| name | agent-coding-workflow |
| description | Use this skill when an AI coding agent edits ametel01 repos, especially with dirty worktrees, generated skills, CLI harness code, or multi-file changes. It encodes read-before-edit, scoped diffs, command verification, validation-before-completion, and user-change preservation habits observed across the org. |
Agent Coding Workflow
Purpose
Make agent edits deliberate, scoped, and verifiable across the org's repos.
When to use
Use before making code or generated-skill edits, when working in a dirty tree, when commands fail, or when the task spans multiple files or repos.
Evidence from repositories
Strength: Moderate.
agent-vitals/AGENTS.md: requires finishing requested parts and validating before completion.
- Current org pack instructions for
org-agent-skills: require reading target file and a related caller/test/type before edits, verifying cwd and command syntax, and adapting after failures.
agents-toolbelt/CONTRIBUTING.md: asks for focused changes, full quality gates, docs for behavior/workflow changes, and logical commits.
orgs-ai-harness/tests/test_org_pack_foundation.py: validates generated pack and skill structure, installation targets, and no unintended global writes in tests.
Standard workflow
- Inspect git status and treat existing changes as user-owned.
- Read the target file before editing. Also read at least one related caller, test, type definition, workflow, or validator.
- Keep edits scoped to the requested behavior and surrounding conventions.
- Before shell commands, confirm
cwd and command syntax. After a failure, read the full error and change approach.
- Run relevant validation before reporting completion.
- Report what changed, what validation ran, and any residual gap.
Required checks
- Do not revert or overwrite unrelated user changes.
- Do not write org-level skills directly into global runtime install targets; use staging or repo paths requested by the task.
- Prefer
rg/rg --files for repo inspection.
- Use existing scripts, validators, and tests rather than inventing one-off checks.
Common pitfalls
- Blind edits in unfamiliar files.
- Retrying a failing command without changing the approach.
- Reporting completion after only a subset of requested work.
- Mixing unrelated cleanup into a focused change.
- Treating generated output as installed runtime state when the harness expects staged output.
Escalation / uncertainty rules
If user-owned changes conflict with the requested edit, stop and explain the conflict. If validation cannot run because a tool is missing or a service is unavailable, state exactly which check was skipped and why.