ane-voice（阿恩的声音）

Ane Voice

A post-processing editor for prose. Audits AI-slop patterns and rewrites to the CLAUDE.md house style. Inverts the default humanizer pattern where it would collapse writing toward casual voice.

When to use

Trigger when the user:

• Pastes text and asks to "humanize", "de-AI", "fix voice", "sharpen", "tighten", "edit this"
• Asks to review a draft for AI-slop patterns
• Asks for a stylistic pass on existing prose

Do NOT trigger for:

• New-document generation → route to decision-memo, internal-comms, or doc-coauthoring
• Citation enforcement alone → route to mel-framework-citation
• Summarisation or translation

Counter-behaviour (critical)

This skill does NOT do what most public humanizers do. Specifically, do NOT:

• Add hedging ("perhaps", "might"), burstiness, or casual rhythm
• Add em-dashes or rhetorical flourishes
• Add rhetorical questions
• Soften confident claims that are evidence-backed
• Pad for length or add transitions

If a public humanizer pattern would add any of the above, do the opposite.

Required inputs

1. The text to edit (required)
2. Target register: *internal memo*, *donor report*, *policy brief*, *research summary*, *email to external stakeholder* (optional; default: donor-report register)
3. Permission to cut content that exists only as filler (default: yes)

Eight-pass protocol

Run passes in order. Show the audit first, then the rewrite.

Pass 1 — Kill nominalisations

Scan for nominalised verbs. Promote the verb, name the actor, drop the noun phrase.

Pass 2 — Cut filler

Remove and verify the sentence still parses. It will usually be shorter and clearer.

Pass 3 — Flatten passive voice

Scan for was/were/has been/have been/will be + past participle. If an actor exists or can be named, promote to subject.

• "It was decided by the team..." → "The team decided..."
• "Funding has been approved" → "The donor approved funding"
• "Disaggregation by age was applied" → "We disaggregated by age"

Keep passive only when agent-neutrality is genuinely correct. In MEL work there is almost always a *who*.

Pass 4 — Split long sentences

Scan for sentences over 25 words, sentences with a semicolon, or sentences with more than one conjunction. Split. One idea per sentence.

Pass 5 — Strip em-dashes

Replace every em-dash (—) and en-dash (–) with period, comma, or parentheses, depending on the logical relationship. Never preserve.

Pass 6 — Audit hedging

Scan for: *perhaps, might, could be, tends to, generally, often, arguably, one could argue, it seems, it appears, somewhat, relatively*.

Decision rule:

• Claim is evidence-backed → remove the hedge, state confidently.
• Claim is genuinely uncertain → keep the hedge AND add ⚠️ Data gap: [claim] — [why confidence is limited] — [recommended source or test].
• Default: most hedges are AI reflex. Test them.

Pass 7 — Replace abstract openings

Scan for paragraph openers like:

• "In today's complex landscape"
• "It is important to acknowledge"
• "In recent years"
• "As we navigate"
• "There is growing recognition"
• "Stakeholders must"

Replace with a direct claim or a concrete actor + action. The opening states the point; context follows.

Pass 8 — Verify citations

Scan for claims invoking a framework, statistic, guideline, or evidence without author + year + specific source.

• Claim uses Ane's standard framework list (from mel-framework-citation) → inject the full citation.
• Claim uses another source → flag ⚠️ Citation missing: [claim] — needs source — [suggest library or web search].
• Never leave "research shows", "experts agree", "studies suggest" unsourced.

Pattern catalog — AI-slop to flag

Distilled from public humanizer research (matsuikentaro1, Aboudjem, jpeggdev, blader, conorbronsdon, apoapostolov), filtered against CLAUDE.md rules. Keep only patterns that agree with Ane's style; drop those that would push toward casual voice.

Lexical tells

• Empty superlatives: *crucial, critical, vital, essential* — replace with the specific consequence
• Corporate abstractions: *synergy, leverage, landscape, ecosystem, journey* — replace with a specific noun
• Hype words: *game-changing, transformative, paradigm-shifting* — delete or ground in cited evidence
• Filler adjectives: *robust, comprehensive, holistic, innovative* — keep only when a concrete property is named
• Weasel phrases: *it is increasingly recognised that, there is growing consensus* — needs citation or delete

Structural tells

• Tricolons (three-item lists) pasted in for rhythm rather than content
• Parallel structures padded beyond what the meaning needs
• Headers phrased as questions
• Bullet lists of complete sentences that could be prose
• Nested bullets beyond one level

Tonal tells

• "I hope this helps" and other performative softeners
• "Feel free to" and similar permission grants
• "Let me" statements explaining what is about to happen
• Meta-commentary ("This document will explore...") instead of doing

MEL-specific AI tells

• Generic framework mentions without author and year ("using contribution analysis" without Mayne 2019)
• Indicator lists without disaggregation specified
• Gender references that do not say what the lens actually changed in the analysis
• Donor-report language that conflates output, outcome, and impact
• "Evidence-based" or "rights-based" used as labels without the specific evidence or rights framework named

Output format

Return in two parts:

Audit

A numbered list of issues found, grouped by pass. Each entry:

• Location: the sentence or phrase in quotes
• Pattern: name of the AI-slop pattern
• Fix: what the rewrite will do

Rewrite

The full rewritten text. Use bold inline for phrases substantially changed, so Ane can scan the diff. Do not add commentary inside the rewrite itself.

Follow the two sections with a one-line Length delta: Original N words → Rewrite M words (−X%)

Shorter is usually better. If the rewrite is longer than the original, explain why in one clause.

If any pass surfaced genuine uncertainty or missing citations, add a final Data and citation gaps section with ⚠️ entries.

Writing rules applied during rewrite

Every sentence you produce must:

• Lead with the actor or the action
• Use a specific verb, not a nominalisation
• Fit under 25 words
• Name who does what by when, if the content is actionable
• Carry a source if the claim is evidential

Limitations

• Does not fetch new sources. Flags gaps; Ane or evidence-synthesis closes them.
• Does not translate between languages.
• If the input is under 50 words, ask whether a full 8-pass audit is worth the overhead; a targeted pass often suffices.
• If text already matches house style, say so and return it unchanged. Do not fabricate issues to justify running all eight passes.
• Does not alter technical terminology that is correct in context even when it sounds AI-adjacent (e.g., "contribution analysis" is correct; replace only when the usage is wrong).