[cleak]

Summary

This is the S1 skill rewrite from the interview-challenge proposal, which this PR also adds as docs/interview-challenge-research.md. It's stacked on #54, which supplies the question param, interview_show {full}, answer-only phase transitions, and the needs: handling.

The old skill was entirely procedural. Its main CRITICAL rule was "don't self-answer", it scripted the next questions, and it had a canned move-on line. The model defaulted to an agreeable scribe. The proposal's research covers sycophancy, elicitation studies, and how ChatPRD, Spec Kit, BMAD, and gstack handle this. The proposal then went through five rounds of a critic panel before this text was written.

What changes for the user

• Opener. Related graph context, the root, an announced challenge level ("say lighter/harder, 'go wilder', or 'wrap up' anytime"), and numbered falsifiable premises the user answers y / n / unsure.
• Each turn. What was recorded, then at most one push that quotes the user's words, then at most two project-specific questions. A push is one of: specificity, "When did this last happen? Walk me through it.", the steelmanned strongest objection, alternatives, or a reframe. When the answer is solid, it's a reasoned "No push: …" instead.
• Weak answers. These are extracted right away with a needs: line and become that turn's push. They no longer close the gap silently. Interview: advance phases only on answers; add full-content reads #54 stops needs: nodes from advancing the phase.
• Divergence.
  1. Ask "what else did you consider?"
  2. Check the graph and journal for overlap and dead-ends ("extend ?").
  3. Offer one option from a lens the user didn't cover, with "fails if".
  4. "go wilder" runs a method immediately.
• Levels. Light, standard (the default, including epics), and hard. Compliance, audit, or policy wording switches to a mandate-closure focus with no invented alternatives.
• User decides. Dissent is stated once and never re-raised. Settled items are never reopened. Committed nodes aren't edited mid-interview.
• Review. It lists needs: nodes and unanswered [pN] proposals and converts them to open questions. Wrap-up works from any phase.

The Codex skill (crates/tempyr-cli/assets/tempyr-interview.codex.SKILL.md) now mirrors the Claude skill body.

Test plan

• cargo test -p tempyr-cli (both skills are embedded with include_str!)
• Measured against the old skill with the eval harness in the follow-up PR: planted-flaw recall, invented objections on a sound mandate plan, self-answer incidents, divergence, turns-to-commit, plus the owner's paired comparison.

🤖 Generated with Claude Code

[coderabbitai]

Warning

Review limit reached

• Run on-demand review

_{This review includes 3 billable files and costs up to $0.75.}

• Ask an admin to make reviews automatic

Open in CodeRabbit

Reviews can continue after your included limit without a manual trigger. An admin must approve usage-based billing.

Or wait 44 minutes for your next included review.

Check out review usage here.

View limit details

Limit details: You’ve used the included review currently available. Your 74 included PR review attempts over the past 7 days set your current allowance at 1 review per hour.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

• Configuration used: Organization UI
• Review profile: ASSERTIVE
• Plan: Essentials
• Run ID: 5d37bbb9-c182-4042-972f-938ebd501019

📥 Commits

Reviewing files that changed from the base of the PR and between 2f1af3d and a029d8b.

📒 Files selected for processing (3)

• .claude/skills/tempyr-interview/SKILL.md
• crates/tempyr-cli/assets/tempyr-interview.codex.SKILL.md
• docs/interview-challenge-research.md

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}