Agent Dev Guardrails
Mental model: The agent is an extremely confident junior dev with amnesia. This skill provides guardrails, progressive context, and automated quality checks.
This skill enforces disciplined development workflows. Essentials are here in SKILL.md; detailed patterns are in references/, loaded only when needed.
Non-Negotiables (Always Apply)
- Plan first, implement second. For non-trivial changes: produce plan, risks, and task checklist before editing.
- Work in small slices. Implement 1–2 checklist items at a time; pause for review/testing between slices.
- Never leave errors behind. Run checks and fix failures before moving on.
- Be explicit about changes. Every response that edits code must include:
- Files changed (paths)
- Why the change is correct
- How it was validated (commands + outcome)
- If stuck for 30 minutes, stop. Ask for narrower scope or propose different approach; don't thrash.
Hook System
Pre-Work Hook (Before Coding)
Before proposing code changes:
- Identify scope — Small fix (no dev-docs) or large task (create dev-docs folder)
- State applicable skills — Which domain skills apply (e.g., python-engineering, security)
- For non-trivial work:
- Ask clarifying questions (see references/planning-protocol.md)
- Produce plan with phases + tasks + risks
- Wait for plan acceptance before implementation
Post-Work Hook (After Coding)
After every code edit:
- Run linting — Use
ReadLintson edited files - Fix failures — Do not leave errors behind
- Self-review — Apply specialized roles (see references/specialized-roles.md)
- Summarize:
- Files changed
- Commands run + outcomes
- Next slice (if applicable)
When to Read Which Reference
| Situation | Read |
|---|---|
| Need to plan or ask clarifying questions | references/planning-protocol.md |
| Complex task (>30 min), need task folders | references/dev-docs-system.md |
| Self-reviewing code after edits | references/specialized-roles.md |
| Setting up new project conventions | references/project-setup.md |
| Checking Definition of Done | references/quality-gates.md |
Quick Patterns
- Token efficiency: Request only the context you need; reference file paths and symbols instead of pasting large code blocks.
- Planning: Ask minimum context → produce plan with risks/tasks → wait for acceptance
- Questions: Ask "What are the tradeoffs? What breaks? What's missing?" not "Is this good?"
- Slicing: One slice = 1–2 checklist items; each slice ends with validation
- Review: Apply reviewer role personas (Architecture, Security, Test Engineer)
- Done: Code compiles, lint passes, edge cases handled, docs updated
Quick Reference / Examples
| Task | Approach |
|---|---|
| Start non-trivial work | Read references/planning-protocol.md, produce plan, wait for approval |
| Create task docs | Read references/dev-docs-system.md, create dev/active/<task-name>/ |
| Self-review code | Read references/specialized-roles.md, apply relevant personas |
| Set up new project | Run the setup script from the skill's scripts/ folder (see references/project-setup.md for path and options) |
| Check if slice is done | Read references/quality-gates.md, verify all criteria |
Pre-work check example:
Scope: Large task (auth feature)
Skills: agent-dev-guardrails + python-engineering
Plan: [3 phases, 8 tasks, 2 risks identified]
Waiting for approval before implementation.
Post-work summary example:
Files changed: src/auth.py, tests/test_auth.py
Validation: ReadLints (0 errors), pytest (12 passed)
Self-review: ✓ Input validation, ✓ No hardcoded secrets, ⚠ Consider rate limiting
Next: Slice 2 (login endpoint)
Workflow
- Starting any change → Check scope, activate skills, follow pre-work hook
- Planning complex work → Read references/planning-protocol.md, then references/dev-docs-system.md
- After every edit → Follow post-work hook, apply references/specialized-roles.md
- Setting up a project → Run the setup script from the skill's
scripts/folder; see references/project-setup.md - Marking work complete → Verify against references/quality-gates.md
Keep SKILL.md lean; load reference files only when relevant to the task.