New Capability: Hone -- Skill & Command Architect

New Capability: Hone -- Skill & Command Architect

Date: April 12, 2026 Origin: Chris Carolan identified structural decay across the skill, command, and script layers. An audit revealed 14 broken agent references, 7 commands missing enforcement, 48 legacy commands with no guardrails, and 5 documentation sources disagreeing on agent count. No agent owned the structural integrity between these layers. Chris asked V and Aegis to design and hire one. Impact: The organization now has an agent whose sole job is ensuring every cross-layer reference resolves, every command loads enforcement, and every documentation count matches filesystem reality -- across 89 agent definitions, 110 commands, and 202 skills.


What Was Built

Hone is a new agent in V's Operations Org (Self-Improvement group) that owns the structural integrity of the skill, command, and script layers. Hone does not author new methodology, write command business logic, or maintain agent definitions. Hone validates the connections between what exists and flags where the wiring is broken.

The system had grown to 89 agent definitions, 110 commands (62 current + 48 legacy), and 202 skills with no agent responsible for verifying that cross-references between these layers actually resolve. Skills were renamed without updating the commands that loaded them. Agent definitions pointed to implementation directories that did not exist. Commands were created without enforcement loading. Legacy commands remained discoverable alongside current ones. Five separate documentation sources each claimed a different agent count. The problem was invisible precisely because it accumulated one reference at a time -- each change working in isolation but collectively creating a brittle surface that broke unpredictably.

The name is deliberate. A hone sharpens the edge of an existing tool. This agent does not build new tools -- it keeps the existing ones precise.

Infrastructure Changes

Change Before After
Cross-layer reference validation Manual. No agent owned it. Broken references discovered only when they caused runtime failures. Hone validates agent-to-implementation, command-to-skill, command-to-script, and skill-to-script references on demand and weekly.
Enforcement coverage 7 commands (including client-facing) had no enforcement loading. No one tracked coverage. Hone audits every command for enforcement presence and categorizes as full, partial, or none. Client-facing gaps flagged as high risk.
Documentation count accuracy 5 sources (Dewey index, INDEX.md header, INDEX.md table rows, .claude/agents/ file count, MEMORY.md) disagreed. Hone reconciles counts across all sources and flags every mismatch with file, line number, and claimed vs. actual count.
Legacy command disposition 48 legacy commands coexisted with current commands. No guardrail. Accidental invocation possible. Hone owns disposition: archive, merge into current command, or retain with deprecation warning.
Dependency mapping Did not exist. "If I change this skill, what breaks?" required manual grep. Hone builds reference maps: which commands load which skills, which commands spawn which agents, which skills reference which scripts.
Script pattern deduplication 5 patterns copy-pasted across 10-52 commands each (datetime, BU health, calendar, social schedule, Sanity query). Hone identifies repeated blocks and proposes shared references to reduce copy-paste drift.
Delegation manifest compliance 47 of 50 current commands ran procedural logic directly. Only 3 prototypes followed the delegation pattern. Hone validates that every command follows the delegation manifest pattern (Team table, named agent spawning, no embedded procedure).

Implementation

File Purpose
.claude/agents/hone.md Agent definition. Ownership, scope, team interactions, delegation contract, invocation triggers, enforcement rules.
agents/skill-command-architect/AGENT.md Implementation. Triggers, workflow (6-step validation process), output format (Layer Consistency Report), run commands for manual verification.
/mnt/d/Leadership/assessments/job-descriptions/hone-skill-command-architect.md Job description. Identity, role type, KVIs, team interactions, current state assessment.

Registration Points

System Record
Agent definition .claude/agents/hone.md
Implementation agents/skill-command-architect/AGENT.md
AI Leadership Roster skills/ai-leadership/INDEX.md (V Ops, Self-Improvement group, 86 total)
Dewey Index .claude/codebase-index.md (sections 330 + 390 updated)
MEMORY.md Agent count updated to 86
Job Description /mnt/d/Leadership/assessments/job-descriptions/hone-skill-command-architect.md
Org Chart /mnt/d/Leadership/assessments/2026-03-15-ai-org-chart.html
Website Data apps/website/src/data/ai-team-structure.ts
Agent Office DB data/agent-office.db
HubSpot Contact 215190664854 (hone@valuefirstteam.com, agent_readiness_level: 4_fully_operational)

Usage

Invocation

Hone is spawned via the Agent tool with subagent_type: "hone".

