- \n
- Agentic coding shifts the engineering problem from generating code to designing the context, workflows, tools, checks, and approval boundaries around coding agents.<\/li>\n
- Repository instructions should hold stable project facts, while Skills should hold repeatable methods that an agent needs only for specific tasks.<\/li>\n
- Claude Code Skills and Codex Skills support a shared Skill structure, but runtime-specific tools, permissions, and instructions still affect portability.<\/li>\n
- Skills work best alongside deterministic scripts, MCP tools, evals, and human review rather than replacing every layer with instructions.<\/li>\n
- Engineering teams should standardize one repetitive, testable workflow before they build a large Skill library.<\/li>\n<\/ul>\n
How Is Agentic Coding Changing Engineering Work?<\/h2>\n
Agentic coding moves AI from one step in development into a workflow that can gather context, plan work, edit files, run tools, test results, and return decisions to an engineer.<\/p>\n
That is a bigger change than better code completion.<\/p>\n
A coding assistant waits for an engineer to decide what happens next. A coding agent can participate in the sequence itself. That makes the surrounding workflow almost as important as the model.<\/p>\n
Anthropic’s harness research<\/a> shows what changes when the workflow gets more deliberate. Anthropic used a planner, generator, and evaluator for long-running application development. It also broke work into smaller chunks and passed structured artifacts between stages rather than asking one agent to carry the whole build in one undifferentiated task.<\/p>\n
The lesson is not that every coding task needs three agents.<\/p>\n
Anthropic found that the evaluator still helped on work near the edge of the model’s capabilities, but became unnecessary overhead on tasks the newer model could already handle reliably.<\/p>\n
That creates a different engineering question:<\/p>\n
How much structure does this task actually need?<\/strong><\/p>\n
Every extra instruction, evaluator, wrapper, or checkpoint creates another assumption the team may eventually need to maintain.<\/p>\n
Agentic coding therefore changes two things at once. The model writes more of the code, while the engineering team designs more of the system around how that work happens.<\/p>\n
Why Does Agentic Coding Need More Than AGENTS.md or CLAUDE.md?<\/h2>\n
Agentic coding needs more than one permanent instruction file because not every engineering procedure belongs in every task’s context.<\/strong><\/p>\n
Repository instructions still have an important job.<\/p>\n
Architecture boundaries, package-management rules, testing commands, naming conventions, and project-wide constraints affect many tasks. They make sense as persistent project context.<\/p>\n
A release-verification procedure does not.<\/p>\n
Neither does a database migration checklist, dependency-upgrade workflow, or project-specific bug-reproduction process.<\/p>\n
The Claude Code Skills docs<\/a> make that distinction explicit. Anthropic recommends creating a Skill when engineers repeatedly paste the same checklist or multi-step procedure, or when a section of
CLAUDE.md<\/code> has grown into a procedure rather than a fact. Claude loads the Skill body only when the workflow uses it.<\/p>\nThat gives teams a cleaner way to organize agent context:<\/p>\n
- \n
- Stable project facts<\/strong> belong in repository instructions because many tasks need them.<\/li>\n
- Repeatable procedures<\/strong> belong in Skills because only relevant tasks need the full method.<\/li>\n
- Mechanical checks<\/strong> belong in scripts or hooks because code can verify them consistently.<\/li>\n
- Current information<\/strong> belongs in tools or MCP connections because a static instruction file will go stale.<\/li>\n<\/ul>\n
This is why context engineering should not mean “write more instructions.”<\/p>\n
It is closer to information architecture for an engineering agent.<\/p>\n
![\"Agent](\"https:\/\/i0.wp.com\/procreator.design\/blog\/wp-content\/uploads\/2026\/10\/01-Agent-context-architecture.png?resize=1024%2C682&ssl=1\")
<\/p>\nThe distinction also becomes easier once teams separate Claude, Codex, and Agent Skills<\/a> by what belongs to the shared format and what depends on the product running it.<\/p>\n
Where Do Claude Code Skills and Codex Skills Fit?<\/h2>\n
Claude Code Skills and Codex Skills sit between persistent project context and the tools that perform work. A Skill gives the agent a reusable method for handling a recognizable task.<\/p>\n
Both ecosystems now support the same basic idea: a
SKILL.md<\/code> file carries instructions, while supporting files can hold references, scripts, examples, or other resources.<\/p>\nClaude Code states that its Skills follow the open Agent Skills standard. OpenAI likewise says its Skills are compatible with the open Agent Skills standard and can package instructions, references, scripts, and assets.<\/p>\n
- Repeatable procedures<\/strong> belong in Skills because only relevant tasks need the full method.<\/li>\n
- Stable project facts<\/strong> belong in repository instructions because many tasks need them.<\/li>\n
![\"agenctic](\"https:\/\/i0.wp.com\/procreator.design\/blog\/wp-content\/uploads\/2026\/10\/02-agenctic-coding-stack.png?resize=1024%2C682&ssl=1\")
<\/p>\n![\"Portable](\"https:\/\/i0.wp.com\/procreator.design\/blog\/wp-content\/uploads\/2026\/10\/03-Portable-method-vs-runtime.png?resize=1024%2C682&ssl=1\")
<\/p>\n![\"Agentic](\"https:\/\/i0.wp.com\/procreator.design\/blog\/wp-content\/uploads\/2026\/10\/04-Agentic-coding-workflow.png?resize=1024%2C682&ssl=1\")
<\/p>\n![\"method](\"https:\/\/i0.wp.com\/procreator.design\/blog\/wp-content\/uploads\/2026\/10\/05-method-vs-context.png?resize=1024%2C682&ssl=1\")
<\/p>\n