Skill: multi-turn-revision-consistency | Skills Pool
技能檔案
Skill: multi-turn-revision-consistency
Use this skill when the user keeps refining the same document over multiple turns. Trigger it for requests like “now shorten the intro,” “keep the new tone but add one caveat,” or “revise the second paragraph without undoing the previous changes.” It is meant for iterative editing where each turn must build on the current state rather than restart from scratch.
Dingxingdi0 星標2026年4月8日
職業
分類
代碼質素
技能內容
1. Capability Definition & Real Case
Professional Definition:The capability to preserve and integrate previously accepted document edits across successive user instructions, applying new revisions without undoing, contradicting, or silently degrading earlier modifications unless explicitly directed to do so.
Dimension Hierarchy: Revision Fidelity->Direct Text Modification->multi-turn-revision-consistency
Real Case
[Case 1]
Initial Environment: The agent is editing a project update email. In turn one, the user asks for a more formal tone. In turn two, the user asks to shorten the introduction. In turn three, the user asks to add a single sentence about delivery risk while preserving the earlier changes.
Real Question: Revise this project update in three passes: first make it more formal, then shorten the introduction, then add one sentence about schedule risk without undoing the previous edits.
Real Trajectory: The agent creates a revision ledger after each turn, updates the working draft incrementally, checks for conflicts between the new request and prior accepted changes, and applies only the new delta at each pass.
相關技能
Real Answer: A final email that is still formal, now has a shorter introduction, and includes one risk sentence without losing the prior tone change.
Why this demonstrates the capability: The central difficulty is cumulative state management. A weak model often reverts tone, re-expands the intro, or rewrites the whole draft from scratch. This case therefore isolates whether the agent can carry forward accepted edits across multiple instructions.
[Case 2]
Initial Environment: The agent is revising a short encyclopedia section. One turn changes the timeline wording, a second turn converts bullet points into prose, and a third turn requests that the concluding sentence stay untouched.
Real Question: Update the timeline dates, rewrite the bullets as prose, and keep the final sentence exactly as it is.
Real Trajectory: The agent stores which edits are already committed, marks the final sentence as locked, merges the prose-conversion request with the timeline update, and verifies that the locked sentence survives unchanged.
Real Answer: A prose paragraph with updated dates and an unchanged final sentence.
Why this demonstrates the capability: The requests operate on overlapping but not identical regions, so careless regeneration easily violates one of them. The agent must integrate compatible edits while respecting frozen text. That makes the case a direct test of iterative revision control rather than single-shot editing.
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:
Phase 1: Environment Exploration
Read the exploration guidelines to discover raw knowledge seeds:
references/EXPLORATION.md
Phase 2: Trajectory Selection
Once Phase 1 is complete, read the selection criteria to evaluate the trajectory:
references/SELECTION.md
Phase 3: Data Synthesis
Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data:
references/SYNTHESIS.md