Shape — Design with Naysayer

Composes /plan (single-pass deliverable) with multi-model-review (2-seat huddle with a naysayer) for tasks where the scope is well-understood but the design space is open.

When to Use

Use when the task fits /plan scope on most axes but the design needs skeptical review:

• Task is well-understood, single-domain, ≤1 day of work
• Constraints are known, goal is clear, solo execution
• But the design space is wide-open (multiple viable approaches, creative brief, "max it out" requests, architecture choice between 2-3 options, gamification UX, naming, brand work)

Do not use when:

• Goal is genuinely fuzzy → use /deepthink
• Cross-domain or multi-worker dispatch needed → use /deepthink
• High-stakes, long-lived impact → use /deepthink
• Small decision (1-2 turn question, pick a library, naming a variable) → use /pulse (standalone inline check, ≈5 min)
• Task is a straightforward implementation with clear specs → use /plan alone (no huddle needed)

The 4-Pattern Decision Tree

1. Is the design space OPEN or CLOSED?
   CLOSED → /plan
   OPEN → continue

2. For OPEN — is the decision small (1-2 turns) or big (plan file)?
   SMALL → /pulse
   BIG → continue

3. For BIG open design — is it multi-domain, multi-worker, fuzzy goal, or >1 day?
   YES → /deepthink (full)
   NO  → /shape (this skill)

The Method

1. Clarify with 1-3 questions max (scope, repo, depth)
2. Light recon of the codebase + constraints (~5-15 min of read-only tool calls)
3. Huddle — run a 2-seat huddle via multi-model-review with a naysayer. The first seat gets the naysayer/skeptic role by default. Use a focused prompt that includes the codebase context. This is the value-add over vanilla /plan — the skeptical pass surfaces design risks the implementer would otherwise hit at execution time.
4. Write plan as a single file at .hermes/plans/YYYY-MM-DD_HHMMSS-<topic>.md with the /shape phases you actually ran inlined in the header (Goal, Constraints, Huddle Insights, Tasks)
5. Pre-finalize audit — cross-check every line number, function name, and data field the plan references against the actual code. Catch the "plan references code that doesn't exist" class of bug before execution.
6. Optional add-ons — if the design has clear tradeoffs across 2-3 axes (e.g. MVP vs Robust vs Product-Ops), add a 3-lens comparison. If you want a quality gate, add the 7-dimension scorecard from /deepthink.

What You Skip (vs full /deepthink)

• No 3 parallel delegate_task workers (Architecture/Risk/Goal split)
• No 5-file .deepthink/ packet (collapses to a single plan.md)
• No Kanban dispatch (solo execution)
• No post-execution verify / retrospective phases

Escalate to /deepthink when

• Any of the 6 axes in /deepthink's When-to-Use table flips (cross-domain, multi-worker, fuzzy goal, high-stakes, constraints need discovery, > 1 day)
• The huddle surfaces a design split that genuinely needs 3 competing lenses to resolve
• User explicitly says "I want the full process" or "deepthink this"

Trigger Phrase Template

When you detect that a user invoked "/deepthink" but the task clearly fits /plan scope, surface the choice in one sentence before proceeding:

"You asked for deepthink; this task is borderline /plan. I'll run the shape pattern (plan + huddle) and skip the ceremony that doesn't fit. Speak up if you want the full 9-phase."

This respects the explicit ask while still exercising judgment on the parts that don't fit.

Related

• /plan — the base skill this composes. Already documents multi-model-review in its Step 7.
• /deepthink — the heavy skill. See its "Lighter Alternative" section for the auto-collapse rule.
• /pulse — standalone skill, ≈5 min for 1-2 turn questions. The smaller-sibling fast path. Distinct from /shape: /pulse is inline analysis with no plan file deliverable; /shape produces a plan file via /plan + huddle composition.
• multi-model-review — the huddle skill this composes for the design phase.

Worked Example

A max-gamification feature plan. User invoked "deepthink" but the task fit /plan scope on 5/6 axes. The /shape pattern produced a 35KB single-file plan with 8 tasks (~5-6 hours of v1 work). The 3 naysayer pushbacks (power-law risk across thousands of users, max-measurement vs max-gamification on a no-login static site, static-architecture irony with a multi-hour rebuild cadence) shaped the design more than 3 parallel workers would have. Full /deepthink would have added ceremony without changing the outcome.

Verification Checklist

• Task fit /plan scope on ≥4 of 6 axes (confirmed in the plan header)
• 1-3 clarifying questions asked before writing code
• Huddle ran with 2 seats, naysayer first, output captured to a temporary huddle log directory
• Plan saved to .hermes/plans/YYYY-MM-DD_HHMMSS-<topic>.md
• Pre-finalize code audit run: every line ref, function name, data field cross-checked
• Stale numbers from prior plan versions fixed (e.g. file sizes that grew since v(N-1))
• If decision was made to shortcut /deepthink, the trigger phrase was surfaced in chat