SuperAPI AI Directory

time-estimation

Office & productivity

View repository

Updated 10-16 03:42

About this skill

Time Estimation

Estimate task duration for better planning.

Estimation Categories

Category Duration Symbol Use For
Quick < 15 min ~15m Simple replies, quick fixes, small updates
Short 15-30 min ~30m Code reviews, short meetings, simple features
Medium 30-60 min ~1h Regular features, documentation, debugging
Long 1-2 hours ~2h Complex features, research, planning
Deep Work 2-4 hours ~4h Architecture, major features, learning

Estimation by Task Type

Code Tasks

Type Typical Estimate
Typo/small fix ~15m
Bug fix (known cause) ~30m
Bug fix (unknown cause) ~1h to ~2h
Small feature ~1h
Medium feature ~2h
Large feature ~4h or break down
Refactoring ~2h to ~4h

Review Tasks

Type Typical Estimate
Small PR (< 100 lines) ~15m
Medium PR (100-500 lines) ~30m
Large PR (500+ lines) ~1h
Document review ~30m
Architecture review ~1h

Communication Tasks

Type Typical Estimate
Quick message ~15m
Email response ~15m
Meeting prep ~30m
Meeting (scheduled) actual duration
Presentation prep ~2h

Research Tasks

Type Typical Estimate
Quick lookup ~15m
Tool evaluation ~1h
Technology research ~2h
Deep learning session ~4h

Estimation Rules

The 2x Rule

If uncertain, estimate 2x your initial guess:

  • "This seems like 30 min" → Estimate ~1h
  • "Maybe an hour" → Estimate ~2h

Break Down Large Tasks

If estimate > 4 hours:

  1. Break into subtasks
  2. Estimate each subtask
  3. Add 20% buffer for coordination

Context Switching Tax

Add time for:

  • Unfamiliar codebase: +50%
  • Interruption-heavy environment: +30%
  • Learning new tool: +100%

Daily Capacity

Realistic productive hours: 4-6 hours/day

Day Type Available
Meeting-heavy day 2-3 hours
Normal day 4-5 hours
Focus day 6-7 hours

Planning Rule

  • Morning focus: Schedule ~2h deep work
  • Afternoon: Schedule ~1h tasks
  • Buffer: Keep 1 hour unscheduled

Estimation Confidence

When presenting estimates:

  • High confidence: "~1h" (done similar tasks)
  • Medium confidence: "~1h-2h" (some unknowns)
  • Low confidence: "needs spike" (too many unknowns)

For low confidence, create a timeboxed research task first: - [ ] Spike: Investigate X ~1h → Then estimate the real task