version: "1.0.1" name: progressive-estimation description: "Estimate AI-assisted and hybrid human+agent development work with explicit PERT assumptions and calibration feedback loops" category: project-management risk: safe source: community date_added: "2026-03-10" author: Enreign tags:

• estimation
• project-management
• pert
• sprint-planning
• ai-agents

tools:

• claude

Progressive Estimation

Estimate AI-assisted and hybrid human+agent development work using three-point estimates, stated uncertainty assumptions and calibration feedback loops.

Overview

Progressive Estimation adapts to your team's working mode — human-only, hybrid, or agent-first — choosing explicit assumptions for each. Estimates remain uncertain and require local calibration.

When to Use This Skill

• Estimating development tasks where AI agents handle part of the work
• Sprint planning with hybrid human+agent teams
• Batch sizing a backlog (handles 5 or 500 issues)
• Staffing and capacity planning with agent multipliers
• Release date forecasting with confidence intervals

How It Works

1. Mode Detection — Determines if the team works human-only, hybrid, or agent-first
2. Task Classification — Categorizes by size (XS–XL), complexity, and risk
3. Formula Application — Applies locally measured multipliers when available, or clearly labeled scenarios
4. PERT Calculation — Produces expected values using three-point estimation
5. Confidence Bands — Generates percentiles only with a specified, calibrated forecast model
6. Output Formatting — Formats for Linear, JIRA, ClickUp, GitHub Issues, Monday, or GitLab
7. Calibration — Feeds back actuals to improve future estimates

Examples

Single task:

"Estimate building a REST API with authentication using Claude Code"

Batch mode:

"Estimate these 12 JIRA tickets for our next sprint"

With context:

"We have 3 developers using AI agents for ~60% of implementation. Estimate this feature."

Estimation procedure

Record scope, dependencies, acceptance criteria and whether durations mean effort or elapsed time. Elicit optimistic O, most likely M and pessimistic P in the same units, with O ≤ M ≤ P. A conventional PERT approximation is E = (O + 4M + P)/6 and SD ≈ (P − O)/6; these are modeling assumptions, not observed certainty. For O=2, M=4, P=12 hours, E=5 hours and SD≈1.67 hours. Record the assumptions behind each bound.

Do not derive P50/P75/P90 directly from the mean or label a range a confidence interval without a stated distribution and method. For a project forecast include dependencies, correlated risks, review/queue time and limited human capacity. Use actual comparable completed tasks to calibrate AI multipliers; absent data, present explicit scenarios rather than claiming a research-backed speedup. Report actuals against the original scope and revise future forecasts. This guide does not bundle an estimation engine or calibration database.

Best Practices

• Start with a single task to calibrate before moving to batch mode
• Feed back actual completion times to improve the calibration system
• Use "instant mode" for quick T-shirt sizing without full PERT analysis
• Be explicit about team composition and agent usage percentage

Common Pitfalls

• Problem: Overconfident estimates

Solution: Use P75 or P90 for commitments, not P50

• Problem: Missing context

Solution: The skill asks clarifying questions — provide team size and agent usage

• Problem: Stale calibration

Solution: Re-calibrate when team composition or tooling changes significantly

Related Skills

• @sprint-planning - Sprint planning and backlog management
• @project-management - General project management workflows
• @capacity-planning - Team velocity and capacity planning

Additional Resources

• Source Repository
• Installation Guide
• Research References

Limitations

• Use this skill only when the task clearly matches the scope described above.
• Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
• Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.