A personal collection of agent skills, workflow references, multi-agent coordination protocols, delivery playbooks, and focused audit checklists for AI coding agents.
The standalone Markdown files directly under skills/ capture reusable agent workflows and extracted runtime behavior:
| Skill or reference | Purpose |
|---|---|
advisor | Configure a smaller Claude model to consult a stronger advisor model at difficult decision points. |
batch | Plan and execute wide, independently mergeable changes through parallel worktree agents and separate pull requests. |
code-review | Run a multi-angle diff review with independent finding verification and optional fixes or pull-request comments. |
deep-research | Produce a cited report through parallel searches, source extraction, and adversarial claim verification. |
goal | Keep an agent working until an evidence-backed completion condition is judged to be satisfied. |
init-verifiers | Generate project-specific functional verifier skills for browser, CLI, and HTTP behavior. |
professional-localization | Design localization as an end-to-end product boundary spanning UI, server, storage, contracts, and CI. |
simplify | Review a diff for reuse, simplicity, efficiency, and abstraction-level improvements, then apply the cleanup. |
testable-frontend | Structure frontend code and test tiers so deterministic automated testing is practical by construction. |
The repository contains both lightweight file-protocol coordination skills and Hive-backed skills that use named registry implementors, parallel reviewer panels, durable workflow artifacts, and bounded multi-repository execution.
| Skill | Purpose |
|---|---|
code-review-loop | Coordinate an implementor and persistent reviewer through session-isolated frozen-diff cycles, NO_FINDINGS work-item boundaries, and caller- or implementor-controlled session completion; the protocol includes TLA+ state-machine verification. |
parallel-code-review-loop | Schedule multiple independent code-review-loop pairs across distinct repository working paths, pass shared and assignment-specific context to implementors, enforce per-pair startup order, and wait for all pairs to terminate. See the one-workspace, N-repository example. |
hive-review-loop | Run one repository through a named hive_registry implementor and a configurable parallel reviewer panel using hive_workflow; it launches no Codex subagents. |
hive-parallel-review-loop | Run multiple disjoint repository review loops as bounded concurrent hive_workflow processes, with shared defaults and per-assignment registry-agent overrides in YAML. |
For one pair, invoke code-review-loop on both agents with the appropriate role:
/code-review-loop implementor
/code-review-loop reviewer
Omitting completion authority on the implementor defaults it to completion-authority=implementor. The reviewer has no completion-authority input. Use completion-authority=caller only on the implementor when an enclosing workflow, such as an epic implementor, must keep the reviewer alive across multiple work items and close the session itself. Continue such a session by resuming the same initialized implementor role; a fresh implementor invocation would perform startup cleanup and must begin a new session instead.
For multiple independent repositories, invoke parallel-code-review-loop on one orchestrator with its input manifest. The orchestrator validates disjoint repository roots and complete-pair capacity, then launches one code-review-loop implementor/reviewer pair per repository. Install the complete directories for both coordination skills because the parallel wrapper depends on the base protocol and its scripts.
For provider-neutral agent panels, use the Hive-backed variants. They call
hive_registry and
hive_workflow directly and
never create Codex subagents.
hive-review-loop runs one registry implementor with one or more parallel
registry reviewers; hive-parallel-review-loop runs that topology across
multiple repositories from one YAML manifest.
| Skill | Purpose |
|---|---|
verified-commit-split | Split uncommitted work, one commit, or a contiguous commit range into logical commits while proving final tree identity against a protected backup. |
skills/product-delivery-skills/ contains four self-contained, installable skills:
| Skill | Purpose |
|---|---|
product-implementation-planner | Turn a product description and repository evidence into a phased implementation plan. |
product-plan-compiler | Compile a detailed product plan into formal models, run relevant provers and model checkers, and refine inconsistencies before implementation. |
parallel-plan-implementation | Implement an approved plan through controlled, dependency-aware parallel phases. |
phase-commit-reviewer | Perform a read-only, phase-aware review of one implementation commit. |
The default delivery sequence is product-implementation-planner → product-plan-compiler → parallel-plan-implementation → phase-commit-reviewer.
Across that sequence, the main stage orchestrator maintains docs/implementation-plan/delivery-status.md as a concise, derived human view of scope, stage outcomes, decisions, blockers, and evidence links. It remains non-authoritative, and the operator is explicitly told its path and any required action after planning, formal verification, implementation, and optional review.
