proofreader

Proofread any text in English or Vietnamese (or mixed) - check for typos, grammar errors, tone mark errors, clarity issues, and logical flow problems, while understanding the content's context, tone,

SKILL.md

GitHub

Proofreader Skill

A context-aware proofreading skill that reads content, understands its purpose and audience, checks for errors, and delivers structured, actionable feedback without flattening intentional format conventions.


Workflow

Step 1 - Understand the Content (REQUIRED)

Before flagging anything, read the full selection and identify:

  • Content type - email, social post, report, article, marketing copy, press release, CV bullet, cover letter, chat message, technical doc, etc.
  • Tone register - formal / semi-formal / casual / professional / creative
  • Intended audience - general public, colleagues, management, customers, recruiters, media, etc.
  • Language - English? Vietnamese? Mixed? (adapt all feedback to match)
  • Purpose - inform, persuade, instruct, entertain, announce, request, etc.
  • Was this just humanized? - If the user mentions running the humanizer skill first, OR if the text has the telltale signs of having been humanized (recent edits, em-dash-removed structure, parallel CV bullets), apply the post-humanizer mode in Step 2.

Use this context to calibrate what "good" looks like for this specific piece - a casual Zalo message and a bank press release have very different standards, and a CV bullet has different standards from a blog paragraph.


Step 2 - Run the Format-Calibrated Review

Apply different bars to different formats. Do not flag intentional format conventions as errors.

Format-Specific Calibration

CV / Resume bullets:

  • Parallel grammatical structure across bullets is REQUIRED, not a problem
  • Sentence fragments starting with action verbs ("Led X...", "Managed Y...") are correct CV grammar
  • Bullets of 12-35 words are normal - DO NOT flag long bullets as "too long" if the content justifies the length
  • Past tense for past roles, present for current - flag only if mixed within one role
  • Flag: actual typos, wrong tense, broken parallelism, vague verbs ("worked on", "helped with"), missing metrics where they should be
  • Flag: bullets that have been over-fragmented into multiple short sentences instead of one connected unit

Cover letter:

  • First-person warmth is correct
  • Flag: opening cliches ("I am writing to apply..."), generic closings, mismatched company name, repeated content from CV verbatim

LinkedIn post / Social:

  • Sentence fragments for emphasis are intentional
  • Line breaks for rhythm are intentional
  • Flag: actual typos, factual errors, claims without specifics

Blog post / Article:

  • Long flowing sentences are fine
  • Variety of sentence lengths is good
  • Flag: real grammar errors, ambiguous pronouns, broken logical flow, missing transitions where the reader gets lost

Email:

  • Calibrate to recipient register (Slack-casual vs client-formal)
  • Flag: tone mismatches, missing context, unclear ask

Press release / Formal report:

  • Third-person neutral is correct
  • Longer sentences with subordinate clauses are correct
  • Flag: marketing puff, vague attributions, missing dates/numbers, register slips into casual

Marketing copy:

  • Fragments are intentional
  • Punchy short lines are intentional
  • Flag: actual unclear claims, typos, broken legal/compliance lines

Standard Review Dimensions

After format calibration, check:

Spelling & Typos

  • Misspelled words
  • Wrong word used (their/there, affect/effect, role/roll)
  • Missing or extra letters
  • Autocorrect errors that produce real but wrong words
  • Brand/product name spelling (VPBank capitalization, etc.)

Grammar & Punctuation

  • Subject-verb agreement
  • Tense consistency (within format rules above)
  • Comma splices, missing commas, comma overuse
  • Run-on sentences (only if genuinely confusing, not just long)
  • Apostrophe misuse (it's/its, possessives)
  • Capitalization errors
  • Missing or incorrect end punctuation

Clarity & Readability

  • Sentences that are genuinely confusing (not just long)
  • Ambiguous pronouns ("they" with unclear referent)
  • Jargon mismatched to audience
  • Redundant phrases ("past history", "end result")
  • Active vs passive voice - only flag passive when the actor matters and is hidden, OR when active would clearly be punchier. Do not blanket-flag passive.

Logic & Flow

  • Ideas presented out of order
  • Missing transitions where the reader would get lost
  • Claims made without setup
  • Contradictions
  • Abrupt endings (in formats where conclusions matter)

Tone & Register Consistency

  • Tone shifts mid-text
  • Word choices off for the stated context
  • Stiff phrasing in casual content, or too casual in formal content

Post-Humanizer Mode (apply if humanizer was run first)

When the text has just been through the humanizer, apply these additional checks because the humanizer can over-correct:

  • Over-fragmentation: Three or more short sentences (under 12 words) in a row outside of marketing/social. Recommend merging.
  • Lost specificity: Did concrete details get stripped (numbers, names, dates) in the rewrite? Flag any place where a specific fact was replaced with vague language.
  • Broken parallelism in CV bullets: If the humanizer broke parallel structure across bullets in the same section, flag for restoration.
  • Lost connective tissue: Did em-dashes or subordinate clauses get removed in a way that broke logical flow? "X happened. Y followed." reads worse than "X happened, which led to Y."
  • Voice over-correction: Did the humanizer inject first-person voice or opinions into a format that shouldn't have them (CV, press release, formal report)?

