# Evidence To Action Writer

> Use when drafting or revising Chinese long-form articles from PDFs, briefs, transcripts, Q&A, policies, course materials, workplace documents, or professional-practice sources, especially when the user asks for more practical detail, less preaching, preserved narrative structure, anonymization, or no fabricated experience.

- Skill: `razer-chan/evidence-to-action-writer` (Agent Skill, multi-file: 6 files)
- Install (CLI): `npx skillmds@latest add razer-chan/evidence-to-action-writer`
- Raw SKILL.md: https://api.skillmd.com/api/skills/razer-chan/evidence-to-action-writer/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: razer-chan (https://skillmd.com/u/razer-chan)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/razer-chan/evidence-to-action-writer

---


# Evidence to Action Writer

## Overview

Write `证据驱动的实务叙事长文`. Keep the reading experience conversational and progressive, but make every major judgment earn its place through evidence and a usable next action. Never impersonate a named writer.

## Required references

Read `references/style-profile.md` completely before drafting or revising.

When the material concerns Australian teacher education, secondary practicum, placement, mentor teachers, APST, classroom observation, lesson planning, school policy, or professional experience, also read `references/secondary-practicum.md` completely.

## Workflow

1. Inventory the supplied sources and the requested output shape. Preserve filenames, series structure, length, language, and formatting when the user specifies them. For anonymized work, also build a denylist of known institutions, people, course codes, platforms, locations, contacts, cohort labels, and local dates.
2. Build an evidence ledger before writing. Separate current official rules, practitioner advice, peer or Q&A material, and unresolved claims. Record who can confirm each conditional point.
3. Choose one narrative spine. Start from a concrete friction, question, document detail, or contradiction. Let each section reveal a deeper layer, then return to the opening image once at the end.
4. Draft in evidence-to-action units. Move from sourced detail to reader consequence, then to one concrete action, wording, sequence, fallback, or verification method. Add the applicability boundary where needed.
5. When revising, replace repeated abstract judgments with practical material. Preserve the useful opening and structure. Do not keep the preaching and append a checklist after it.
6. Run the validator, fix hard errors, read every warning, then perform the judgment pass in `references/style-profile.md`.

## Truth and identity boundary

- Use first-hand experience only when the user supplied or confirmed it.
- Never invent teaching, student response, mentor dialogue, outcomes, credentials, or personal history for vividness.
- When lived experience is absent, use source-based narration, attributed examples, or explicit conditional framing.
- Do not copy a named writer's identity, signature, contact details, biography, or fixed verbal tics. Transfer techniques, not persona.
- Keep a quoted or reconstructed example visibly distinct from source verbatim.

## Default article contract

Unless the user asks otherwise, write 4,000–8,000 Chinese characters, use one H1 title, short paragraphs, no H2/H3 headings, no bullet lists, no large bold blocks, no colon, em dash, or double quotation marks. Use `「」` for quotations. User-requested structure overrides these house defaults.

Use specific product, document, role, or process names when they are relevant and safe. Do not hide uncertainty behind confident prose.

## Validator

Run:

```bash
python3 "${CODEX_HOME:-$HOME/.codex}/skills/evidence-to-action-writer/scripts/check_article.py" \
  /absolute/path/article.md
```

For an anonymized secondary-practicum article, run:

```bash
python3 "${CODEX_HOME:-$HOME/.codex}/skills/evidence-to-action-writer/scripts/check_article.py" \
  /absolute/path/article.md --anonymize \
  --deny-term "Source University" --deny-term "COURSE1234"
```

Repeat `--deny-term` for every known source identifier; built-in patterns are a safety net, not proof of anonymity. Use `--json` for machine-readable findings. Use `--allow-h1-count`, `--allow-subheads`, `--allow-lists`, `--allow-formulaic-wording`, or `--allow-house-punctuation` only when the user explicitly requested those departures from the default house style.

For named-style work, pass one `--protected-writer-name "Name"` value per source writer. The option is repeatable. Treat every resulting mention warning as an attribution check and fix every impersonation error before delivery.

## Delivery gate

Deliver only when the source tiers remain distinct, no unsupported experience is implied, each core judgment has a practical consequence, repeated conclusions have been removed, the ending closes once, and the latest validator run reports no hard errors.