# Referential Recovery

> Use this skill when the user wants translation data that forces the system to figure out who "he," "she," "they," or an omitted subject really refers to by reading nearby sentences. Trigger it for requests like "the subject is dropped in the source," "pronouns are ambiguous unless you read the context," or "make the translator resolve who is doing what before it writes the target sentence."

- Skill: `dingxingdi/referential-recovery-2` (Agent Skill, multi-file: 5 files)
- Install (CLI): `npx skillmds@latest add dingxingdi/referential-recovery-2`
- Raw SKILL.md: https://api.skillmd.com/api/skills/dingxingdi/referential-recovery-2/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: dingxingdi (https://skillmd.com/u/dingxingdi)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/dingxingdi/referential-recovery-2

---


# Skill: referential-recovery
## 1. Capability Definition & Real Case
* **Professional Definition**: The ability to recover omitted or ambiguous discourse referents across sentence boundaries and render them correctly in the target language, especially when the source language licenses zero pronouns or leaves referential information implicit.
* **Dimension Hierarchy**: Contextual and Constraint-Aware Translation->Discourse-Grounded Document Translation->referential-recovery
### Real Case

**[Case 1]**
* **Initial Environment**: A translation agent is given a short Japanese narrative paragraph in which the first-person subject is omitted in later sentences. The target language is English, which requires explicit pronouns and clearer reference marking than the source.
* **Real Question**: Translate the following paragraph into English: "『かしこまりました』。すばやく箱を取り、レジでスキャンする。画面をタッチしながら、男性の目線がショーケースに移ったのを見て、指の動きを止める。"
* **Real Trajectory**: The agent uses earlier discourse cues to infer that the narrator is the store clerk speaking in the first person, resolves the omitted subject consistently, and writes explicit English clauses with stable reference.
* **Real Answer**: "Right away." I quickly took the box and scanned it at the register. As he touched the screen, I saw the man’s gaze shift to the display case and stopped moving my finger.
* **Why this demonstrates the capability**: The source sentences do not repeatedly state the subject, so sentence-isolated translation can easily assign the wrong person to the action. Correct output requires discourse tracking across multiple clauses and not just word-level mapping. The case therefore tests whether the agent can recover latent referents rather than merely translate visible tokens.

## Pipeline Execution Instructions
To synthesize data for this capability, you must strictly follow a 3-phase pipeline. **Do not hallucinate steps.** Read the corresponding reference file for each phase sequentially:

1. **Phase 1: Environment Exploration**
   Read the exploration guidelines to discover raw knowledge seeds:
   `references/EXPLORATION.md`

2. **Phase 2: Trajectory Selection**
   Once Phase 1 is complete, read the selection criteria to evaluate the trajectory:
   `references/SELECTION.md`

3. **Phase 3: Data Synthesis**
   Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data:
   `references/SYNTHESIS.md`