human-eyes

name	human-eyes
description	Audits and edits prose that uses writing habits commonly associated with machine-generated text. Use for audits, suggestions, rewrites, or new drafts when the user wants to remove AI-feeling phrasing, rhythm, structure, or voice.

Human-eyes

Human-eyes is an opinionated writing-cleanup skill. It identifies specific constructions the reader wants removed and never classifies authorship.

The writer controls every edit. Audit is the default action. Suggestions, Rewrite, and Write run only when requested.

Resolve the skill directory

Resolve SKILL_DIR to the absolute directory containing this SKILL.md before running any command. Call scripts and read references through SKILL_DIR; do not assume the current working directory is the repository.

Choose the action

Action	Use when	Result
Audit	The user supplies prose for review or does not request an edit	Complete deterministic and semantic findings
Suggestions	The user asks how to fix audited prose	One context-validated replacement per finding
Rewrite	The user asks for revised source prose	Source Audit, revised prose, fresh output Audit, change report
Write	The user supplies a brief without a draft	Draft, complete Audit, change report

If intent is ambiguous, default to Audit. Never treat a surface-only scan as an Audit.

Run a complete Audit

Save the exact input to a temporary UTF-8 Markdown file. Preserve source punctuation and line breaks.

Create a private work bundle:

python3 "$SKILL_DIR/scripts/grade.py" preflight "$INPUT_PATH" --work-bundle "$WORK_PATH"

When preserved slide or caption boundaries are available, supply a structure manifest:

python3 "$SKILL_DIR/scripts/grade.py" preflight "$INPUT_PATH" --work-bundle "$WORK_PATH" --structure-manifest "$STRUCTURE_PATH"

Read the complete input, scripts/judgement.json, and the bundle's semantic_candidates. Fill semantic_answers with exactly one answer for every current registry record. Candidates focus attention but do not limit the reading.

Each semantic answer contains only:

• id: the registry ID.
• status: clear or flagged, derived from flagged_when.
• answer: the exact shape required by answer_schema.
• evidence: a list of exact input substrings. Use an empty list for a clear answer with no quoted evidence.

For a list answer, every item's first field is an exact input substring. Explain the missing antecedent, criterion, information, context, or logical relationship in the item's second field. Do not add severity; the grader owns severity.

Write the completed bundle back to WORK_PATH, then render the Audit:

python3 "$SKILL_DIR/scripts/grade.py" audit "$INPUT_PATH" --work-bundle "$WORK_PATH" --format markdown --depth balanced

Use --full-report when the user asks for complete coverage tables. Use --format json when a later action needs structured findings. Print Markdown output without paraphrasing quoted evidence.

If the bundle is missing, partial, stale, malformed, or contradictory, stop and repair the semantic reading. Do not render a completed-looking Audit.

Use surface-only mode

Surface-only mode exists for deterministic development checks:

python3 "$SKILL_DIR/scripts/grade.py" audit "$INPUT_PATH" --surface-only --format markdown --depth all

The result is incomplete. Do not offer or run Suggestions, Rewrite, or Write from it.

If Python or the grader is unavailable, report the limitation and stop. A manual scan cannot satisfy complete Audit coverage.

Produce Suggestions

Run a complete source Audit first. Read references/alternatives.md for lexical findings and references/process.md for structural findings.

For each proposed replacement:

1. Apply the replacement to a copy of its surrounding sentence or paragraph.
2. Save the changed context as a new input.
3. Create and complete a new work bundle.
4. Run a complete Audit of the changed context.
5. Return the suggestion only when it clears the target finding without introducing another required finding.

Keep quotations, facts, stance, and factual qualifications unchanged. A suggestion count matches the finding count unless a finding cannot be repaired without changing meaning; name the constraint instead of inventing a replacement.

Produce a Rewrite

Run a complete source Audit, then read references/process.md and references/voice.md.

Use Balanced depth unless the user requests All. Balanced repairs hard failures and strong warnings and reviews contextual findings against meaning. All attempts to repair every finding while preserving the source.

Save the rewrite as a new file. Create a new work bundle and complete a fresh Audit. Never reuse the source bundle. Revise and repeat for no more than three passes.

Record the required finding IDs after each pass. Stop if a pass introduces a new required finding or if the sets alternate between two states. Report the regression or oscillation instead of continuing.

Return:

• The source Audit.
• The rewritten prose.
• A compact change report using Changes made, Preserved, and Remaining findings.

If required findings remain after the third pass, list them. Do not claim completion.

Produce new writing

Draft from the supplied brief without inventing facts, personal experience, opinions, emotion, humour, uncertainty, or names. Read references/voice.md before drafting.

Save the draft, create a work bundle, complete every semantic answer, and run a full Audit. Revise and re-audit for no more than three passes. Return the draft and compact change report only after a clean required-finding result, or return the residual findings at the limit.

Record required finding IDs after every pass and stop on newly introduced findings or alternating sets.

Preserve meaning

The source or brief controls argument, facts, examples, stance, genre, point of view, and factual uncertainty. Do not manufacture irregularity to make prose appear human. Do not add a parenthetical doubt, forced register change, anecdote, or personal detail.

Every generated heading must contain no parentheses. Every generated demonstrative and definite description needs an identifiable referent. Remove performed candour, relevance padding, redundant recap, counted slogans, unsupported evaluation, empty connections, staccato filler, and prose aimed at absent chat context.

Keep the product boundary

Describe the construction, reading problem, and edit. Do not estimate provenance, assign a probability, accuse the writer, or call the result proof of authorship.