, methods paragraph, draft abstract, full intro - produce:

1. text - Nature-level English rewrite.
2. text_compact - tighter 60-70% length version for comparison.
3. summary - metadata including:
   • logic_line: the reasoning behind the rewrite
   • rules_applied: rule IDs (R001-) from knowledge/writing_rules.tsv
   • patterns_used: phrase IDs (P001-) from knowledge/phrase_bank.tsv
   • ai_tells_avoided: AI-tell patterns removed
   • section: target section type
   • domain: inferred or provided domain

When target=multi_section, the orchestrator instead produces per-section blocks plus a top-level meta line containing version_used, logic_line, entity_registry, cross_section audit, and an optional degraded flag. The agent (or human-facing layer) is responsible for stitching the per-section text and text_compact strings into two final full-paper versions.

When to use this skill

Trigger when the user asks for any of:

• Nature-style writing, polish, rewrite, translate (Chinese -> Nature English), restructure
• Academic writing upgrade ("make it sound like Nature")
• Section drafts: abstract, introduction, methods, results, discussion, figure legend
• Sentence-level polish: turn a single Chinese sentence into Nature-grade English
• Multi-section writing: when the user supplies a full draft (Chinese or English) and wants a Nature-style full paper; produces per-section Markdown blocks + a summary line that an external agent stitches into two full-paper versions
• Anti-AI cleanup: remove delve into, navigate complexities, etc.
• Domain-specific drafting (physics / biology / medicine / materials / etc.)

Calling convention

The skill expects input shaped like:

input: <the text to rewrite>
target: <abstract | introduction | methods | results | discussion | figure_legend | sentence>
domain: <physics_condensed | biology_neuro | medicine_oncology | ... | general_nature>  (optional)

For sentence-level polish, target = sentence and input can be a single sentence or short phrase.

For full-paper orchestration:

target: multi_section
input: <full draft, Chinese or English, no section markers required>
domain: <optional>

The multi_section target invokes prompts/sections/multi_section.txt as an orchestrator. It auto-selects a section sequence (SP001 default), forces a logic_line + story_arc + entity_registry pre-step, invokes the per-section prompts with a shared context, and emits per-section Markdown blocks plus a summary line. The external agent (or human-facing layer) stitches the per-section text and text_compact into two final full-paper versions.

Operating procedure (for the agent)

1. Load prompts/system_writer.txt and prompts/style_guide.txt.

2. Load the section-specific prompt from prompts/sections/<target>.txt.

3. Read knowledge/writing_rules.tsv (60 rules), phrase_bank.tsv (122 phrases), anti_ai_patterns.tsv (132 blacklist), domain_register.tsv (if domain specified).

4. Rewrite the input following the section architecture and hedge tiers.

5. Anti-AI scan: substitute or delete any phrase matching anti_ai_patterns.tsv.

6. Render the result with the full output contract (text, text_compact, summary).

7. Self-check against the criteria in style_guide.txt "Quick Sanity Check Before Returning".

8. Multi-section orchestration (only when target=multi_section). Read prompts/sections/multi_section.txt and knowledge/cross_section_rules.tsv. Then:

   • Select the section sequence via the orchestrator's auto-version-selection (SP001 default).
   • Produce logic_line + story_arc + entity_registry as a hard prereq before any section writing.
   • Build a shared_context object and prepend it (fenced) to each section's input slice.
   • Invoke the matching per-section prompt for each section in version order.
   • Run the existing Verification Layer per section; re-run once on failure with corrected context.
   • After all sections, run the cross-section audit (entity diff, hedge ladder, citation gaps, abbreviation drift, transition breaks, numeric format, reference density) and emit meta.cross_section.
   • Return inline Markdown blocks in the same shape as the single-section example, with per-section ## text / ## text_compact / ## summary followed by a ## meta block listing version_used, logic_line, entity_registry, cross_section, degraded. Do not stitch prose; the agent does that.