Each product-delivery skill includes its supporting references, templates, schemas, scripts, and a dedicated README.
skills/audit/ contains focused audit playbooks. Each one is intended to produce concrete findings with file paths, line numbers, impact, and suggested fixes.
| Skill | Purpose |
|---|---|
audit-input-validation | Audit client and server validation for user-controlled input. |
audit-auth-and-access-control | Audit auth flows, sessions, roles, admin routes, and cross-user data isolation. |
audit-secrets-and-config | Audit hardcoded secrets, environment validation, and webhook signatures. |
audit-rate-limiting | Audit API-route rate-limit coverage and policy fit. |
audit-cors | Audit CORS configuration and origin allowlists. |
audit-database-performance | Audit indexes, pagination, unbounded queries, and connection-pool risks. |
audit-resilience-and-observability | Audit error boundaries, external I/O, health checks, logs, and backups. |
audit-asset-pipeline | Audit uploads, storage, CDN/object storage, and asset delivery. |
audit-type-safety | Audit TypeScript safety bypasses and unchecked external data. |
security-review | Run an umbrella security review composed from focused web skills. |
ios-prelaunch-checklist | Audit iOS App Store assets, technical setup, legal readiness, and signing. |
Clone the repository and copy the skill or reference you need into the directory expected by your agent runtime:
skills/<standalone-workflow>-skill.md
skills/code-review-loop/
skills/parallel-code-review-loop/
skills/hive-review-loop/
skills/hive-parallel-review-loop/
skills/verified-commit-split/
skills/product-delivery-skills/<skill-name>/
skills/audit/<skill-name>/
skills/*-skill.md Standalone workflow skills and references
skills/code-review-loop/ Two-agent implementation/review coordination skill
skills/code-review-loop/verification/ TLA+ model, checks, and reproducibility notes
skills/parallel-code-review-loop/ Multi-repository orchestration over isolated review pairs
skills/hive-review-loop/ Hive-backed implementation and reviewer-panel loop
skills/hive-parallel-review-loop/ Bounded multi-repository Hive workflow launcher
skills/verified-commit-split/ Content-preserving Git commit-splitting skill
skills/audit/ Focused audit skills
skills/product-delivery-skills/ Self-contained product-delivery skills
LICENSE MIT license
Python
89.7%
TLA
8.4%
Shell
1.9%
A personal collection of agent skills, workflow references, multi-agent coordination protocols, delivery playbooks, and focused audit checklists for AI coding agents.
The standalone Markdown files directly under skills/ capture reusable agent workflows and extracted runtime behavior:
| Skill or reference | Purpose |
|---|---|
advisor | Configure a smaller Claude model to consult a stronger advisor model at difficult decision points. |
batch | Plan and execute wide, independently mergeable changes through parallel worktree agents and separate pull requests. |
code-review | Run a multi-angle diff review with independent finding verification and optional fixes or pull-request comments. |
deep-research | Produce a cited report through parallel searches, source extraction, and adversarial claim verification. |
goal | Keep an agent working until an evidence-backed completion condition is judged to be satisfied. |
init-verifiers | Generate project-specific functional verifier skills for browser, CLI, and HTTP behavior. |
professional-localization | Design localization as an end-to-end product boundary spanning UI, server, storage, contracts, and CI. |
simplify | Review a diff for reuse, simplicity, efficiency, and abstraction-level improvements, then apply the cleanup. |
testable-frontend | Structure frontend code and test tiers so deterministic automated testing is practical by construction. |
The repository contains both lightweight file-protocol coordination skills and Hive-backed skills that use named registry implementors, parallel reviewer panels, durable workflow artifacts, and bounded multi-repository execution.
| Skill | Purpose |
|---|---|
code-review-loop | Coordinate an implementor and persistent reviewer through session-isolated frozen-diff cycles, NO_FINDINGS work-item boundaries, and caller- or implementor-controlled session completion; the protocol includes TLA+ state-machine verification. |
parallel-code-review-loop | Schedule multiple independent code-review-loop pairs across distinct repository working paths, pass shared and assignment-specific context to implementors, enforce per-pair startup order, and wait for all pairs to terminate. See the one-workspace, N-repository example. |
hive-review-loop | Run one repository through a named hive_registry implementor and a configurable parallel reviewer panel using hive_workflow; it launches no Codex subagents. |
hive-parallel-review-loop | Run multiple disjoint repository review loops as bounded concurrent hive_workflow processes, with shared defaults and per-assignment registry-agent overrides in YAML. |
For one pair, invoke code-review-loop on both agents with the appropriate role:
/code-review-loop implementor
/code-review-loop reviewer
Omitting completion authority on the implementor defaults it to completion-authority=implementor. The reviewer has no completion-authority input. Use completion-authority=caller only on the implementor when an enclosing workflow, such as an epic implementor, must keep the reviewer alive across multiple work items and close the session itself. Continue such a session by resuming the same initialized implementor role; a fresh implementor invocation would perform startup cleanup and must begin a new session instead.
