Style Harmonizer

Style Harmonizer (de-slot editor)

Triggers & routing

• Trigger: style harmonizer, de-template stems, remove slot phrases, discourse stems, 写作风格统一, 去槽位句式, 去生成器味.
• Use when: writer-selfloop is PASS but output/WRITER_SELFLOOP_TODO.md flags Style Smells (e.g., repeated count-based openers), or the draft reads like many sections share the same rhythm.

Purpose: remove subtle generator-voice signals that can survive structural gates.

This skill is not a full rewrite. It is a targeted rewrite queue:

• only touch the specific sections/*.md files flagged under ## Style Smells
• keep facts and citation keys unchanged

Inputs

Required:

• output/WRITER_SELFLOOP_TODO.md (Style Smells section)
• the referenced sections/*.md files

Optional (helps you stay in-scope while rewriting):

• outline/writer_context_packs.jsonl (allowed citations + opener_mode hints)

Output

Run this targeted pass after writer-selfloop and before opener-variator and logic polish. If major edits happened since the report, rerun writer-selfloop so ## Style Smells reflects current text. The deterministic script is a certification adapter: it refuses to create the marker while the latest report still names flagged files or predates any sections/*.md file. Perform the semantic rewrite through this Skill or the responsible upstream writer, rerun writer-selfloop, then retry the adapter. A passing marker records the certified Section-tree SHA256.

• Updated sections/*.md files (same filenames; still body-only; no headings)
• Re-running writer-selfloop is the audit trail (Style Smells should shrink).
• Create sections/style_harmonized.refined.ok (empty file) when you are done (pipeline contract signal; required if this unit is marked DONE).

Role cards (use explicitly)

Style Harmonizer (editor)

Mission: remove slot phrases and stem repetition while keeping meaning unchanged.

Do:

• rewrite the surface form (opener/closer/cadence), not the claim
• keep each paragraph content-bearing (argument bridge, not navigation)
• prefer small local edits over global style refactors

Avoid:

• adding new factual claims or new citations
• moving citations to different paragraphs or different subsections
• rewriting a thin section instead of routing upstream for more evidence

Evidence Steward (skeptic)

Mission: prevent style work from becoming content drift.

Do:

• after each rewrite, spot-check that every cited claim still matches the same sentence
• if you feel forced to add new material to make prose sound better, stop and route upstream

Common style smells and how to fix them

1) Count-based opener slots (Two limitations..., Three takeaways...)

Why it is high-signal: it creates a reusable sentence slot that repeats across H3s.

Rewrite moves (choose one):

• Integrate the caveat into a contrast paragraph (last sentence): state the boundary that changes interpretation.
• Use a single caveat sentence opener (no counting), but rotate across H3s (avoid repeating the same stem): "These results hinge on ..." / "Interpretation depends on ..." / "Evidence is thin when ..." / "A caveat is that ..." (use sparingly).
• If enumeration is truly needed, hide the count: use two coordinated clauses in one sentence, or vary the syntax (do not repeat across sections).

Mini example (paraphrase only):

• Bad: Two limitations stand out. First, ...
• Better: These results hinge on ...; this matters because it changes how results transfer across protocols.

2) Reused discourse stems (The key point is that ...)

Rewrite moves:

• Replace with one of: "A practical implication is that ...", "One takeaway is that ...", "A useful way to read these results is ...".
• Change cadence: split into a short claim sentence plus a follow-up sentence with the condition/why.

3) Same opener cadence across many H3s

Rewrite moves:

• Switch opener mode for the section (tension-first / decision-first / protocol-first / contrast-first).
• Replace generic connectors (Additionally/Moreover) with content-bearing pivots ("At the protocol level, ...", "Under budget constraints, ...").

4) Overview / narration openers ("This section provides an overview ...")

Why it is high-signal: it reads like a generated ToC narration rather than a paper argument.

Rewrite moves:

• Replace "overview" narration with a content-bearing lens: tension/decision/failure/protocol/contrast.
• Keep the first sentence falsifiable: name the constraint and why it matters (not what you are about to do).

Mini example (paraphrase only):

• Bad: This section provides an overview of tool interfaces for agents.
• Better: Tool interfaces define what actions are executable; interface contracts therefore determine which evaluation claims transfer across environments.

5) Paragraphs repeatedly starting with connector adverbs (Moreover, In addition, Therefore, Overall, As a result)

Why it is high-signal: the prose starts to sound mechanically stitched (each paragraph begins with the same connective), even when the content is solid.

Rewrite moves:

• Keep the logical relation, but move the connector into the sentence: start with the subject (e.g., "Tool catalogs also ..."), then add the relation mid-sentence ("..., which in turn ...").
• Use clause shapes instead of adverb openers: "While ... , ..." / "Although ... , ..." / "Because ... , ...".
• When summarizing, avoid "Overall," as a default label; state the conclusion directly as a claim sentence.

Mini example (paraphrase only):

• High-signal smell: "Moreover, ..." (repeated across multiple paragraphs)
• Better: start with the content noun phrase ("One implication is ..." / "A practical constraint is ..."), then express the relation inside the sentence.

6) Internal shorthand leaking into paper voice ("token(s)" used as a protocol noun)

Why it is high-signal: outside of NLP contexts (token budget/context window), "token" reads like internal shorthand. In this pipeline it often originates from packs/schemas and gets copied into prose, which makes the draft feel like an intermediate artifact.

Rewrite moves:

• Replace "X tokens" with reader-facing nouns: "X protocol details/assumptions/fields/parameters/dimensions".
• If you truly mean language-model tokens, keep it numeric and specific (e.g., "a 60k-token context window", "token budget"); avoid using "token" as a generic label for protocol metadata.
• Avoid "three tokens: ..." slots; either (a) state the conclusion directly, or (b) use "three reporting fields" and embed them naturally in the sentence.

Mini example (paraphrase only):

• Bad: Overall, self-improvement should be reported as a protocol with three explicit tokens: the feedback channel, the update rule, and the accounting rule.
• Better: Self-improvement results are easiest to compare when papers make three reporting fields explicit: the feedback channel, the update rule, and the accounting assumptions.

Workflow (minimal)

1. Read output/WRITER_SELFLOOP_TODO.md

• Find ## Style Smells and the file list.

2. Rewrite only the flagged files

• Prefer fixing the upstream writer skill that created the cadence problem rather than relying on a local regex cleanup.
• If a local rewrite is still necessary, make small edits: opener/closer stems, sentence shape, connector variety.
• Best-of-2 rewrite (recommended): for any paragraph you touch, draft 2 alternative phrasings and keep the one that (a) removes the slot stem, and (b) does not introduce a new repeated cadence across H3s.
• If needed, consult outline/writer_context_packs.jsonl for opener_mode hints and to stay citation-scope safe while rewriting.
• Do not touch citation keys.

3. Re-run writer-selfloop

• Expect: PASS remains PASS.
• Expect: Style Smells section is shorter (or disappears).

Done checklist

• The same slot phrase does not repeat across multiple H3s (especially count-based openers).
• No citation keys were added/removed/moved.
• writer-selfloop still reports PASS, and Style Smells shrinks.

Script

Quick Start

• uv run python .codex/skills/style-harmonizer/scripts/run.py --workspace <workspace>

All Options

• --workspace <dir> (required)
• --unit-id <U###>
• --inputs <semicolon-separated>
• --outputs <semicolon-separated>
• --checkpoint <C#>

Examples

• Harmonize style in a survey workspace:
  • uv run python .codex/skills/style-harmonizer/scripts/run.py --workspace workspaces/survey-llm-agents