9. Verify (Verification Layer, anti-fabrication): run the Verification Layer defined in prompts/system_writer.txt (Rule 1a / 1b / 2 / 2.5 / 3). Two tiers of traceability:

   • Rule 1a (HARD, must trace to input): any data claim about the user's study — data, statistics, sample sizes, time points, figure / citation tokens, user-specific identifiers (gene variants, cell line IDs, dataset accessions), user-stated conditions. A non-traceable data token IS fabrication; drop it.
   • Rule 1b (SOFT, allow freely): established scientific knowledge (mechanisms like alternative splicing or gene regulation, well-known pathways like Wnt or MAPK, model organisms as background, generic tissue / cell-type categories, generic phenotype / process terms). These set up the user's specific finding and are NEVER fabrication.
   • Decision rule: if the token would still be a true statement if you replaced the user's specific gene / number / cell with someone else's — it is Rule 1b (allow). If removing the user's input would make the sentence false — it is Rule 1a (must trace).
   • Treat em-dash (U+2014) as fabrication (Rule 2). The audit field is summary.untraceable_tokens — empty array [] = pass; non-empty = fabrication was found, the rewrite must be redone. In multi_section mode, the offending section is set to text=null and listed in meta.degraded. See the worked examples for delete-only, delete-plus-rewrite, and general-knowledge-allowed cases.

10. Em-dash scrub (deterministic): after the Verification Layer passes, run the strip_em_dash() function from prompts/system_writer.txt Rule 2.5 on text and text_compact independently. This is a hard post-process, not a soft rule, and runs AFTER every LLM pass. The returned text and text_compact MUST contain zero U+2014. If the substitution breaks a sentence (e.g., starting with , ), re-split at the sentence boundary. Do not invent new content to fix artifacts.

Output contract

## text
<Nature-level prose>

## text_compact
<tighter 60-70% length version>

## summary
- section: abstract
- domain: biology_neuro
- logic_line: <1-2 sentences>
- rules_applied: [R001, R005]
- patterns_used: [P005, P008]
- ai_tells_avoided: [delve_into, in_this_paper]

Default is inline Markdown. Switch to JSON or write to file only on explicit user request (e.g., "output as JSON", "save to .docx").

When target=multi_section, the output shape changes:

## text
<abstract text> ... <methods text>     (stitched by the orchestrator)

## text_compact
<abstract text_compact> ... <methods text_compact>

## summary
- mode: multi_section
- version_used: SP001
    "logic_line": "<single sentence>",
    "entity_registry": [...],
    "cross_section": {"entity_diff": [], "hedge_violations": [], "citation_gaps": [], "abbreviation_drift": [], "transition_breaks": [], "numeric_format_drift": [], "reference_density": [], "overall": "PASS"},
    "degraded": null
  }
}

The agent (or human-facing layer) iterates sections in order and concatenates each section's text (separated by blank lines) into one final text, then concatenates each section's text_compact into one final text_compact. Render both stitched versions for the user to read.

Example

User input (Chinese):

我们用CRISPRi 筛选了 287 个转录因子，找到 ZNF219 这个之前没报道过的转录因子，它抑制神经分化。 Skill output (Nature): Here we applied CRISPRi screening to a panel of 287 transcription factors (TFs) and identified ZNF219, a previously uncharacterized TF, as a repressor of neural differentiation.

Summary: Logic: open with Here + method + scale; lead finding identifies a previously uncharacterized TF. Rules applied: R005 (Here we), R008 (hedge), R017 (avoid novel overuse). Patterns: P010, P013. AI-tells avoided: none (input was clean). Section: sentence. Domain: biology_neuro. Verification: 0 em-dash; 0 fabricated tokens; pass. All tokens trace to the input (287, CRISPRi, ZNF219, transcription factor, repressor, neural differentiation).

Figure and Citation Placeholders

When the user input has no figure numbers or citations, insert explicit placeholders in the prose:

• Figures: (Fig. X), (Extended Data Fig. X), (Fig. 1a, b) (use literal X if not assigned).
• Citations: [citation needed], [1], [2], or numeric superscript placeholders like [1-].
• Statistics: [n = TBD], [P = TBD].

Never silently omit. Placeholders are honest; omission invites fabrication risk.

text vs text_compact

• text: full prose with hedges, parenthetical asides, calibrated redundancy.
• text_compact: 60-70% length. Drops parenthetical caveats and one of paired redundant clauses. Keeps lead claim, mechanism, anchor citations, and all anti-AI fixes.

The two versions must differ measurably - at least 20% length difference or 2 substantive cuts.

For multi-section orchestration, the stitched text is the concatenation of every section's text, and the stitched text_compact is the concatenation of every section's text_compact; the overall length ratio is preserved.

Anti-AI precision

The word key is allowed in some Nature contexts (key result, key step, key parameter). It is AI-tell ONLY when collocated with key player, key regulator, key insight, key step toward, key role. See knowledge/anti_ai_patterns.tsv for the precise regex.