Research / oh-my-codex and AI Agent Workflow Patterns

oh-my-codex and AI Agent Workflow Patterns

oh-my-codex and AI Agent Workflow Patterns: What Can We Leverage?

2026-04-03 · agentic-ai tools-infrastructure ai-architecture · medium · source → · wiki →
key claims
  1. OMX separates operating contract from execution surface — AGENTS.md is the top-level workspace contract; role-specific prompts are narrower execution surfaces that must follow it, not override it. This layered approach prevents role-specific rules from conflicting with global constraints and is directly applicable to the Research repo's copilot-instructions.md and skills structure
  2. The Clarify, Plan, Execute, Verify (CPEV) workflow is the most transferable OMX pattern — $deep-interview clarifies scope before implementation; $ralplan approves the plan; $ralph or $team executes; a verifier role confirms completion. Adopting a scope-clarification gate before research investigation would reduce scope drift on ambiguous questions
  3. AGENTS.md is a cross-agent standard adopted by 15 or more tools — Unlike copilot-instructions.md (Copilot-only), AGENTS.md is readable by Codex CLI, Cursor, Claude Code, Devin, and others. Its adoption in each active davidamitchell repo would extend current guidance to all these tools
  4. Three-tier boundary statements (always/ask first/never) outperform free-form constraints — Empirical analysis of 2,500 or more agent files found this pattern consistently prevents destructive actions. The current copilot-instructions.md uses prose constraints, which are less reliable
  5. Executable commands must appear early in agent files — Agents reference commands frequently; burying them later causes agents to miss them. The six core areas for high-performing agent files are: commands, testing, project structure, code style, git workflow, and boundaries
  6. The davidamitchell Skills repo has 16 skills but no discoverability index — There is no map of skill name to trigger conditions to expected output, analogous to OMX's `/skills` browse surface. This makes the skill library harder to use reliably, particularly for new agents or sessions
  7. Model routing by task complexity is a low-cost, high-value improvement — OMX maps lightweight tasks (explore, search) to efficient models and heavy tasks (synthesis, architecture) to capable models, reducing token use and improving output quality. Adding similar routing guidance to copilot-instructions.md is a simple text addition
  8. Persistent machine-readable session state is absent from the Research repo — OMX stores plans, logs, and mode state in `.omx/` for cross-session continuity. The Research repo uses progress/ for human-readable session logs but has no machine-readable state store

Research Question

What patterns from oh-my-codex (OMX) and similar AI agent workflow projects (AGENTS.md, SKILL.md, etc.) are most applicable to improving the instructions, skills, agents, and tooling across davidamitchell's repositories — and which specific changes would deliver the highest value?

Findings

(Populated from §6 Synthesis above.)

Executive Summary

oh-my-codex (OMX) provides a mature reference architecture for AI agent workflow orchestration built around a cross-agent operating contract (AGENTS.md), four canonical workflow skills ($deep-interview, $ralplan, $ralph, $team), a role catalog with model routing, and persistent state management. The davidamitchell Research repo has strong domain-specific skill coverage (16 skills) but is Copilot-specific and lacks several structural patterns from the OMX model. The highest-value improvements are: adding a root AGENTS.md to each active repo, adopting three-tier boundary statements, adding a scope-clarification skill for ambiguous research questions, adding model routing guidance, and creating a skills index for discoverability. These improvements are additive and do not require removing existing infrastructure.

Key Findings

  1. OMX separates operating contract from execution surface — AGENTS.md is the top-level workspace contract; role-specific prompts are narrower execution surfaces that must follow it, not override it. This layered approach prevents role-specific rules from conflicting with global constraints and is directly applicable to the Research repo's copilot-instructions.md and skills structure. [fact/inference]

  2. The Clarify, Plan, Execute, Verify (CPEV) workflow is the most transferable OMX pattern — $deep-interview clarifies scope before implementation; $ralplan approves the plan; $ralph or $team executes; a verifier role confirms completion. Adopting a scope-clarification gate before research investigation would reduce scope drift on ambiguous questions. [fact/inference]

  3. AGENTS.md is a cross-agent standard adopted by 15 or more tools — Unlike copilot-instructions.md (Copilot-only), AGENTS.md is readable by Codex CLI, Cursor, Claude Code, Devin, and others. Its adoption in each active davidamitchell repo would extend current guidance to all these tools. [fact]

  4. Three-tier boundary statements (always/ask first/never) outperform free-form constraints — Empirical analysis of 2,500 or more agent files found this pattern consistently prevents destructive actions. The current copilot-instructions.md uses prose constraints, which are less reliable. [fact]

  5. Executable commands must appear early in agent files — Agents reference commands frequently; burying them later causes agents to miss them. The six core areas for high-performing agent files are: commands, testing, project structure, code style, git workflow, and boundaries. [fact]

  6. The davidamitchell Skills repo has 16 skills but no discoverability index — There is no map of skill name to trigger conditions to expected output, analogous to OMX's /skills browse surface. This makes the skill library harder to use reliably, particularly for new agents or sessions. [fact/inference]

  7. Model routing by task complexity is a low-cost, high-value improvement — OMX maps lightweight tasks (explore, search) to efficient models and heavy tasks (synthesis, architecture) to capable models, reducing token use and improving output quality. Adding similar routing guidance to copilot-instructions.md is a simple text addition. [inference]

  8. Persistent machine-readable session state is absent from the Research repo — OMX stores plans, logs, and mode state in .omx/ for cross-session continuity. The Research repo uses progress/ for human-readable session logs but has no machine-readable state store. [fact/inference]

Assumptions

Analysis

The davidamitchell Research repo is already operating at a high level of agent-instruction sophistication relative to most public repositories. The Skills submodule with 16 domain-specific skills, the structured research-prompt.md process, and the copilot-instructions.md document represent meaningful prior investment. The gap relative to OMX is structural rather than depth: the current setup is Copilot-optimised and lacks the cross-agent portability, explicit role routing, and workflow-progression gates that OMX provides.

The highest-leverage changes are additive: adding AGENTS.md does not require removing copilot-instructions.md; adding a skills index does not require rewriting skills; adding three-tier boundaries is a text addition to an existing file. The one area that requires more design work is the pre-research clarification gate, which would need to integrate with or replace part of the existing research-prompt.md process.

Cross-repo consistency is a second-order issue: the Multi-Agent-Testing, Agent-Evaluation, and Latest-developments- (trailing hyphen is the actual repo name) repos likely benefit from the same AGENTS.md additions, but each has different stacks and purposes.

Risks, Gaps, and Uncertainties

Open Questions

sources

← newerThe shape of organisations when software is no longer the constraint
older →Anthropic Claude Code leak

Connected items

Loading…

View full knowledge graph →