estimate-calibrator

Estimate Calibrator

Replaces single-point guesses with structured three-point estimates: decomposes work into atomic units, estimates best/likely/worst case for each, identifies unknowns and assumptions, calculates aggregate ranges using PERT, and assigns confidence levels with explicit rationale.

Reference Files

Prerequisites

• Work item description (feature, task, project)
• Decomposed tasks (or use task-decomposer skill first)
• Context: team familiarity, tech stack, existing codebase

Workflow

Phase 1: Decompose Work

If the work item is not already decomposed into atomic units:

1. Break into tasks — Each task should be estimable independently.
2. Right granularity — Tasks should be 1 hour to 3 days. Larger tasks have higher uncertainty; break them down further.
3. Identify dependencies — Tasks on the critical path determine the minimum duration.

Phase 2: Three-Point Estimate

For each task, estimate three scenarios:

Key rule: Worst case is NOT "everything goes wrong." It's the realistic bad scenario (90th percentile), not the apocalyptic one (99th percentile).

Phase 3: Identify Unknowns

Categorize unknowns that affect estimates:

Phase 4: Calculate Ranges

For individual tasks, use the PERT formula:

Expected = (Best + 4 × Likely + Worst) / 6
Std Dev = (Worst - Best) / 6

For aggregate (project) estimates:

• Sum of expected values for total expected duration
• Root sum of squares of std devs for aggregate uncertainty

Phase 5: Assign Confidence

Output Format

## Estimate: {Work Item}

### Summary
| Scenario | Duration |
|----------|----------|
| Best case | {time} |
| Likely case | {time} |
| Worst case | {time} |
| **PERT expected** | **{time}** |
| **Confidence** | **{High/Medium/Low}** |

### Task-Level Estimates

| # | Task | Best | Likely | Worst | PERT | Unknowns |
|---|------|------|--------|-------|------|----------|
| 1 | {task} | {time} | {time} | {time} | {time} | {key unknown or "None"} |
| 2 | {task} | {time} | {time} | {time} | {time} | {key unknown} |
| | **Total** | **{sum}** | **{sum}** | **{sum}** | **{pert}** | |

### Key Unknowns

| # | Unknown | Category | Impact on Estimate | Mitigation |
|---|---------|----------|-------------------|------------|
| 1 | {unknown} | {Technical/Scope/External} | +{time} if realized | {spike, prototype, early test} |

### Assumptions
- {Assumption 1 — what must be true for this estimate to hold}
- {Assumption 2}

### Risk Factors
- {Risk}: If realized, adds {time}. Likelihood: {High/Medium/Low}.

### Confidence Rationale
**{High/Medium/Low}** because:
- {Specific reason — e.g., "Team has built 3 similar features"}
- {Specific reason — e.g., "External API is a new integration"}

### Recommendation
{Commit to PERT expected with {X}% buffer, or spike the top unknown first.}

Calibration Rules

1. Three points, not one. Single-point estimates are always wrong. Three points communicate uncertainty — the most important part of any estimate.
2. Worst case is the 90th percentile, not the 99th. "Asteroid hits the office" is not a useful worst case. "The API documentation is wrong and we need to reverse-engineer the protocol" is realistic worst case.
3. Unknowns inflate estimates more than known difficulty. A hard but well-understood task is more predictable than an easy but novel one.
4. Estimates are not commitments. Communicate ranges, not deadlines. If stakeholders need a single number, give the PERT expected plus a buffer for confidence level.
5. Spike unknowns early. If a single unknown dominates the estimate range, invest 1-2 days spiking it before estimating the rest.

Error Handling

When NOT to Estimate

Push back if:

• The work is exploratory (research, spikes) — timebox instead of estimating
• Requirements are completely undefined — define scope first
• The user wants precision (hours) for a large project — provide ranges, not false precision
• The estimate will be used as a commitment without acknowledging uncertainty