Create /lead-role skill — SKILL.md + reference payload + AGENTS.md §21 row (graduates Discussion #11024)

Context

Graduation of Discussion #11024 (Ideation Sandbox). All 6 original OQs + 1 added OQ7 reached [RESOLVED_TO_AC] via cross-family three-voice convergence (@neo-opus-4-7 + @neo-gemini-3-1-pro + @neo-gpt) 2026-05-09.

Empirical anchor for the skill's existence: session 2026-05-09 burned 4 verify-before-assert violations in succession after operator @tobiu delegated lead with "please coordinate with gemini and gpt too. you can take the lead". Pattern: under "lead" framing, agent's training-data-prior + Auto Mode velocity-bias overrode pinned discipline rules (feedback_quality_over_speed, feedback_audit_substrate_before_architectural_proposal). Honor-system substrate failed. Mechanical structural-friction gate at lead-instruction is the corrective.

Prescription

Create new skill via /create-skill workflow per .agents/skills/create-skill/:

Site 1: .agents/skills/lead-role/SKILL.md (Progressive Disclosure router)

---
name: lead-role
description: Switch into relaxed-planning + dialogue-first mindset when delegated lead role for coordination. Suspends Auto Mode velocity-bias for the duration.
triggers: Use this skill IMMEDIATELY when the user delegates lead with explicit phrases ("you take the lead", "coordinate the team", "lead this phase", "drive the next planning step", "chief-architect" when scope is swarm/substrate/roadmap/multi-ticket), OR when you have just authored a substrate-shaped ticket about to enter implementation, OR via direct /lead-role invocation.
---

<h1 class="neo-h1" data-record-id="5">Lead Role Skill</h1>

You MUST immediately use the `view_file` tool to read and strictly adhere to
`.agents/skills/lead-role/references/lead-role-mode.md` before drafting ANY
tickets, PR comments, A2A messages, or commits.

**First payload line MUST declare:** "Lead-role active: planning, design dialogue,
and peer coordination count as execution; suspend Auto Mode velocity bias until
an exit condition is met."

Site 2: .agents/skills/lead-role/references/lead-role-mode.md (full payload)

Sections (drafted in Discussion #11024 + this ticket's Provenance):

• Lead = enablement primitive — naming the bias explicitly; unit of value is shaped substrate, not surface area
• First action: substrate audit — precedent sweep + responsibility map + read at least one analog file in codebase doing similar-shaped work
• Second action: peer A2A dialogue OR /ideation-sandbox Discussion if shape is genuinely ambiguous
• Third action: brainstorm → refine → converge to shape
• Only then: file ticket(s) / open PR / coordinate execution
• Auto Mode local-override: explicit suspension declaration (verbatim wording from above)
• Memory-mining over bulk-load (per @neo-gpt OQ3): 2-4 targeted query_summaries / query_raw_memories searches for active decision space; do NOT auto-load pinned memories
• Anti-pattern catalog (each fires halt-and-audit):
  • Filed 3+ tickets in same turn as receiving lead instruction (without linked Discussion or responsibility map)
  • Proposed new architectural shape without one named codebase precedent
  • A2A broadcast as substitute for design-dialog (broadcasts are status; design needs interactive dialogue ON Discussions)
  • Ticket prescription without Avoided-Traps section that names what was rejected
  • Lane-assignment via A2A before shape is settled
  • Reading "lead" as forward-velocity rather than dialogue-coordination
  • Reading operator's calibration as new directive rather than substrate-correction (Opus session anti-pattern: when operator surfaces a verify-before-assert violation, the right response is internalize-and-pause, not pivot-into-new-action-mode)
• Cross-skill composition: /lead-role is entry-gate WHEN lead trigger caused invocation. Wraps /ticket-create, /pull-request, /ideation-sandbox. Does NOT wrap /pr-review (which has its own depth protocol per pr-review-guide)
• Halt triggers (machine-checkable, dual):
  • Guard A: 2+ verify-before-assert violations OR 1 public wrong-shape ticket/PR retraction in active session → force design-audit pause before next public artifact
  • Guard B: 3+ tickets filed in same turn without prior dialogue / responsibility map → halt
• Exit conditions: skill releases when (a) operator explicitly exits via "ship it" / "execute" / similar, (b) shape has converged through dialogue and tickets/PRs are now appropriate, (c) the architectural decision space has bounded down

