What Are the Outputs Created by build-iterated-agentic-loop? A Complete File Reference

The build-iterated-agentic-loop skill generates four core repository-local artifacts: a skill definition (SKILL.md), a GitHub Actions workflow, an agent memory file, and optional reference files that together enable automated, self-improving coding-agent loops.

The build-iterated-agentic-loop skill in the humanlayer/skills repository scaffolds a complete iterated agentic loop infrastructure directly into your codebase. Unlike one-shot AI prompts, this pattern creates persistent, schedulable automation that improves across runs using feedback memory. The generated files live in standardized locations so any compatible coding agent can discover and execute them.

The Four Core Outputs of build-iterated-agentic-loop

Each execution of the skill produces a consistent set of files. According to the source code at [plugins/build-iterated-agentic-loop/skills/build-iterated-agentic-loop/SKILL.md](https://github.com/humanlayer/skills/blob/main/plugins/build-iterated-agentic-loop/skills/build-iterated-agentic-loop/SKILL.md), these are the mandatory and optional outputs.

1. Skill Definition (SKILL.md)

The skill definition is the primary artifact that encodes the agent's judgment, scope, and step-by-step completion criteria. It resides at:

.claude/skills/<skill-name>/SKILL.md

This file uses frontmatter for metadata and markdown for instructions. The build-iterated-agentic-loop skill populates it with a structured template including iteration boundaries, validation steps, and success criteria.

# .claude/skills/fix-eslint-issues/SKILL.md

---
name: fix-eslint-issues
description: Fix ESLint violations automatically
---

# Iterate through src/ and fix all auto-fixable ESLint errors

# STOP after 20 files or if any test fails

# VALIDATE by running `npm run lint` and `npm test`

The skill definition is read by the coding agent on every invocation, making it the single source of truth for task behavior.

2. GitHub Actions Workflow (agent-<task-name>.yml)

The workflow file automates execution through GitHub's CI infrastructure. Generated at:

.github/workflows/agent-<task-name>.yml

This workflow supports scheduled triggers, manual dispatch, or both. The build-iterated-agentic-loop skill uses the skeleton from [references/workflow-template.yml](https://github.com/humanlayer/skills/blob/main/plugins/build-iterated-agentic-loop/skills/build-iterated-agentic-loop/references/workflow-template.yml) and parameterizes it with your task name and agent configuration.

# .github/workflows/agent-fix-eslint.yml

name: Agent – Fix ESLint Issues
on:
  schedule:
    - cron: "0 3 * * MON"   # weekly on Monday

  workflow_dispatch:
jobs:
  run-agent:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v3
      - name: Run Coding Agent
        run: |
          # headless agent command injected here

      # PR creation and notification steps follow

The workflow enforces validation gates and creates pull requests for human review, preventing unbounded autonomous changes.

3. Agent Memory File (<task-name>.md)

The memory file provides persistent feedback that survives across runs, enabling true iteration. Located at:

.github/agent-memory/<task-name>.md

On each execution, the coding agent reads this file before starting work and appends new learnings afterward. This pattern prevents repeated mistakes and captures institutional knowledge like false-positive patterns or permanent exclusions.

# .github/agent-memory/fix-eslint.md

# Persistent feedback for the ESLint-fix loop

## Standing Instructions

- Exclude `src/generated/` from all linting operations
- Rule `no-console-log` is a known false-positive in test files

## Learned from 2024-01-15 run

- Prefer `logger.debug()` over manual `// eslint-disable` comments

The [references/memory-template.md](https://github.com/humanlayer/skills/blob/main/plugins/build-iterated-agentic-loop/skills/build-iterated-agentic-loop/references/memory-template.md) provides the initial structure.

4. Optional Reference Files

The references directory holds supplementary templates that keep the main skill file concise:

.claude/skills/<skill-name>/references/
├── prompt-template.md
├── response-template.md
├── agent-iteration.ts
└── workflow-template.yml   # if customized

These files are referenced from SKILL.md using relative paths. The build-iterated-agentic-loop skill seeds common templates from its own references/ directory, including:

• agent-runner-templates.md – Commands and secrets for various coding agents (CodeLayer, Claude Code, etc.)
• workflow-template.yml – The parameterized CI/CD skeleton

How the Outputs Work Together

The build-iterated-agentic-loop skill creates a closed feedback system:

1. Workflow triggers on schedule or manual dispatch
2. Agent reads SKILL.md for instructions and agent-memory/<task-name>.md for context
3. Agent executes bounded by iteration limits and validation rules
4. Agent appends new learnings to the memory file
5. Pull request captures changes for human review

This architecture separates what to do (skill definition), when to do it (workflow), and what we've learned (memory)—enabling progressively better automated assistance without prompt drift.

Summary

The build-iterated-agentic-loop skill generates these repository-local outputs:

• .claude/skills/<skill-name>/SKILL.md – Core task definition with scope, boundaries, and validation
• .github/workflows/agent-<task-name>.yml – Automated trigger and execution environment
• .github/agent-memory/<task-name>.md – Persistent feedback that improves across iterations
• .claude/skills/<skill-name>/references/ – Optional templates and supporting materials

These artifacts implement the iterated agentic loop pattern directly in your codebase, making AI-assisted automation schedulable, bounded, and self-improving.

Frequently Asked Questions

Where does the build-iterated-agentic-loop skill place the generated skill definition?

The skill definition is placed at .claude/skills/<skill-name>/SKILL.md relative to repository root. The <skill-name> is derived from your task identifier, creating a namespaced location that Claude Code and compatible agents automatically discover.

How does the agent memory file persist learning across runs?

The memory file at .github/agent-memory/<task-name>.md is read before each agent execution and appended to afterward. Because it lives in version control, feedback survives workflow restarts and can be manually edited by maintainers to correct or enrich the agent's context.

Can I customize the generated GitHub Actions workflow?

Yes. The skill uses [references/workflow-template.yml](https://github.com/humanlayer/skills/blob/main/plugins/build-iterated-agentic-loop/skills/build-iterated-agentic-loop/references/workflow-template.yml) as a starting point, but the generated .github/workflows/agent-<task-name>.yml is a standard workflow file you can modify. Changes persist across skill re-runs unless you explicitly regenerate the workflow.

What coding agents are supported by the generated workflow?

The build-iterated-agentic-loop skill includes runner templates for multiple agents in [references/agent-runner-templates.md](https://github.com/humanlayer/skills/blob/main/plugins/build-iterated-agentic-loop/skills/build-iterated-agentic-loop/references/agent-runner-templates.md). The generated workflow can be configured for CodeLayer, Claude Code, or other headless coding agents by substituting the appropriate command block and secrets configuration.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}