Agent Skills: Checkpoint

Checkpoint

Pause, assess, and surface intelligent next-step options to the user.

When to Trigger Proactively

Suggest this skill (without being asked) when detecting:

• Completion signal: A feature, fix, or milestone just finished
• Uncertainty signal: Requirements unclear, multiple valid paths, or low confidence in current direction
• Complexity signal: Scope expanded, unexpected dependencies emerged, or task is taking longer than expected
• Drift signal: Work may have diverged from user's original intent
• Quality signal: Code works but may benefit from review, testing, or refactoring

Workflow

1. Context Assessment

Silently evaluate the current state across these dimensions (do not output this analysis):

• Progress: What has been accomplished? What remains?
• Quality: Is the work solid, or are there rough edges?
• Alignment: Does recent work match what the user actually wants?
• Uncertainty: What assumptions were made? What's unclear?
• Risk: What could go wrong? What hasn't been tested?
• Efficiency: Is there a better path forward?

2. Generate Options

Based on assessment, generate 2-5 contextually-appropriate options. Draw from (but don't limit to) these archetypes:

| Archetype | When Relevant | |-----------|---------------| | Commit progress | Meaningful progress made, good stopping point | | Systems audit | Complex changes, potential for bugs or regressions | | Prioritize/plan | Multiple pending tasks, unclear what matters most | | Re-evaluate decisions | Low confidence in recent choices, new information available | | Clarify with user | Assumptions made, requirements ambiguous | | Test/verify | Code works but edge cases untested | | Refactor/clean up | Code functional but messy | | Document | Complex logic that needs explanation | | Step back | May be overcomplicating or missing simpler solution | | Continue current path | Clear next step, no reason to pause |

Option generation principles:

• Options should be meaningfully different, not variations of the same thing
• Include at least one "continue forward" option when momentum is valuable
• Include at least one "pause and verify" option when risk is present
• Avoid analysis paralysis - fewer sharp options beat many vague ones

3. Select Recommendation

Choose one option as recommended. The recommendation should reflect:

• What would a thoughtful senior engineer do here?
• What reduces risk of wasted effort or rework?
• What serves the user's underlying goals (not just stated requests)?

4. Present via AskUserQuestion

Use AskUserQuestion with this structure:

Question: "What would you like to do next?"
Header: "Next step" (or contextually appropriate 1-2 words)
Options: [generated options with descriptions]

Option format:

• label: Action verb phrase (e.g., "Commit current progress", "Run systems audit")
• description: 1 sentence explaining what this involves and why it might be valuable

Recommendation:

• Place recommended option FIRST in the list
• Append "(Recommended)" to its label
• Include rationale in the description

Example Output

For a scenario where a feature was just implemented but with some shortcuts:

AskUserQuestion:
  question: "Feature implementation complete. What would you like to do next?"
  header: "Next step"
  options:
    - label: "Review and refactor (Recommended)"
      description: "Clean up the shortcuts taken during implementation before they become technical debt. The core logic works but could be more maintainable."
    - label: "Add test coverage"
      description: "Write tests for the new feature to catch edge cases and prevent regressions."
    - label: "Commit and move on"
      description: "The feature works - commit it and tackle the next task. Can refactor later if needed."
    - label: "Walk me through what was built"
      description: "Explain the implementation so you can verify it matches your expectations before proceeding."

Anti-Patterns

• Don't overthink: This skill should take seconds, not minutes
• Don't list every possible option: Curate the most valuable 2-5
• Don't recommend "ask user" when the situation is clear: Have a point of view
• Don't trigger too frequently: Reserve for genuine decision points, not every minor step
• Don't explain the assessment process: Just present the options naturally