Create Code

Create Code Skill

The create-code skill is the primary coding engine for the Horus persona. it orchestrates a multi-stage "Hardened Coding" workflow that ensures code is researched, isolated, tested, and brutally critiqued before completion.

Horus Coding Workflow

The skill follows a strict 6-stage pipeline:

1. Idea & Initial Scoping

Horus starts with a high-level idea or requirement.

• Stage 1 Gate: Clarification interview to define scope and constraints.

2. Deep Research (/dogpile)

Horus research existing implementation patterns, libraries, and potential pitfalls.

• Calls /dogpile search "<idea> implementation patterns"
• Aggregates context into a research summary.

3. Isolated Execution & Digital Twin (/battle)

Horus spins up an isolated environment to safely test code or run adversarial simulations. This leverages a Digital Twin strategy for high-fidelity testing.

• Tools: .pi/skills/battle for Digital Twin orchestration (isolation). Optionally use .pi/skills/hack for security audits.
• Modes:
  • git_worktree: For repository-level isolation.
  • docker: For containerized environment testing.
  • qemu: For hardware/microprocessor emulation (firmware).

4. Structured Implementation (/task, /orchestrate)

Implementation is driven by 0N_TASKS.md files with enforced quality gates.

• Uses /task to break down the idea into actionable items.
• Uses /orchestrate to execute tasks, requiring:
  • Sanity Tests: Every task must have a verification script.
  • Assertions: Core logic must be asserted.
  • Definition of Done (DoD): Strict criteria for task completion.

5. Brutal Code Review (/review-code)

Horus submits the code for a multi-round "Brutal Review".

• Calls /review-code using Copilot GPT-5 (or highest reasoning model available).
• Focuses on "No-Vibes" technical correctness, efficiency, and Horus's specific persona standards.

6. Final Research & Consolidation (/dogpile)

Horus performs a final dogpile search with the working code and full context to find any last-minute edge cases or optimizations.

• Consolidation of the project into the Horus/Memory knowledge graph.

Usage

When to use sub-skills

• /dogpile: use when the idea or tasks are unclear, novel, or library-dependent; automatically triggered in Stage 2.
• /battle: use when performance, resilience, or adversarial testing is needed (e.g., scale, concurrency, fuzz); triggered in Stage 3.
• /hack: use when security-sensitive surfaces exist (auth, crypto, network, plugins, serialization, execution); auto-suggested by heuristics and runnable via --security-audit.
• /anvil: use for heavy builds or reproducible environment provisioning; pair with /battle when complex dependencies are needed.

Retry policies

By default, skills have these retry counts with exponential backoff:

• dogpile: 3 retries (research may be flaky)
• battle: 2 retries (isolation can fail transiently)
• review-code: 2 retries (provider may timeout)
• hack: 1 retry (security audits are generally deterministic)
• orchestrate: 1 retry (task execution should be idempotent)

Override via environment or direct run_skill calls.

# Start a new coding project from an idea
./run.sh start "Implement a high-performance vector store with ArangoDB"

# Resume an existing creation (picks up from last saved stage)
./run.sh resume --project-dir /path/to/project

# Run specific stages (supports project-local logging in .create-code.log)
./run.sh research "idea"
./run.sh review --provider github --model gpt-5 --yes
./run.sh sandbox --mode docker --yes
./run.sh implement --yes

Commands

Key Principles

1. Isolation First: Never run untrusted or experimental code on the host.
2. Quality Gates: No code proceeds without passing sanity tests and DoD.
3. Brutal Critique: Embody Horus's uncompromising standard for technical excellence.
4. Memory Integration: Always check and update the knowledge graph.