For multiple independent repositories, invoke parallel-code-review-loop on one orchestrator with its input manifest. The orchestrator validates disjoint repository roots and complete-pair capacity, then launches one code-review-loop implementor/reviewer pair per repository. Install the complete directories for both coordination skills because the parallel wrapper depends on the base protocol and its scripts.
For provider-neutral agent panels, use the Hive-backed variants. They call
hive_registry and
hive_workflow directly and
never create Codex subagents.
hive-review-loop runs one registry implementor with one or more parallel
registry reviewers; hive-parallel-review-loop runs that topology across
multiple repositories from one YAML manifest.
| Skill | Purpose |
|---|---|
verified-commit-split | Split uncommitted work, one commit, or a contiguous commit range into logical commits while proving final tree identity against a protected backup. |
skills/product-delivery-skills/ contains four self-contained, installable skills:
| Skill | Purpose |
|---|---|
product-implementation-planner | Turn a product description and repository evidence into a phased implementation plan. |
product-plan-compiler | Compile a detailed product plan into formal models, run relevant provers and model checkers, and refine inconsistencies before implementation. |
parallel-plan-implementation | Implement an approved plan through controlled, dependency-aware parallel phases. |
phase-commit-reviewer | Perform a read-only, phase-aware review of one implementation commit. |
The default delivery sequence is product-implementation-planner → product-plan-compiler → parallel-plan-implementation → phase-commit-reviewer.
Across that sequence, the main stage orchestrator maintains docs/implementation-plan/delivery-status.md as a concise, derived human view of scope, stage outcomes, decisions, blockers, and evidence links. It remains non-authoritative, and the operator is explicitly told its path and any required action after planning, formal verification, implementation, and optional review.
Each product-delivery skill includes its supporting references, templates, schemas, scripts, and a dedicated README.
skills/audit/ contains focused audit playbooks. Each one is intended to produce concrete findings with file paths, line numbers, impact, and suggested fixes.
| Skill | Purpose |
|---|---|
audit-input-validation | Audit client and server validation for user-controlled input. |
audit-auth-and-access-control | Audit auth flows, sessions, roles, admin routes, and cross-user data isolation. |
audit-secrets-and-config | Audit hardcoded secrets, environment validation, and webhook signatures. |
audit-rate-limiting | Audit API-route rate-limit coverage and policy fit. |
audit-cors | Audit CORS configuration and origin allowlists. |
audit-database-performance | Audit indexes, pagination, unbounded queries, and connection-pool risks. |
audit-resilience-and-observability | Audit error boundaries, external I/O, health checks, logs, and backups. |
audit-asset-pipeline | Audit uploads, storage, CDN/object storage, and asset delivery. |
audit-type-safety | Audit TypeScript safety bypasses and unchecked external data. |
security-review | Run an umbrella security review composed from focused web skills. |
ios-prelaunch-checklist | Audit iOS App Store assets, technical setup, legal readiness, and signing. |
Clone the repository and copy the skill or reference you need into the directory expected by your agent runtime:
skills/<standalone-workflow>-skill.md
skills/code-review-loop/
skills/parallel-code-review-loop/
skills/hive-review-loop/
skills/hive-parallel-review-loop/
skills/verified-commit-split/
skills/product-delivery-skills/<skill-name>/
skills/audit/<skill-name>/
skills/*-skill.md Standalone workflow skills and references
skills/code-review-loop/ Two-agent implementation/review coordination skill
skills/code-review-loop/verification/ TLA+ model, checks, and reproducibility notes
skills/parallel-code-review-loop/ Multi-repository orchestration over isolated review pairs
skills/hive-review-loop/ Hive-backed implementation and reviewer-panel loop
skills/hive-parallel-review-loop/ Bounded multi-repository Hive workflow launcher
skills/verified-commit-split/ Content-preserving Git commit-splitting skill
skills/audit/ Focused audit skills
skills/product-delivery-skills/ Self-contained product-delivery skills
LICENSE MIT license
Python
89.7%
TLA
8.4%
Shell
1.9%