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.
Hand off product-interface microcopy to product-ui-copy when the target text
is a label, CTA, state, help text, onboarding line, settings explanation,
trust/privacy disclosure, permission prompt, or other short UI copy. In that
route, this skill protects product meaning and escalation boundaries; it does
not become the wording owner.
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
- Classify the task:
polish, edit, rewrite, summarize, translate,
layout, render, or source-faithful reconstruction.
- 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.
- 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.
- 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.
- 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.
- 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.
- 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.
- 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.
- 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
product-ui-copy when the main request is to improve product-interface
microcopy: labels, buttons, CTA/action text, empty/loading/error/success
states, onboarding, settings help, trust/privacy disclosure, permission
prompts, browser-extension popup copy, desktop/WebView settings copy, mobile
UI wording, landing-page interface copy, or locale-specific UI wording.
- 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.
- Keep
writing-fidelity active as a guardrail for Product UI Copy whenever
product state, eligibility, user consequence, payment, subscription,
deletion, retention, consent, privacy, safety, external visibility,
reversibility, required trust/legal disclosure, or exact product/brand/feature
identity could drift.
- 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.
Product UI Copy Handoff
For interface copy, protect meaning before improving wording. Product UI Copy
may make text shorter, clearer, more natural, or more locale-appropriate, but
it must not silently change:
- who can take the action;
- whether the action is reversible;
- whether data becomes public, shared, retained, deleted, synced, charged, or
sent to another system;
- whether a state is final, pending, diagnostic, failed, optional, required, or
unavailable;
- pricing, cancellation, subscription, consent, privacy, safety, legal, trust,
or eligibility meaning;
- exact product, brand, feature, policy, or technical identities.
If those facts are missing or contradictory, classify the issue as
PRODUCT SEMANTICS or LEGAL/TRUST/SAFETY and ask product/domain/legal
authority to resolve it. Do not launder uncertainty into polished copy.
Frontend Design owns content architecture and rendered acceptance. If the
problem is that text is in the wrong place, too much is visible, neighboring
copy conflicts, disclosure is at the wrong level, or the viewport makes the
relationship unreadable, route that part back as CONTENT ARCHITECTURE instead
of treating it as a wording-only edit.
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.
1---2name: writing-fidelity3description: Preserve facts, corrections, labels, structure, equations, citations, version authority, and final artifact identity during writing edits. Route Chinese natural-prose passes to chinese-prose, source-faithful structural scientific/technical rewrites to scientific-rewrite, and English scientific style passes to scientific-prose.4status: active5provenance: user-authored6trusted: false7requires_network: false8writes_files: true9executes_code: false10secrets_needed:11last_reviewed: 2026-07-3012profile_tags:13 - writing14 - global15recommended_scope: global16icon_small: assets/app-facing.svg17icon_large: assets/app-facing.svg18metadata:19 source: https://skillmd.com/skills/yuukias/writing-fidelity20---2122# Writing Fidelity2324Use this skill as a guardrail before editing or finalizing user-facing writing.25Its purpose is to stop Codex from creating a cleaner-looking substitute that26violates the user's source, constraints, correction history, or requested27artifact.2829This is the preservation layer, not the style layer. Use it before or alongside30style work when facts, protected spans, version labels, page/rendered artifact31identity, or user corrections could be lost. Route Chinese natural-prose passes32to chinese-prose. Hand off natural Chinese prose, reader-facing "say it33plainly" rewrites, and ordinary Chinese de-AI/template cleanup to34`chinese-prose`. When the source is existing Chinese or35Chinese-dominant scientific/technical material and the user asks for36reorganization, structural rewrite, or document-level rewrite while preserving37facts, numbers, formulas, citations, comparisons, conditions, limitations,38paths, commands, or configuration details, hand off to `scientific-rewrite`39instead of making `chinese-prose` the main route. Hand off English scientific prose,40evidence-strength calibration, and defensive/self-undermining wording to41`scientific-prose`.4243Hand off product-interface microcopy to `product-ui-copy` when the target text44is a label, CTA, state, help text, onboarding line, settings explanation,45trust/privacy disclosure, permission prompt, or other short UI copy. In that46route, this skill protects product meaning and escalation boundaries; it does47not become the wording owner.4849## Non-Negotiable Rule5051Preserve first. Improve second.5253Do not delete, rewrite, translate, reorder, rename, compress, or normalize54protected content unless the user explicitly asked for that exact operation.5556If the user's constraints conflict, report a blocked or partial status. Do not57silently satisfy one constraint by violating another.5859## Workflow60611. Classify the task: `polish`, `edit`, `rewrite`, `summarize`, `translate`,62 `layout`, `render`, or `source-faithful reconstruction`.632. Extract the user's latest corrections as hard constraints. Pay special64 