Writing Fidelity

Writing Fidelity

Use this skill as a guardrail before editing or finalizing user-facing writing. Its purpose is to stop Codex from creating a cleaner-looking substitute that violates the user's source, constraints, correction history, or requested artifact.

This is the preservation layer, not the style layer. Use it before or alongside style work when facts, protected spans, version labels, page/rendered artifact identity, or user corrections could be lost. Route Chinese natural-prose passes to chinese-prose. Hand off natural Chinese prose, reader-facing "say it plainly" rewrites, and ordinary Chinese de-AI/template cleanup to chinese-prose. When the source is existing Chinese or Chinese-dominant scientific/technical material and the user asks for reorganization, structural rewrite, or document-level rewrite while preserving facts, numbers, formulas, citations, comparisons, conditions, limitations, paths, commands, or configuration details, hand off to scientific-rewrite instead of making chinese-prose the main route. Hand off English scientific prose, evidence-strength calibration, and defensive/self-undermining wording to scientific-prose.

Non-Negotiable Rule

Preserve first. Improve second.

Do not delete, rewrite, translate, reorder, rename, compress, or normalize protected content unless the user explicitly asked for that exact operation.

If the user's constraints conflict, report a blocked or partial status. Do not silently satisfy one constraint by violating another.

Workflow

1. Classify the task: polish, edit, rewrite, summarize, translate, layout, render, or source-faithful reconstruction.
2. Extract the user's latest corrections as hard constraints. Pay special attention to complaints about deletion, automatic rewriting, language changes, title changes, unreadable output, collisions, missing glyphs, OCR errors, formula spacing, or prior failed attempts.
3. Mark protected spans before editing. For polish, layout, and source-faithful reconstruction, this includes titles, headings, section order, labels, numbers, dates, units, formulas, variables, notation, code, paths, citations, Chinese/English language spans, user comments, caveats, examples, and quoted source text. For explicit structural rewrite routes, apply the structural rewrite handoff below instead of treating headings and source order as protected by default.
4. Apply only the requested operation. For polishing, preserve substantive content. For layout, do not change wording. For OCR cleanup, fix corruption without canonicalizing valid source notation.
5. When the deliverable is rendered or formatted, verify the artifact itself: page count/size, readable font, glyphs, formulas, collisions, clipping, whitespace, and absence of stray QA notes.
6. Before finalizing Markdown, LaTeX, PDF, slides, or reports, check artifact authority: title, method name, version label, selected row, final/best designation, and visible headline must match the latest user decision and evidence ledger. Do not invent lifecycle labels such as start <project> or final <project> unless the source explicitly defines them.
7. Finish with precise status and final artifact path. Do not treat reports, previews, audits, or side-by-side candidates as substitutes for the requested final artifact.
8. For Chinese, Markdown, PDF, slide, or report acceptance, verify readability authority before status: the first visible conclusion must be human-readable, and machine fields, paths, commands, status tokens, and audit trails must support rather than replace the conclusion.
9. If a Chinese Markdown/PDF/slide/report starts with machine fields, paths, commands, status tokens, audit trails, or mixed English process labels before a readable Chinese judgment, classify the deliverable as qa_failed until it is rewritten with a reader-facing conclusion first.

Hand Off

• Use chinese-prose when the main request is "中文说人话", "改自然一点", "不要 AI 味", "别像日志", "普通英文能翻就翻", or "别每句话一个 bullet".
• Use scientific-rewrite when the user provides existing Chinese or Chinese-dominant scientific/technical material and asks to reorganize, structurally rewrite, or rewrite the document into clearer Chinese while preserving facts, numbers, formulas, citations, comparisons, conditions, limitations, paths, commands, configuration keys, or other exact details.
• Use scientific-prose when the main request is polishing English Results, captions, rebuttals, slide text, or scientific reports without overclaiming or sounding defensive.
• Keep writing-fidelity active as a guardrail when those style passes must preserve numbers, equations, citations, paths, headings, user-approved rows, final/best labels, or rendered artifact identity.
• Do not treat "去 AI 味" as detector evasion, source laundering, or permission to hide authorship. The allowed goal is clearer prose that preserves facts and evidence boundaries.

Version and Label Fidelity

When a document contains dates, months, leaderboard rows, draft labels, model nicknames, or result variants, preserve the user's intended authority rather than normalizing names into a cleaner but false story.

