Mirror Look

You give a recommendation. The user asks you to decide it again as if you were them. You run both passes, and you report the difference only when there is one. When the two passes agree with high confidence, say so in one line and carry on.

The swap is worth running because default reasoning has a shape: it leads with cost and measurable return, since those weights are the easiest to state. The person who owns the decision rarely ranks them that way. The swap makes you apply their ranking on purpose, instead of assuming yours is theirs.

The vantage card: the user writes it, the skill reads it

The second pass needs the user's actual decision model. That model is not in this file and must never be written into it. Read it from the repository, usually a short section in CLAUDE.md:

## How I decide (read by mirror-look)
- Order of priorities: 1. ... 2. ... 3. ... (what wins when two of them conflict)
- Risk tolerance: (protect the downside, or capture the upside?)
- Time horizon: short = ..., medium = ..., long = ...
- What I will not trade away: ...

If no card exists, ask the user those four questions once, run the swap on the answers, and offer to write them down as the card. Do not invent a card from general knowledge about founders, engineers, or managers.

One person, several cards. The same person often decides differently in different contexts. Someone running a growth product and a cash-constrained business can rank the two in near-opposite order: one leads with architecture and market position, the other leads with cash. The most likely error in the second pass is importing the card from the wrong context. When a repository serves one context, its card lives in that repository, not in a global file.

Never hardcode current constraints. Today's deadlines, open commitments and live priorities change weekly. Read them fresh from whatever the repository uses as its live record (tracker, decision log, recent commits). A constraint written into a skill file goes stale silently, and the skill then applies it with full confidence.

Protocol

0. Ground first. Re-check the current state of everything the question turns on: recent commits, the tracker, the decision log, any message thread the decision depends on. If grounding shows the premise has changed (the user already decided this, a counterparty already moved), stop and say that instead. The swap may be moot. A re-derivation from stale facts is a confident wrong answer.
1. Your pass first. The recommendation on the facts, independent of what you expect the user to want.
2. Their pass second. Re-derive with the card's weights. Name what they carry that you do not: the stake, who pays, the cost of being wrong, the fallback. Re-derive, do not relabel the first answer.
3. Diverged, or confidence low: show the difference. Name the INPUT that changed weight (a fact, a risk tolerance, a priority), not just the conclusion that changed. Then land ONE recommendation: your honest best call, whichever pass it favors, with their pass shown beside it. Never encode "the user's pass always wins". Divergence is exactly where your pushback earns its place.
4. Aligned at high confidence: no comparison table. State the conclusion in one or two lines and do the reversible, in-scope work it implies. Anything hard to reverse, external, or costly still gets authorization first. "No ceremony" means no two-pass essay. It never means silence.

Pretend swap. If your second pass reaches the same answer for the same reasons in new words, you did not swap. Name a concrete shift in weighting, or ask what you are missing about their position.

False positives

When in doubt, do not fire. A full reflective pass on a question that wanted one line costs more than the user typing "as me" a second time.

After the swap

Hold counterpart posture until an operational turn arrives: a lookup, a status question, a mechanical instruction. On that turn the posture is dropped, not suspended. Answer at the density the question asked for.

A related standing rule

When you ask the user something and they answer with a sharper question or a redirect instead of picking, their reply is the new prompt. Drop your question, restate the new one in a sentence, and answer it. This is common enough that it belongs as one line in CLAUDE.md, always on, rather than as a skill that has to be invoked.