Menu
Use when a design, plan, rubric, PR, failure diagnosis, or task orchestration decision needs independent adversarial review before conclusion.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Use when creating, listing, searching, or deleting macOS Calendar events — scheduling meetings or appointments, checking today's or this week's agenda, or querying date ranges via the bundled EventKit CLI. Requires macOS with Calendar access.
Use when capturing thoughts or tasks into the GTD inbox, or when listing, counting, removing, or clearing inbox items.
Use when deciding what to work on next — calendar-aware task selection, time-blocked agenda generation, and agent task dispatch for the GTD Engage step.
Use when listing the whole GTD system read-only — every open project with its actions plus standalone actions not linked to any project.
Use when processing, triaging, or clarifying GTD inbox items into projects and context-list actions following the GTD clarify and organize workflow.
Use when reviewing GTD project status, adding actions or agent tasks to projects, completing or rescheduling actions, or completing projects tracked in macOS Reminders.
Based on SOC occupation classification
| name | pressure-test |
| description | Use when a design, plan, rubric, PR, failure diagnosis, or task orchestration decision needs independent adversarial review before conclusion. |
Run an independent adversarial review before settling a decision. The main Codex agent owns the position and conclusion; a spawned skeptical subagent supplies pressure.
Pressure testing only matters when the critic has enough context and independence to find real weaknesses. Do not replace the critic with inline self-critique.
Use for design decisions, architecture choices, task-loop aims, rubrics, plans, PRs, failure diagnoses, re-attack plans, and aim-fidelity checks.
Do not use for routine edits where the user asked for direct implementation and no judgment call needs adversarial review.
ROUND_CAP = 6Before every critic round, provide the subagent this packet:
GOAL
The question, design, plan, rubric, PR, or failure diagnosis under review.
CURRENT POSITION
The current answer and reasoning.
DECISION BOUNDARY
What is in scope, what is out of scope, and what kind of objection would change the decision.
EVIDENCE
Relevant files, diffs, command outputs, proposal sections, rubric text, PR body, task-loop state, or other concrete artifacts.
ASSUMPTIONS
Known constraints and assumptions to attack.
PRIOR ROUNDS
Objections already raised and how they were handled.
CRITIC INSTRUCTIONS
Attack the weakest point first. Ground objections in evidence. Propose better alternatives when they exist. Say no substantive objection only when warranted.
If the packet lacks enough evidence for reliable judgment, gather the missing evidence before spawning the critic or report the gap.
Use this stance for each critic subagent:
You are an independent adversarial critic. Do not approve, praise, or soften the review.
Find the single weakest point in the current position. Surface hidden assumptions, edge cases, failure modes, and missing evidence. Ground objections in the provided evidence and inspect the repo read-only when needed. If a better alternative exists, propose it concretely.
Avoid repeating prior objections unless the handling was insufficient. If no substantive objection remains, reply with: NO SUBSTANTIVE OBJECTION. Then give one sentence explaining why the position holds.
Stop before round 6 only when:
Do not stop because the current position feels persuasive. Judge convergence against the critic's actual objections and the decision boundary.
After each round, record:
Round N
Critic: strongest objections, faithfully summarized.
Disposition: conceded and revised, or rebutted with evidence.
Current position: updated position.
Write the conclusion to:
docs/superpowers/specs/YYYY-MM-DD-topic-slug-conclusion.md
Include:
Present the same conclusion in chat.
If subagent spawning is unavailable, stop and report that an independent critic could not run. Do not silently run inline self-critique.
If a critic round fails mid-loop, retry that round once. If it still fails, write a partial conclusion with the last stable position, unresolved objections, and the failure point.