• Do not use month, season, stage, or old-draft labels as a substitute for the actual selected version. If the best or author-approved row is known, name that row by metric, absolute date, and evidence path.
• Do not call a result final, complete, official, best, or the project method unless the evidence supports that exact status. Use bounded labels such as partial audit, leaderboard row by author rule, draft-only, or not verifiable when that is the real state.
• Do not restore stale numbers, stale headings, inactive blocks, or old draft labels just because they are polished or familiar. Prefer the latest explicit user decision and the best-supported evidence.
• When two rows are both permissible, distinguish the best row from an alternate historical row; do not hide that distinction behind month shorthand.
• A report, audit note, preview, candidate output, old draft, or side-by-side comparison is not the final deliverable unless the user explicitly selected it as final. Name the selected artifact and the authority for that selection.
• Treat audit result, leaderboard row, best row, author decision, old draft, and month/date labels as different authority classes. Do not promote one class into another for cleaner prose.
• If the user's correction names a phrase as wrong or unnatural, treat that phrase as protected-negative text: avoid reintroducing it in headings, PDF titles, captions, status summaries, or final answers.

Structural Rewrite Handoff

When a route such as scientific-rewrite explicitly declares STRUCTURAL_REWRITE_AUTHORIZED_BY_TASK, fidelity protects the content/evidence graph rather than the source's reader-facing structure.

This handoff is valid only for explicit rewrite / structural rewrite / heavy scientific-rewrite tasks. It does not apply to ordinary polishing, layout, OCR cleanup, source-faithful reconstruction, or user-protected outlines.

Under this handoff, these may change unless the user explicitly protected them:

• reader-facing headings;
• paragraph grouping;
• paragraph order;
• section order;
• table organization;
• where a local explanation is introduced.

When the structural route provides a Reader Plan, fidelity ownership follows reader-question bundles, not the original paragraph positions. A bundle may combine non-contiguous source spans, split one dense source span into a list, table, or formula walkthrough, or move a local explanation near the claim it supports, provided the evidence and caveat that bind the claim move with it.

These must still be preserved:

• claims and polarity;
• evidence, attribution, and citations;
• conditions, comparators, caveats, uncertainty, and negative findings;
• formulas, numbers, variables, datasets, methods, metrics, paths, and code according to their literal location roles;
• decision boundaries and conclusion strength.

The acceptance criterion is complete proposition/evidence coverage with lower reader inference burden, not source order or compression. A structural rewrite may pass fidelity when headings or paragraph order change, but it fails if any proposition is omitted, duplicated as new ownership, reattributed, strengthened, weakened without authorization, or moved away from the caveat that limits it.

Classify exact items as:

• inline-critical: exact material that belongs in the reader-facing scientific argument, such as numbers, formulas, metrics, datasets, method names, comparison-defining identifiers, and citations that support nearby claims.
• relocatable-trace: exact material that must remain somewhere in the final deliverable but can move to a technical/evidence appendix, such as repository paths, checkpoint paths, exhaustive file identities, implementation locators, and detailed audit trails.
• internal-workflow-trace: exact material that describes the current automation/review workflow rather than the scientific or technical subject, such as Reviewed Handoff state, Gate numbers, Planner/Reviewer/Executor status, CI/test summaries, Git commit hashes, branch/worktree status, GitHub Actions run ids, task-local results/, exports/private/, automation/reviewed_handoff/, .local-runtime/, plugin-cache paths, or CURRENT.json / RESULT.md / FINAL_REPORT.md.

Relocation cannot hide or delete limitations, negative results, uncertainty, contradicting evidence, decision conditions, attribution, or comparison boundaries. Ordinary reader-facing headings and internal workflow labels are not literal-protected by default. An inline-critical item is not preserved if it appears only in a technical appendix, token inventory, receipt, or trace list; it must remain in the reader-facing scientific context.

For a clean reader-facing scientific/technical rewrite, internal-workflow-trace is not preserved by dumping it into prose. Use it only to understand artifact authority and to avoid false claims. Omit it from the final candidate unless the user explicitly requests an audit log, release handoff, or repository status report. A final candidate fails fidelity/readability if it mixes scientific conclusions with workflow state, CI results, commit ids, task paths, or review status as if those were part of the scientific argument.

Red Flags

• The user asked for polishing, but content was removed.
• Chinese was converted to English or mixed-language text was normalized.
• Titles, labels, equations, notation, citations, or ordering changed.
• A visual/style change touched wording or structure.
• A generated PDF/slide/document was not rendered and inspected.
• The user repeated a correction and the workflow did not change.
• The final artifact uses month shorthand, stage shorthand, or invented start/final project labels instead of the best-supported version.
• Machine fields, paths, status tokens, or checklist fragments appear before the human-readable conclusion in a user-facing report.
• A Chinese Markdown/PDF/slide/report is technically rendered but its first visible section cannot be understood by the intended reader without reading audit logs, file paths, branch names, or English process tokens.

References

• Read references/failure-summary.md when the user has complained about prior writing failures, source fidelity, deletion, layout collisions, unreadable output, OCR/math corruption, or false completion.

Completion Standard

Use precise status:

• complete: requested artifact exists and checks pass.
• partial_complete: useful output exists, but some criteria are not met.
• qa_failed: output exists but fails fidelity, readability, rendering, or content checks.
• blocked: cannot proceed without user input or external state.
• blocked_target_not_met: a page/space target cannot be met without unapproved deletion or unreadable output.