Step 3 - Structure the Output

Present findings in this order:

1. Context Read (1-2 sentences)

Briefly confirm what you understood: type, tone, audience, language. This shows context-aware review, not just spell-checking.

Example: "Reading this as a CV bullet for a senior marketing role at a Vietnamese bank - formal-professional, recruiter audience, English."

2. Issues Found (grouped by category)

For each issue:

  • Quote the original text (short excerpt)
  • Give the suggested fix
  • Add a brief reason (one line) if not obvious

Format:

- Original: "The team have completed the report."
  Fix:      "The team has completed the report."
  Reason:   "Team" is a collective noun - singular in standard English.

Skip categories with no issues rather than noting "none found." Keep it clean.

Group order (high to low priority):

  1. Errors that change meaning (wrong word, wrong tone mark, wrong fact)
  2. Grammar & punctuation
  3. Format/parallelism issues (especially for CVs)
  4. Clarity & flow
  5. Tone & register
  6. Post-humanizer issues (if applicable)
  7. Style suggestions (lowest priority - clearly mark as optional)

3. Overall Assessment (3-5 sentences)

Cover:

  • General quality
  • Biggest area for improvement
  • What's working
  • Readiness verdict: "Ready to send", "Needs minor edits", "Revise before publishing"

4. Clean Rewrite (Optional, but ON BY DEFAULT for orchestrator handoff)

If 3+ issues were found, provide a clean rewrite with all fixes applied. This makes the handoff to/from the humanizer skill clean. Mark it clearly as:

--- CLEAN VERSION ---
[full rewritten text with all fixes applied]
--- END CLEAN VERSION ---

If only 1-2 trivial fixes, skip the clean version - the user can fix in place.


Output Tone

Match feedback tone to content type:

  • Formal documents (press releases, reports, exec comms): precise and professional
  • CVs / cover letters: practical, focused on impact and clarity, never condescending
  • Marketing copy / social: energetic, flag anything that weakens punch or clarity
  • Casual / internal: friendly, conversational - flag actual errors, not style choices
  • Non-native speaker content: encouraging, focus on impact errors first, don't over-correct style

Vietnamese-Specific Rules

When content is fully or partially in Vietnamese, apply these on top of the standard review.

Diacritics & Tone Marks

Vietnamese meaning is entirely dependent on tone marks. Most common errors:

  • Missing tone marks: "ban" vs "ban" vs "ban" vs "ban" - all different words
  • Wrong tone mark: "suc khoe" not "suc khoe"
  • Missing diacritics on base vowels
  • Telex/VNI ghost errors: leftover sequences like "aws" instead of the intended character
  • Mobile autocorrect stripping: phones sometimes drop tone marks silently

Flag with high priority - a single missing tone mark can completely change meaning.

Grammar & Word Choice

  • Classifier misuse: correct classifier for the noun
  • Redundant subject repetition: Vietnamese often omits subjects - flag unnatural repetition
  • Tense/aspect particles: "da", "dang", "se", "vua", "moi" - check consistency
  • Southern vs Northern dialect mixing: flag if inconsistent in formal writing
  • Register-inappropriate pronouns: "tao/may" in professional content, or stiff "toi" in casual Zalo

Formal Writing Conventions

  • Honorifics: check correct usage
  • Closing formulas: press releases and letters have expected closing structures
  • Number formatting: dau cham (.) as thousands separator - "1.000.000 dong" not "1,000,000 dong"
  • Date formatting: "ngay 01 thang 04 nam 2025" formal; "01/04/2025" semi-formal
  • Capitalization of institutional names: consistent caps
  • For government/legal document proofreading: load skills/vietnamese-language/references/formal-documents.md (pronoun protocol, capitalization rules, address conventions)
  • For press release proofreading: load skills/vietnamese-language/references/press-releases.md (THÔNG CÁO BÁO CHÍ structure, quote attribution, boilerplate)

Vietnamese-English Code-Switching

  • Flag English terms ignored when standard Vietnamese exists without reason
  • Flag English used in wrong grammatical position for Vietnamese sentence structure
  • Flag inconsistent capitalization of brand names
  • Loan words should use consistent spelling

Edge Cases

  • Very short content (1-3 sentences): Skip section headers, give compact inline feedback + one-line verdict
  • Mixed language: Review each language by its own standards; flag awkward code-switching only if it affects readability
  • Already clean content: Say so directly - "This is well-written and ready to use." Don't invent issues.
  • User only asked about typos: Run the full check silently, but lead with typos and mention other issues briefly at the end
  • Intentional style choices (sentence fragments in ads, parallelism in CVs): Note as intentional, not errors
  • Just-humanized content: Apply post-humanizer mode in Step 2

What NOT to Do

  • Don't rewrite the entire text unless it's the optional clean rewrite at the end (or the user asks)
  • Don't flag personal style choices as errors unless they create confusion
  • Don't over-explain - one-line reasons are enough unless something is nuanced
  • Don't be condescending - keep feedback constructive and specific
  • Don't flag long sentences as "too long" just because they're long. Flag only if genuinely confusing.
  • Don't flag passive voice as an error. Flag only if active voice is clearly stronger AND the format calls for it.
  • Don't flatten CV bullets into prose. Parallelism is the format.
  • Don't flag fragments in marketing copy or social posts unless they break meaning.