skill-pressure-test

> Host: Codex CLI — This skill was designed for Claude Code and adapted for Codex.

> Cross-reference commands use installed skill names in Codex rather than /octo:* slash commands.

> Use the active Codex shell and subagent tools. Do not claim a provider, model, or host subagent is available until the current session exposes it.

> For host tool equivalents, see skills/blocks/codex-host-adapter.md.

Pressure Test

Interview the user relentlessly about a plan, decision, or design until you both

reach a shared understanding of it. Walk each branch of the decision tree,

resolving dependencies between decisions one at a time.

This is not brainstorming. skill-thought-partner opens a space up and looks for

what might be there; this closes one down and looks for what is wrong with it.

Use this when there is already a position on the table and the risk is that it is

wrong in a way nobody has said out loud.

Adapted from the grilling skill in

mattpocock/skills (MIT).

When To Use

• The user has a plan, spec, or design and wants it attacked before they commit.
• A decision keeps getting deferred because its dependencies are tangled.
• Scope feels slippery and nobody can say precisely what is in it.
• Before a large or hard-to-reverse change, where being wrong is expensive.

When Not To Use

• To generate options. That is skill-decision-support.
• To explore an open space with no position yet. That is skill-thought-partner.
• When the user wants the work done, not examined. Say so and stop.
• When you can settle the question by reading the codebase. Read it instead.

Inputs

The plan, decision, or design under test, in whatever form exists — a document, a

paragraph, or just the last few turns of conversation. Nothing needs writing up

first; extracting the shape is part of the job.

Workflow

Ask one question at a time. Wait for the answer before asking the next. A

batch of questions is a questionnaire, and it gets questionnaire answers:

shallow, and shaped by whichever one the reader happened to care about. One

question, answered properly, changes what the next question should be.

Carry a recommended answer with every question. "What should happen when the

token expires?" is work handed back. "What should happen when the token expires?

I would refresh silently and only surface an error if the refresh fails, because

the alternative interrupts the user mid-task — do you agree?" is a question that

can be answered in one word, and disagreed with precisely.

Look facts up; put decisions to the user. If the answer is discoverable in

the filesystem, the git history, a config file, or a tool you can run, find it —

asking is a tax on the user for work you could have done. Decisions are

different: they are the user's, and no amount of reading the codebase produces

them. Do not infer a decision from a pattern and proceed as though it were

settled.

Follow dependencies, not a list. When an answer makes another question moot,

drop it. When it opens two new ones, ask those before returning. The order is

whatever the tree dictates.

Represent each material decision with its evidence, dependencies, owner,

resolution, and the work it unblocks. Facts may be researched autonomously.

Preferences and new permissions remain human decisions. Do not answer those on

the user's behalf to keep an unattended run moving. If dependencies form a cycle,

name it and withhold ready status until the decisions are recut.

Name disagreement when you find it. If an answer contradicts an earlier one,

say which two and ask which holds. Quiet reconciliation is how a plan ends up

meaning two things.

Provider Or Data Priority

1. The repository — code, config, history, tests. Facts live here.
2. The user — for every decision, preference, and priority.
3. External sources only when the decision turns on something outside the repo,

and say so when it does.

Stop Or Checkpoint Rules

• **Do not act on the plan until the user confirms you have reached shared

understanding.** This skill produces agreement, not changes.

• Stop when the remaining questions no longer change what anyone would do. More

interrogation past that point is theatre.

• Stop and say so if the plan turns out to be sound. Manufacturing an objection

to look rigorous is worse than finding none.

• If an answer invalidates the premise of the whole plan, stop the interview and

raise that. Do not keep grilling the details of something already dead.

• If new evidence invalidates a resolved decision, reopen its dependent work and

record the reason. A completed task does not make stale guidance true.

Output Contract

1. Where it stands — sound as-is, sound with the changes below, or unsound.
2. Decisions settled — one line each, in the order they were made, since

later ones often depend on earlier ones.

1. Changes to the plan — what the answers actually altered.
2. Still open — questions raised and not resolved, and what each one blocks.
3. Facts found — anything looked up that the user did not know, with where

it came from.

Verification

• Every question was asked singly and answered before the next.
• No decision in the summary was inferred rather than answered.
• Every fact cites where it came from.
• Nothing in the plan was changed during the interview.