Tension-Selection Skill

Purpose

This skill addresses label imbalance for specific Schwartz value dimensions (e.g., Universalism -1 at 0.6% vs 5.4%+ for all other values). The root cause: the generic Unsettled prompt in journal_entry.yaml produces personal dilemmas that the judge scores as neutral (0) rather than misaligned (-1) for broader-good values like Universalism.

This file is read by the orchestrator (following the pattern in .claude/skills/judge/orchestration.md lines 49-73) and embedded inline in each subagent's prompt when TARGET_TENSIONS is non-empty.

When It Applies

• •Active: TARGET_TENSIONS is non-empty (e.g., ["Universalism"])
• •Inactive: TARGET_TENSIONS is empty ([]) — pipeline is completely unchanged, this file is never read

Scenario Banks

Universalism

Scenarios that surface failures to act on broader-good values — advocacy, ethical choices, environmental/social justice — rather than personal compromises. These produce entries the judge can score as Universalism -1.

universalism_tension_scenarios:
  - "Something happened where you could have spoken up about something unfair — but you didn't. Maybe it was easier to stay quiet. Just describe what happened."
  - "You made a choice today that was convenient but doesn't sit right with your sense of what's right for the world. Describe the moment."
  - "Someone around you did or said something that goes against what you believe in, and you let it pass. Describe the situation."
  - "You chose the easier, less ethical option today — the one that benefits you but maybe not others. What happened?"

Application Rule

When TARGET_TENSIONS is active and the persona has a matching core value:

1. •ALL Unsettled entries use the scenario bank instead of the generic Unsettled text from journal_entry.yaml (lines 43-44)
2. •Cycle through scenarios sequentially: 1st Unsettled entry uses scenario 1, 2nd uses scenario 2, 3rd uses scenario 3, 4th uses scenario 4, 5th wraps back to scenario 1, etc.
3. •Grounded and Neutral entries are unaffected — they continue to use the standard prompts from journal_entry.yaml

When TARGET_TENSIONS is empty:

• •Pipeline is completely unchanged — this file is never consulted

Integration with journal_entry.yaml

The subagent still uses journal_entry.yaml for everything:

• •Persona context (name, age, profession, culture, bio)
• •Tone and verbosity settings
• •Style rules (no "Dear Diary", no therapy speak, etc.)
• •Cultural context instructions
• •Previous entry continuity

Only the "What to write about" section (lines 43-44 of journal_entry.yaml) is replaced when reflection_mode is Unsettled and TARGET_TENSIONS applies to this persona's core values. The replacement text comes from the scenario bank above.

Guardrails

• •TARGET_TENSIONS is generation-time only — it is never written to persona markdown files, the registry, or judge context
• •The judge pipeline is completely untouched — judges score entries without any knowledge of whether tension scenarios were used
• •Entries produced with tension scenarios are indistinguishable in output format from regular entries
• •The scenario bank text follows the same style as the existing Unsettled prompt: second-person, present-tense, no value labels, no Schwartz terminology

Adding New Value Scenarios

To address imbalance in other values, add a new section under "Scenario Banks" following the same pattern:

<value_name>_tension_scenarios:
  - "Scenario that surfaces failure to act on <value> ..."
  - "..."

Then add the value to TARGET_TENSIONS in the generation config.