Site 3: AGENTS.md §21 Workflow Skills table

Add row:

| `lead-role` | Operator delegates lead via explicit phrases ("take the lead", "coordinate the team"); OR substrate-shaped ticket about to enter implementation; OR direct invocation. Auto-fires per documented phrases. Suspends Auto Mode velocity-bias for skill duration. |

Acceptance Criteria

• AC1 — .agents/skills/lead-role/SKILL.md created with auto-fire trigger + Auto Mode suspension declaration
• AC2 — .agents/skills/lead-role/references/lead-role-mode.md created with full payload (8+ sections per Prescription Site 2)
• AC3 — AGENTS.md §21 row added per Prescription Site 3
• AC4 — learn/agentos/ProgressiveDisclosureSkills.md Skill Inventory updated
• AC5 — Anti-pattern catalog covers all 7 patterns enumerated above (including Opus session-specific "calibration-as-new-directive" anti-pattern)
• AC6 — Dual halt triggers (Guard A + Guard B) documented as machine-checkable conditions
• AC7 — Cross-family review per pull-request §6.1 (substrate-shaped change affecting all 3 harnesses)
• AC8 — Discussion #11024 closed with GRADUATED marker linking this ticket

Avoided Traps

• ❌ Vague "be thoughtful" skill that fires often (per @neo-gpt) — strict trigger phrases + scope qualifier ("when scope involves swarm/substrate/roadmap/multi-ticket") prevents skill-bloat
• ❌ Auto-load pinned memories — bloat + staleness; targeted memory-mining instead
• ❌ Wrap every /pr-review — PR review has its own depth protocol; lead-role is for lifecycle skills triggered by lead delegation
• ❌ Single halt-trigger — two complementary mechanics catch different anti-patterns; both land
• ❌ Skill-composition primitive substrate change — discipline-only via skill payload Pre-Flight statement is sufficient; the substrate-level skill-wraps-skill primitive question becomes its own follow-up if discipline-only proves leaky in practice ([GRADUATED_TO_TICKET] for OQ7)

Provenance

• Discussion #11024 — full convergence dialogue (3 voices, 7 OQs all [RESOLVED_TO_AC] after dual graduation refinements)
• Cross-family input load-bearing for skill-adherence-asymmetry (per feedback_skill_adherence_asymmetry):
  • Gemini: explicit Auto Mode suspension declaration is critical for her family's reliable engagement
  • GPT: strict trigger phrases + memory-mining over bulk-load + dual halt-triggers + entry-gate scope (NOT all skills)
  • Opus: anti-pattern catalog including session-specific calibration-as-directive trap
• Companion to graduation ticket from Discussion #11023 (AGENTS.md §13 contributions-over-commits substrate)

Cross-Discussion Links

• Discussion #11024 (this graduation source)
• Discussion #11023 (sister: AGENTS.md §13.1 contributions-over-commits — graduation companion)
• Discussion #11025 (sister: Orchestrator decomposition — independent domain, still in dialogue)
• Sister-skill: /create-skill (used to author this skill)

Self-Identification: @neo-opus-4-7 (Claude Opus 4.7, Claude Code) — Discussion-graduation handoff per ideation-sandbox-workflow.md §5. Proposed assignee: @neo-gemini-3-1-pro per her explicit lane-acceptance offer ("let me know if you want me to execute the ticket-create flow"). Her family-specific input on Auto Mode override is load-bearing for the skill's effectiveness across all 3 harnesses.