attention to complaints about deletion, automatic rewriting, language65 changes, title changes, unreadable output, collisions, missing glyphs, OCR66 errors, formula spacing, or prior failed attempts.673. Mark protected spans before editing. For `polish`, `layout`, and68 `source-faithful reconstruction`, this includes titles, headings, section69 order, labels, numbers, dates, units, formulas, variables, notation, code,70 paths, citations, Chinese/English language spans, user comments, caveats,71 examples, and quoted source text. For explicit structural rewrite routes,72 apply the structural rewrite handoff below instead of treating headings and73 source order as protected by default.744. Apply only the requested operation. For polishing, preserve substantive75 content. For layout, do not change wording. For OCR cleanup, fix corruption76 without canonicalizing valid source notation.775. When the deliverable is rendered or formatted, verify the artifact itself:78 page count/size, readable font, glyphs, formulas, collisions, clipping,79 whitespace, and absence of stray QA notes.806. Before finalizing Markdown, LaTeX, PDF, slides, or reports, check artifact81 authority: title, method name, version label, selected row, final/best82 designation, and visible headline must match the latest user decision and83 evidence ledger. Do not invent lifecycle labels such as `start <project>` or84 `final <project>` unless the source explicitly defines them.857. Finish with precise status and final artifact path. Do not treat reports,86 previews, audits, or side-by-side candidates as substitutes for the requested87 final artifact.888. 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.899. 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.9091## Hand Off9293- Use `chinese-prose` when the main request is "中文说人话", "改自然一点",94 "不要 AI 味", "别像日志", "普通英文能翻就翻", or "别每句话一个 bullet".95- Use `product-ui-copy` when the main request is to improve product-interface96 microcopy: labels, buttons, CTA/action text, empty/loading/error/success97 states, onboarding, settings help, trust/privacy disclosure, permission98 prompts, browser-extension popup copy, desktop/WebView settings copy, mobile99 UI wording, landing-page interface copy, or locale-specific UI wording.100- Use `scientific-rewrite` when the user provides existing Chinese or101 Chinese-dominant scientific/technical material and asks to reorganize,102 structurally rewrite, or rewrite the document into clearer Chinese while103 preserving facts, numbers, formulas, citations, comparisons, conditions,104 limitations, paths, commands, configuration keys, or other exact details.105- Use `scientific-prose` when the main request is polishing English Results,106 captions, rebuttals, slide text, or scientific reports without overclaiming or107 sounding defensive.108- Keep `writing-fidelity` active as a guardrail when those style passes must109 preserve numbers, equations, citations, paths, headings, user-approved rows,110 final/best labels, or rendered artifact identity.111- Keep `writing-fidelity` active as a guardrail for Product UI Copy whenever112 product state, eligibility, user consequence, payment, subscription,113 deletion, retention, consent, privacy, safety, external visibility,114 reversibility, required trust/legal disclosure, or exact product/brand/feature115 identity could drift.116- Do not treat "去 AI 味" as detector evasion, source laundering, or permission117 to hide authorship. The allowed goal is clearer prose that preserves facts and118 evidence boundaries.119120## Product UI Copy Handoff121122For interface copy, protect meaning before improving wording. Product UI Copy123may make text shorter, clearer, more natural, or more locale-appropriate, but124it must not silently change:125126- who can take the action;127- whether the action is reversible;128- whether data becomes public, shared, retained, deleted, synced, charged, or129 sent to another system;130- whether a state is final, pending, diagnostic, failed, optional, required, or131 unavailable;132- pricing, cancellation, subscription, consent, privacy, safety, legal, trust,133 or eligibility meaning;134- exact product, brand, feature, policy, or technical identities.135136If those facts are missing or contradictory, classify the issue as137`PRODUCT SEMANTICS` or `LEGAL/TRUST/SAFETY` and ask product/domain/legal138authority to resolve it. Do not launder uncertainty into polished copy.139140Frontend Design owns content architecture and rendered acceptance. If the141problem is that text is in the wrong place, too much is visible, neighboring142copy conflicts, disclosure is at the wrong level, or the viewport makes the143relationship unreadable, route that part back as `CONTENT ARCHITECTURE` instead144of treating it as a wording-only edit.145146## Version and Label Fidelity147148When a document contains dates, months, leaderboard rows, draft labels, model149nicknames, or result variants, preserve the user's intended authority rather150than normalizing names into a cleaner but false story.151152- Do not use month, season, stage, or old-draft labels as a substitute for153 the actual selected version. If the best or author-approved row is known,154 name that row by metric, absolute date, and evidence path.155- Do not call a result `final`, `complete`, `official`, `best`, or the project156 method unless the evidence supports that exact status. Use bounded labels such157 as `partial audit`, `leaderboard row by author rule`, `draft-only`, or158 `not verifiable` when that is the real state.159- Do not restore stale numbers, stale headings, inactive blocks, or old draft160 labels just because they are polished or familiar. Prefer the latest explicit161 user decision and the best-supported evidence.162- When two rows are both permissible, distinguish the best row from an alternate163 historical row; do not hide that distinction