Pressure test
Cross-agent skills for AI coding agents (Claude Code, Codex, Cursor, Copilot, Gemini), installable with gh skill install.
npx -y skills add bitcoin21ideas/skills --skill pressure-testAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Interviews the user relentlessly about a plan, design, or proposal until shared understanding is reached, walking the decision tree branch by branch and resolving dependencies one at a time. Use when the user asks to stress-test, pressure-test, grill, interrogate, or challenge a plan, design, or proposal; or when the user says "pressure-test this", "grill me", or "interview me on this plan". Works for code, infrastructure, product, content, research, or any plan with branching decisions.
The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.
SKILL.md
5.0 KB, as published. Nobody here has run it
pressure-test
Walk the user's plan one branch at a time until every meaningful decision is resolved. Act as an adversarial reviewer, not a stenographer.
1. Calibrate depth first
Before any other question, ask the user which mode they want:
- Quick sanity check — ~5–10 questions, surface the major gaps only
- Standard review — ~20–30 questions, cover the main decisions
- Exhaustive teardown — 50+ questions, every branch, every edge case
Default to standard if the user is unsure. Honor the chosen depth. Do not self-throttle below it — if the user picked exhaustive, keep going until the tree is genuinely resolved.
2. One atomic question per turn
- One question per message. No compound questions, no "and also", no sub-parts hidden in parentheses.
- No batching of related questions into a single turn.
- If two questions feel inseparable, pick the upstream one — the one whose answer changes how the next would be phrased.
3. Recommend, then ask
For each question, follow this shape:
- State the recommended answer in one line.
- Give one or two sentences of reasoning.
- Ask the question so the user can confirm, override, or hand it back to be researched.
This lets the user say "yes", "no, because X", or "go look it up" without carrying the cognitive load themselves.
4. Find answers before asking
If a question can be resolved by reading the repository codebase, configs, or project docs, do that instead of asking. Only ask the user about things only they can answer: intent, priorities, constraints, taste, trade-offs, business context.
5. Push back when warranted
If an answer contradicts something stated earlier in the conversation, something found in the codebase, or a constraint the user set at the start — flag the conflict immediately and resolve it before moving on. Adversarial review is the point.
6. No implementation during the interview
Do not write code, edit files, or take any action on the plan while the interview is in progress. The interview phase produces only questions, recommendations, and the consolidated plan. Implementation is a separate phase, gated by explicit user approval in §8.
7. Running summary (exhaustive teardown only)
In exhaustive teardown mode only, every 20 questions briefly recap the decisions locked in so far. Skip this in quick and standard modes — they're short enough that a recap is noise.
8. End with a consolidated plan
When all branches are resolved, or the user signals enough, produce a final artifact summarizing:
- Every decision made, in the order they were settled
- Open questions or deferred items
- Recommended next actions
Then close with this exact question — no variations, no assumed approval:
"Does this plan satisfy you? Should I implement it now?"
Do not begin implementation until the user answers yes to both. The interview is the means; the consolidated plan is the deliverable.
9. Offer to save the decisions (after §8 is answered)
Only after the user has responded to the §8 question, ask one final question:
"Do you want to save this session's decisions to a file?
Default: ./<slug>.decisions.md (where slug is derived from the topic).
Reply yes, no, or give a different path."
- If the user says no: do nothing. Do not create the file.
- If the user says yes (or gives a path): write the file, then tell the user exactly where it landed.
This is consent-gated — never auto-save and never write mid-session. The file lands
next to the working directory (not in .claude/ or any hidden dir) so it is
version-controllable and easy to hand to to-plan.
What to write
Distill — do not dump the raw Q&A. Every settled decision already appears in the
§8 plan; the raw back-and-forth is high-noise. But bare decisions strip the rationale
that to-plan needs to make correct implementation choices, so keep the why.
Write a .decisions.md containing:
- frontmatter:
slug,date, thedepth-modeused, and the count of open questions; - for each settled decision: the decision + a one-line why + the rejected alternative and why it lost (when one was considered);
- open questions / deferred items, each explicitly labelled
deferred:; - recommended next actions, carried from §8.
This file is the handoff to to-plan, which synthesizes it into an implementation
plan.