A collection of multi-role agent skills for Claude Code. Each skill decomposes a complex task into isolated roles — independent subagents with scoped visibility and fixed responsibilities — that coordinate through structured handoffs.
The core principle across all skills: an agent that produced something cannot be an unbiased evaluator of that thing. Role isolation prevents anchoring, scope creep, and self-assessment bias.
Built on ideas from:
- Agentic Self-Correction — generate, validate, correct, repeat
- Measuring Context Pollution — detecting semantic drift from original intent
- Context Management — maintaining coherence over long-running workflows
A self-correcting architecture where agents iteratively validate and improve an artifact until it achieves both structural validity and coherence for its audience. The base template has four roles — Orchestrator, Validator, Optimizer, Fixer. Skills may add domain-specific roles (e.g., blockkit-builder adds a Generator for interactive/auto Block Kit production).
Two agents assess the artifact in parallel from opposite directions: the Validator checks constraints (rules only, no domain knowledge), the Optimizer assesses fitness (domain and audience only, no rules). The Orchestrator merges both reports and resolves conflicts. The Fixer applies surgical corrections. The loop repeats until both lenses come back clean or the iteration bound is hit.
Orchestrator dispatches Validator + Optimizer (parallel)
→ Both produce identically structured findings reports
→ Orchestrator merges into single fix report
→ Fix report empty → done
→ Fix report has findings → Fixer works
→ Orchestrator dispatches Validator + Optimizer again
→ Repeat until clean or bound hit (default: 3 rounds)
→ Unresolved findings → surface to human as residual
Instances:
math-verify/— verifies mathematical correctness of journal papers against 23 constraintsblockkit-builder/— builds and validates Slack Block Kit JSON (5 roles — adds Generator with interactive/auto modes)
Tooling:
create-convergence-skill/— generates new convergence skills from an interview process; bundles the canonical template atassets/template/scan-convergence-opportunities/— surveys a project for artifacts that would benefit from convergence skills
A five-role agent system (Orchestrator, Planner, Auditor, Executor, Drift Monitor) for planning and executing any complex multi-step task. Not a convergence skill — different architecture, different roles, different purpose.
The Planner locks in a detailed step-by-step plan with the user and is wiped after handoff. The Executor does the work. The Auditor judges every step against the plan before and after execution. The Drift Monitor measures context pollution — the semantic distance between the approved plan and the Orchestrator's accumulated state — and triggers re-anchoring when the Orchestrator drifts. The Orchestrator brokers all communication, owns all state, and is the only agent that decides whether to course correct, re-anchor, or escalate to the human.
All roles are independent subagents. The invoking agent never assumes any role. This prevents anchoring — an agent that produced the artifact can't be an unbiased validator.
Scoped visibility. Each agent sees only what it needs. Agents that assess from different lenses never see each other's reports. Agents that execute never see evaluation criteria.
Per-finding confidence scoring. Every finding carries two independent scores: issue confidence (is this a real problem) and fix confidence / correctability (is the proposed fix correct). The orchestrator uses both to decide what gets auto-approved, escalated, or dropped.
Orchestrator-mediated communication. No agent talks directly to another. The Orchestrator brokers all messages and filters what each agent sees.
Gotchas grow from real runs. Each skill maintains a gotchas file that accumulates operational intelligence — mistakes the agents make that get corrected.
The fastest path:
- Install the
create-convergence-skillskill - Describe your artifact, rules, and audience
- Review the drafted constraints
- The skill generates a complete, ready-to-run convergence skill
Or manually:
- Copy
create-convergence-skill/assets/template/to your skills directory and rename it - Fill in
references/constraints.mdwith your structural rules - Fill in
references/domain.mdwith your field knowledge - Fill in
references/context.mdwith your audience and thresholds - Fill in
references/log-format.mdwith your log directory and naming convention - Update
SKILL.mdfrontmatter and pre-flight checks
- Claude Code (CLI, desktop app, or IDE extension)
- Subagent support (the roles run as independent subagents)
MIT