# Full cross-layer audit
"Check layer consistency"

# Specific dependency check after modifying an agent
"Run dependency audit on sentinel"

# Validate cross-references after creating a new skill
"Validate cross-references"

# Weekly cadence (part of operational rhythm)
"Skill health check"

# After discovering a documentation count mismatch
"Command audit"

What Hone Returns

A Layer Consistency Report containing:

Section Content
Broken References Source file, expected path, status (MISSING/STALE)
Count Mismatches Source, claimed count, actual count, delta
Commands Missing Enforcement Command name, type, risk level
Legacy Command Disposition Command, recommendation (archive/merge/retain), reason
Pattern Deduplication Repeated block, command count, proposed extraction
Delegation Manifest Compliance Commands without Team tables, commands with embedded procedure

Validation Commands (manual)

# Count agent definitions
ls .claude/agents/*.md | wc -l

# Count agent implementations
ls -d agents/*/AGENT.md 2>/dev/null | wc -l

# Find agent defs referencing non-existent AGENT.md paths
for f in .claude/agents/*.md; do
  dir=$(grep -oP 'agents/[a-z-]+/AGENT\.md' "$f" | head -1 | sed 's|/AGENT.md||')
  if [ -n "$dir" ] && [ ! -d "$dir" ]; then
    echo "BROKEN: $f -> $dir"
  fi
done

# Find commands without enforcement sections
for f in .claude/commands/*.md; do
  if ! grep -qi "enforcement\|vf-platform-context\|vf-self-correction" "$f"; then
    echo "NO ENFORCEMENT: $f"
  fi
done

Baseline Audit Findings (Day 1 Work Queue)

# Finding Severity Status
1 14 agent definitions reference non-existent AGENT.md directories Medium Logged
2 3 broken skill references (provost-context.md, marquee-context.md) High Logged
3 7 commands missing enforcement (including client-facing) High Logged
4 48 legacy commands still discoverable, no guardrail Medium Logged
5 5 documentation sources disagree on agent count Medium Partially remediated
6 No dependency map exists High Not started
7 No automated validation scripts High Not started
8 5 repeated script patterns across 10-52 commands Low Not started
9 2 enforcement loading patterns coexist Low Not started
10 47 of 50 current commands lack delegation manifests High Not started

Marshal's Strategic Priorities

  1. Audit full chain for broken links (command -> agent -> skill -> script)
  2. Rewrite 47 remaining procedural commands to delegation manifest pattern (current commands only, never legacy)
  3. Close enforcement gap beyond V (Sage and Pax have no structural enforcement hooks)
  4. Reduce spawning cost (diagnose friction that makes delegation lose to inline work)
  5. Measure delegation health, not just existence (ongoing chain monitoring)

Leader Applications

V (Operations)

Primary beneficiary. Every time V's commands break because a skill path moved or an agent was renamed, Hone catches it first. V delegates layer health checks to Hone instead of running them inline -- consistent with the delegation enforcer hook that prevents V from operating as a solo operator. Hone's dependency map answers "if I change this skill, what breaks?" -- a question V currently cannot answer without manual grep across 202 skill files and 110 commands.

Sage (Customer)

Indirect beneficiary. Sage's commands (/relationship-brief, /meeting-prep, /client-comms, /sentinel-check) depend on skills and agents that Hone validates. When Hone identifies that a Sage command is missing enforcement or referencing a stale skill, Sage gets flagged. No direct interaction.

Pax (Finance)

Indirect beneficiary. Pax's commands (/revenue-brief, /investment-brief, /capacity-brief, /1099-prep) are in Hone's validation scope. Same model as Sage -- no direct interaction, but structural integrity validated.


Related Incidents

Aegis Onboarding Gap (CAR-2026-04-12-001)

During Hone's onboarding, the process exposed that Aegis had no codified onboarding checklist -- 7 of 10 registration points were missed on first pass. This resulted in a corrective action report and a new 12-point checklist added to Aegis's agent definition. Hone's commissioning improved not just the system's structural integrity but the process for adding future agents.

org-roster.yaml Deprecation

The deprecated agents/org-roster.yaml file was deleted during this session. Seven scripts still reference it -- logged as HubSpot task 107902503504 assigned to Hone as the first remediation work item.


Dependencies

Dependency Status Notes
.claude/agents/hone.md Confirmed Created and verified
agents/skill-command-architect/AGENT.md Confirmed Created and verified
INDEX.md roster entry Confirmed V Ops, Self-Improvement group
HubSpot Contact record Confirmed ID 215190664854
Agent Office DB registration Confirmed Registered
Org chart update Confirmed Added to Cytoscape.js visualization
Website data update Confirmed Added to ai-team-structure.ts
Job description Confirmed Filed
Marshal's Top 5 in agent def Pending Approved, not yet embedded
org-roster.yaml script rewrites Pending HubSpot task 107902503504

Verification

# Confirm agent definition exists
head -5 .claude/agents/hone.md

# Confirm implementation exists
head -5 agents/skill-command-architect/AGENT.md

# Confirm HubSpot Friend record
node scripts/hubspot/api.js search contacts --query "Hone" --properties firstname,email,agent_readiness_level

Spawning test: In a future session, invoke subagent_type: "hone" with "validate cross-references for the sentinel agent definition." Hone should load its startup protocol, read enforcement, and return a structured Layer Consistency Report.