01 / THE APPROACHBACK TO INDEX

THE MAKING OF A CONSIDERED CONVERSATION

SCROLL TO UNZIP

01 / 06   —   START WITH THE PERSON

Defining
the brief.

Help an advisor find the right words. Keep the final choice human.

  • One client and one current intention.
  • Three Chinese WeChat drafts, across six scenarios.
  • Fictional facts. Clear boundaries. No automatic sending.
Explore this chapter
THE OUTPUT CONTRACT
03CHINESE
WECHAT DRAFTS
  • 01Birthday and festive greetings
  • 02Product recommendation
  • 03Post-purchase care
  • 04Reactivation
  • 05Event invitation
  • 06Price objection

SIX SCENARIOS. ONE CLIENT AT A TIME.

02 / 06   —   OBSERVE. DERIVE. TEST.

From research
to rules.

Research defines the risks. Fictional cases make them testable.

  • Client journey and CRM shape the communication need.
  • Twenty principles connect facts, service and WeChat tone.
  • Research supplies a lens, never a real client's record.
Explore this chapter
Fictional design workbench with blank paper and leather samples
20 WORKING PRINCIPLES

Research.
Rules.
Real questions.

FICTIONAL CASES / SUPPLIED FACTS

03 / 06   —   A HOUSE WITH A POINT OF VIEW

Building
PUNTO.

A fictional leather house, built around use, material and continuing care.

  • A fictional leather house, with ten supplied products.
  • Use, patina and repair guide the brand voice.
  • A segment profile is a reference, not a personal fact.
Explore this chapter
An original fictional PUNTO leather concept on stone
USE. PATINA. REPAIR.FICTIONAL EDITORIAL CONCEPT

04 / 06   —   FROM CONTEXT TO A DRAFT

Connecting
the pieces.

Keep the orchestration explicit, and every attempt traceable.

  • Route the current task, then load its skill.
  • Generate with supplied context; check and repair once.
  • One shared service for the API, CLI and test batches.
Explore this chapter

05 / 06   —   MEASURE WHAT ACTUALLY HAPPENED

Putting it
to the test.

A completed generation is not the same thing as a good message.

  • Fifty fixed cases cover all six scenarios.
  • Rule checks, AI reviews and human judgement stay separate.
  • 150/150 drafts have imported human ratings.
Explore this chapter
SELECTED BATCH / FIXED SCENARIOS
150DRAFTS
REVIEWED BY AI
GOOD50
NEEDS EDITING61
AI-FLAGGED39

150 HUMAN RATINGS IMPORTED. AI VERDICTS ARE NOT HUMAN CONCLUSIONS.

06 / 06   —   TURN REVIEW INTO BETTER RULES

Reviewing
and refining.

Human feedback shapes the changes. New outputs still need a fresh review.

  • 150 drafts reviewed; 30 judged not acceptable.
  • Update the rules for relevance, tone and pricing.
  • Separate implemented changes from verified quality.
Explore this chapter
FROM HUMAN REVIEW TO REFINEMENT

150 SAVED DRAFTS / HUMAN REVIEW

30 need a rethink.

01 / OBSERVE

Relevance. Tone. Facts.

02 / REFINE

Clearer response rules.

v7.8 IMPLEMENTED. NEW OUTPUT QUALITY REVIEW PENDING.