behind month shorthand.164- 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.165- 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.166- If the user's correction names a phrase as wrong or unnatural, treat that167 phrase as protected-negative text: avoid reintroducing it in headings, PDF168 titles, captions, status summaries, or final answers.169170## Structural Rewrite Handoff171172When a route such as `scientific-rewrite` explicitly declares173`STRUCTURAL_REWRITE_AUTHORIZED_BY_TASK`, fidelity protects the content/evidence174graph rather than the source's reader-facing structure.175176This handoff is valid only for explicit rewrite / structural rewrite / heavy177scientific-rewrite tasks. It does not apply to ordinary polishing, layout, OCR178cleanup, source-faithful reconstruction, or user-protected outlines.179180Under this handoff, these may change unless the user explicitly protected them:181182- reader-facing headings;183- paragraph grouping;184- paragraph order;185- section order;186- table organization;187- where a local explanation is introduced.188189When the structural route provides a Reader Plan, fidelity ownership follows190reader-question bundles, not the original paragraph positions. A bundle may191combine non-contiguous source spans, split one dense source span into a list,192table, or formula walkthrough, or move a local explanation near the claim it193supports, provided the evidence and caveat that bind the claim move with it.194195These must still be preserved:196197- claims and polarity;198- evidence, attribution, and citations;199- conditions, comparators, caveats, uncertainty, and negative findings;200- formulas, numbers, variables, datasets, methods, metrics, paths, and code201 according to their literal location roles;202- decision boundaries and conclusion strength.203204The acceptance criterion is complete proposition/evidence coverage with lower205reader inference burden, not source order or compression. A structural rewrite206may pass fidelity when headings or paragraph order change, but it fails if any207proposition is omitted, duplicated as new ownership, reattributed, strengthened,208weakened without authorization, or moved away from the caveat that limits it.209210Classify exact items as:211212- `inline-critical`: exact material that belongs in the reader-facing scientific213 argument, such as numbers, formulas, metrics, datasets, method names,214 comparison-defining identifiers, and citations that support nearby claims.215- `relocatable-trace`: exact material that must remain somewhere in the final216 deliverable but can move to a technical/evidence appendix, such as repository217 paths, checkpoint paths, exhaustive file identities, implementation locators,218 and detailed audit trails.219- `internal-workflow-trace`: exact material that describes the current220 automation/review workflow rather than the scientific or technical subject,221 such as Reviewed Handoff state, Gate numbers, Planner/Reviewer/Executor222 status, CI/test summaries, Git commit hashes, branch/worktree status,223 GitHub Actions run ids, task-local `results/`, `exports/private/`,224 `automation/reviewed_handoff/`, `.local-runtime/`, plugin-cache paths, or225 `CURRENT.json` / `RESULT.md` / `FINAL_REPORT.md`.226227Relocation cannot hide or delete limitations, negative results, uncertainty,228contradicting evidence, decision conditions, attribution, or comparison229boundaries. Ordinary reader-facing headings and internal workflow labels are not230literal-protected by default. An `inline-critical` item is not preserved if it231appears only in a technical appendix, token inventory, receipt, or trace list;232it must remain in the reader-facing scientific context.233234For a clean reader-facing scientific/technical rewrite, `internal-workflow-trace`235is not preserved by dumping it into prose. Use it only to understand artifact236authority and to avoid false claims. Omit it from the final candidate unless the237user explicitly requests an audit log, release handoff, or repository status238report. A final candidate fails fidelity/readability if it mixes scientific239conclusions with workflow state, CI results, commit ids, task paths, or review240status as if those were part of the scientific argument.241242## Red Flags243244- The user asked for polishing, but content was removed.245- Chinese was converted to English or mixed-language text was normalized.246- Titles, labels, equations, notation, citations, or ordering changed.247- A visual/style change touched wording or structure.248- A generated PDF/slide/document was not rendered and inspected.249- The user repeated a correction and the workflow did not change.250- The final artifact uses month shorthand, stage shorthand, or invented251 `start/final` project labels instead of the best-supported version.252- Machine fields, paths, status tokens, or checklist fragments appear before253 the human-readable conclusion in a user-facing report.254- A Chinese Markdown/PDF/slide/report is technically rendered but its first255 visible section cannot be understood by the intended reader without reading256 audit logs, file paths, branch names, or English process tokens.257258## References259260- Read `references/failure-summary.md` when the user has complained about prior261 writing failures, source fidelity, deletion, layout collisions, unreadable262 output, OCR/math corruption, or false completion.263264## Completion Standard265266Use precise status:267268- `complete`: requested artifact exists and checks pass.269- `partial_complete`: useful output exists, but some criteria are not met.270- `qa_failed`: output exists but fails fidelity, readability, rendering, or271 content checks.272- `blocked`: cannot proceed without user input or external state.273- `blocked_target_not_met`: a page/space target cannot be met without274 unapproved deletion or unreadable output.