From 1a2cfea593dfa72255beb05077ce8315c6ab3bad Mon Sep 17 00:00:00 2001 From: citizen204 Date: Fri, 5 Jun 2026 22:40:00 +0930 Subject: [PATCH] fix(writing-skills): replace count anchors with standard-based language Replace '3+ combined pressures' count targets with qualitative standards throughout SKILL.md and testing-skills-with-subagents.md (5 sites). The count was doing gating work the actual standard could carry alone: 'scenarios that reliably trigger the rationalizations you want to prevent.' A scenario with two strong, well-aligned pressures can trigger the rationalization more reliably than four weak unaligned ones. Closes #1487 --- skills/writing-skills/SKILL.md | 2 +- skills/writing-skills/testing-skills-with-subagents.md | 10 +++++----- 2 files changed, 6 insertions(+), 6 deletions(-) diff --git a/skills/writing-skills/SKILL.md b/skills/writing-skills/SKILL.md index d7393dc790..b818fc4987 100644 --- a/skills/writing-skills/SKILL.md +++ b/skills/writing-skills/SKILL.md @@ -598,7 +598,7 @@ Deploying untested skills = deploying untested code. It's a violation of quality **IMPORTANT: Create a todo for EACH checklist item below.** **RED Phase - Write Failing Test:** -- [ ] Create pressure scenarios (3+ combined pressures for discipline skills) +- [ ] Create pressure scenarios that reliably trigger the rationalizations you want to prevent (combine pressures that make the rationalization feel genuinely compelling — the test for "enough" is whether an agent without the skill consistently takes the rationalization path, not a count) - [ ] Run scenarios WITHOUT skill - document baseline behavior verbatim - [ ] Identify patterns in rationalizations/failures diff --git a/skills/writing-skills/testing-skills-with-subagents.md b/skills/writing-skills/testing-skills-with-subagents.md index a5acfeac86..fb17482616 100644 --- a/skills/writing-skills/testing-skills-with-subagents.md +++ b/skills/writing-skills/testing-skills-with-subagents.md @@ -48,7 +48,7 @@ This is identical to TDD's "write failing test first" - you MUST see what agents **Process:** -- [ ] **Create pressure scenarios** (3+ combined pressures) +- [ ] **Create pressure scenarios** that reliably trigger the rationalizations you want to prevent (combine pressures that make the rationalization feel genuinely compelling) - [ ] **Run WITHOUT skill** - give agents realistic task with pressures - [ ] **Document choices and rationalizations** word-for-word - [ ] **Identify patterns** - which excuses appear repeatedly? @@ -137,7 +137,7 @@ Forces explicit choice. | **Social** | Looking dogmatic, seeming inflexible | | **Pragmatic** | "Being pragmatic vs dogmatic" | -**Best tests combine 3+ pressures.** +**Best tests combine pressures that make the rationalization feel genuinely compelling.** The diagnostic: does an agent without the skill, run on this scenario, consistently take the rationalization path you're trying to close? A scenario with two strong, well-aligned pressures can do this more reliably than four weak unaligned ones. **Why this works:** See persuasion-principles.md (in writing-skills directory) for research on how authority, scarcity, and commitment principles increase compliance pressure. @@ -310,7 +310,7 @@ Meta-test: "Skill was clear, I should follow it" Before deploying skill, verify you followed RED-GREEN-REFACTOR: **RED Phase:** -- [ ] Created pressure scenarios (3+ combined pressures) +- [ ] Created pressure scenarios that reliably trigger the rationalizations you want to prevent - [ ] Ran scenarios WITHOUT skill (baseline) - [ ] Documented agent failures and rationalizations verbatim @@ -340,8 +340,8 @@ Running only academic tests, not real pressure scenarios. ✅ Fix: Use pressure scenarios that make agent WANT to violate. **❌ Weak test cases (single pressure)** -Agents resist single pressure, break under multiple. -✅ Fix: Combine 3+ pressures (time + sunk cost + exhaustion). +Agents resist single pressure, break under combined pressures. +✅ Fix: Combine pressures that make the rationalization feel genuinely compelling — the standard is whether an agent without the skill consistently takes the rationalization path, not a count. **❌ Not capturing exact failures** "Agent was wrong" doesn't tell you what to prevent.