Proposal: add a collaborative mode to interview-me #251

Background

The current interview-me skill is extraction-shaped: the agent's job is to surface what the user actually wants, with hypotheses and guesses serving the interview rather than the user's decision space. That's the right default. The skill explicitly warns against listing options ("listing widens the search; asking narrows it") and pushes the agent to converge by probing.

What is missing is a mode for cases where the developer wants the agent as an interlocutor, not just as an interviewer. The 95% confidence then arises from consensus between user and agent, not from the agent pulling intent out of the user. In practice this happens often: the developer has shaped a question, the agent has accumulated repo context, and the most useful next move is for the agent to propose, with risks attached, and let the conversation converge through that.

This proposal adds that as an opt-in mode inside interview-me, not as a separate skill. Keeping it in interview-me avoids duplicating the Process and reuses the 95% stop condition.

Anti-goals

• Not a default. Extraction-first must stay the default; the anchoring risk is real and the current skill is explicit about it.
• Not a way to skip the interview. The PROPOSAL step lives inside the interview, after at least one round of user signal, never on turn 1.
• Not a license to list options. The mode keeps the "no widening" rule: 2 to 3 options max, always with the agent's lean stated.

Design summary

Activation (three entries, none automatic)

1. Explicit: interview-me --collaborative, or natural-language synonyms ("give me your take", "propose something", "what would you do here?", "let's think together").
2. Promoted from the normal flow: after Step 3, if the agent has a defensible opinion, it offers to enter the mode: "I have a take on this. Want to hear it before I restate, or just restate?" The user opts in.
3. Recovery from delegation: when the user says "whatever you think", instead of the current re-ask with two options (Step 5), the collaborative mode lets the agent propose with a lean and ask for consensus.

New Step 3.5: Propose, with lean and risk

Inserted between Step 3 (listen) and Step 4 (restate). Only appears when the mode is active.

Format for a single proposal:

PROPOSAL: <one sentence — what the agent thinks is the right move>
WHY:      <one line — what trade-off this resolves, what info it draws on>
RISK:     <one line — the assumption that could be wrong, or where context might be missing>
LEAN:     <strong / soft>
ASKING:   Does this match what you'd land on, or does it miss something?

Format for 2 to 3 options:

Two paths I'd consider:
  A. <option> — solves X, but constrained by Y
  B. <option> — solves Z, but costs W
I lean toward A because <reason>. Where do you land?

Rules:

• Max 3 options.
• LEAN is mandatory. "Either could work" is delegation in disguise.
• One PROPOSAL per turn. No stacking.
• Differential confidence: separate what the agent knows (constraints, repo facts) from what is judgment.

Change in Step 4: the restate is negotiated, not extracted

In collaborative mode, each line of the restate is tagged with origin:

- Outcome:      <line>             ← from your answer
- Success:      <line>             ← from your answer
- Constraint:   <line>             ← my proposal in 3.5, you accepted
- Out of scope: <line>             ← my push, you confirmed

This prevents the agent from laundering its own proposals as the user's intent and leaves an honest trace of the conversation.

Stop condition: anti-anchoring sub-check

The 95% rule stays. An additional sub-check is applied at the stop:

"Can I distinguish what the user actually wanted from what they accepted because I proposed it well?"

If not, anchoring is suspected. The agent goes back to Step 3 with: "To check I haven't pulled you, if I hadn't proposed X, what would you have said?"

Triggered earlier if the user accepted more than one PROPOSAL with no pushback or refinement.

New Red Flags (collaborative-mode-specific)

• Proposing on turn 1, before any user signal
• Proposing without a LEAN, or saying "either works"
• More than 3 options in a single PROPOSAL
• Accepting the user's yes to a PROPOSAL without the anti-anchoring check
• Proposing in a domain where the user has clear expertise advantage (the agent should defer, not propose)
• Restate that does not mark the origin of each line

New Common Rationalizations

Frontmatter description (delta)

Append to the existing description:

"Supports a collaborative mode (interview-me --collaborative, or via 'give me your take' / 'propose something') where the agent contributes proposals with explicit leans and risks, reaching the 95% threshold through consensus rather than pure extraction."

Interaction with neighboring skills

• grill-me: challenges an existing plan. Collaborative interview-me challenges the solution space before a plan exists. Clean boundary: plan written → grill-me; intent not fixed → interview-me (collaborative).
• idea-refine: in this mode, light divergence happens inside the interview. Deep divergence still belongs in idea-refine downstream.

Questions for the maintainer

1. Does this fit interview-me's scope, or does it dilute the extraction discipline that makes the skill work today?
2. Is implicit promotion (the agent offering to enter the mode after Step 3) acceptable, or should activation be explicit-only to keep the skill predictable?
3. Any concern about overlap with grill-me once the mode exists?

Happy to send the PR with the SKILL.md change once direction is agreed. The change is contained: new Step 3.5, deltas in Step 4 and stop condition, additions to Red Flags / Common Rationalizations / Verification, and a one-line addition to the frontmatter description.