Agent Skills: Agent Behavior

Use at the start of any coding session to establish activity tracking and working patterns. Defines how the agent should document its work and communicate progress.

UncategorizedID: bfmcneill/agi-marketplace/agent-behavior

Install this agent skill to your local

pnpm dlx add-skill https://github.com/bfmcneill/agi-marketplace/tree/HEAD/core/skills/agent-behavior

Skill Files

Browse the full folder contents for agent-behavior.

Download Skill

Loading file tree…

core/skills/agent-behavior/SKILL.md

Skill Metadata

Name
agent-behavior
Description
Use at the start of any coding session to establish activity tracking and working patterns. Defines how the agent should document its work and communicate progress.

Agent Behavior

Rules for how the agent operates, tracks work, and communicates during coding sessions.

Activity Tracking

Every session must track what was done.

Log activity to docs/activity/ using your judgment on hierarchy:

docs/activity/
├── 2025-01-15-feature-auth.md      # By date + feature
├── 2025-01-15-bugfix-login.md      # By date + type
└── sessions/
    └── 2025-01-15-session-1.md     # By session

Activity Log Format

# [Date] - [Brief Description]

## What Was Done
- Bullet list of changes made
- Files modified
- Decisions made

## Why
- Reasoning behind approach
- Trade-offs considered

## What's Next
- Remaining work
- Known issues
- Questions for human

When to Log

| Situation | Action | |-----------|--------| | Starting work | Create/update activity log | | Completing a task | Summarize what was done | | Making a decision | Document the reasoning | | Hitting a blocker | Note the issue and questions | | Ending session | Final summary of state |

Working Patterns

Ask vs. Proceed

| Situation | Action | |-----------|--------| | Clear requirements | Proceed | | Multiple valid approaches | Ask | | Destructive operation | Ask | | Unclear scope | Ask | | Simple fix | Proceed |

Subagent Usage

Use subagents when:

  • Task is independent and parallelizable
  • Deep exploration needed without polluting main context
  • Multiple files need searching/analysis

Do directly when:

  • Simple, quick operation
  • Context is already loaded
  • Sequential dependency on previous work

Communication

  • Be concise - Don't over-explain obvious things
  • Show progress - Use todo lists for multi-step work
  • Surface blockers early - Don't spin on problems
  • Summarize at end - What was done, what's next

Quality Expectations

Before marking work "done":

  • [ ] Code runs without errors
  • [ ] Tests pass (if applicable)
  • [ ] Activity log updated
  • [ ] No obvious issues left unaddressed