pmo-pert-estimate

PMO PERT Estimate — Three-point estimation workbooks

1. Overview

This skill produces PMI-compliant PERT three-point estimation workbooks through a multi-phase agentic pipeline. Starting from project documents (SoW, RFP, scope descriptions), it interactively builds a WBS, resource breakdown, risk register, and three-point estimates, then generates a fully automated Excel workbook with live formulas (PERT, SUM rollups, cross-sheet references, confidence intervals). The skill follows three key principles: progressive disclosure of reference documents (loaded only when the relevant phase begins, never all at once), adaptive interaction across three levels (Formative / Collaborative / Autonomous) chosen by the user with dynamic adjustment, and strict input/formula separation where Excel formula cells are always injected as strings and never overwritten with computed values.

Bundled assets:

assets/
  pert-template.xlsx              <- bundled reference template
scripts/
  generate_excel.py               <- JSON -> Excel generator (openpyxl)
  validate_template.py            <- custom template validator
  helpers/
    __init__.py
    wb_wbs.py                     <- WBS sheet logic
    wb_timeline.py                <- Timeline/Gantt sheet logic
    wb_resources.py               <- Resources matrix sheet logic
    wb_risks.py                   <- Risk register sheet logic
    wb_summary.py                 <- Summary sheet logic (cross-refs)
    formatting.py                 <- Shared styles, colors, fonts, formats
references/
  workflow.md                     <- detailed agentic flow description
  pmi-methodology.md              <- PMI guide for formative mode
  template-criteria.md            <- criteria for custom templates
  excel-schema.md                 <- column/formula/cross-ref schema per sheet
  interaction-levels.md           <- description of 3 interaction levels

Workspace directory: docs/pert-workspace/ (created during Phase 1) Output directory: configurable per project (default: docs/outbox/)

---

2. Phase 0 -- First Run Setup

When running in a repository for the first time, check whether CLAUDE.md contains a ## pmo-pert-estimate Configuration section. If it does not exist, guide the user through setup.

Steps

1. Ask the user for project-specific values:

   • Language -- output language (default: en)
   • EffortUnit -- pd / hours / story_points (default: pd)
   • DurationUnit -- days / weeks / sprints (default: d)
   • PrimaryColor -- hex color for Excel formatting (default: 1B4FA5)
   • Currency -- currency code (default: EUR)
   • PeriodType -- weekly / biweekly / monthly (default: biweekly)
   • AvgRate -- average daily rate for cost calculations (optional, default: none)
   • ManagementReservePct -- management reserve percentage (default: 10)
   • OutputDir -- where to save generated workbooks (default: docs/outbox/)

2. Ask about template. Offer four options:

   • Use bundled -- inform that the base template is at assets/pert-template.xlsx (relative to this skill directory) and criteria are documented in references/template-criteria.md
   • Use custom -- run validation:

     cd <skill-dir>/scripts && python3 validate_template.py --template <user_path>
     

     If valid: record path. If invalid: show specific errors, offer to fall back to bundled.
   • Customize now -- copy bundled template to project directory for user modification, then validate
   • Generate empty only -- produce the base template and stop (inspection mode)

3. Write the configuration section into CLAUDE.md:

## pmo-pert-estimate Configuration

| Field | Value |
|-------|-------|
| Language | en |
| EffortUnit | pd |
| DurationUnit | d |
| PrimaryColor | 1B4FA5 |
| Currency | EUR |
| PeriodType | biweekly |
| AvgRate | (none) |
| ManagementReservePct | 10 |
| OutputDir | docs/outbox/ |
| CustomTemplate | (bundled) |

4. Confirm to the user that future invocations will use these values automatically.

If the section already exists, read values from it and proceed silently.

---

3. Phase 1 -- Interactive Setup

This phase runs in-skill (no agent call). Collect all inputs needed for subsequent phases.

Steps

1. Ask for input documents -- path(s) to SoW, RFP, scope description, or pasted text
2. Ask if a reference folder exists with additional materials (contracts, previous estimates, org charts)
3. Ask interaction level:
   • (A) Formative -- full guidance, explains every PMI decision (for users new to PERT/PMI)
   • (B) Collaborative -- agent proposes, user validates (default)
   • (C) Autonomous -- agent decides, user reviews final output (for experienced PMOs)
4. Ask if there are target values for total effort and/or duration
   • If yes: record targets for reconciliation in Phase 4
5. Create working directory: docs/pert-workspace/
   • If it already exists: ask whether to reuse or create fresh

Store all choices (document paths, reference folder, interaction level, targets) for use in subsequent phase prompts.

---

4. Phase 2 -- Context Analysis

Progressive disclosure: Read references/workflow.md Phase 2 section before constructing the agent prompt.

Launch an agent to analyze the input documents and extract project context.

Agent(model="opus")

Agent prompt must include:

• Exact file paths to read: the input document(s) and reference folder contents collected in Phase 1
• The interaction level chosen by the user
• Instruction to extract: scope and boundaries, constraints (time/budget/regulatory), assumptions, stakeholders, deliverables, role references, phases and milestones mentioned, risks already identified in input
• For Level A: include methodology explanation sections in the output (what scope/constraints/assumptions mean, why they matter)
• For Level B: present complete context analysis, ask "Does this capture everything?"
• For Level C: analyze silently, present summary for acknowledgment
• Output format: structured markdown saved to docs/pert-workspace/project-context.md

User validation checkpoint: present the context analysis for review. The user may request additions or corrections.

Error recovery:

• If input document is ambiguous: ask clarifying questions
• If reference folder files are unreadable: skip with warning, proceed with available data

---

5. Phase 3 -- WBS + RBS Construction

Progressive disclosure: Read references/pmi-methodology.md before constructing agent prompts. For Level A interactions, the agents use this document to explain decomposition decisions and RACI concepts.

Launch two agents in parallel:

WBS Builder

Agent(model="opus")

Agent prompt must include:

• File to read: docs/pert-workspace/project-context.md
• The interaction level chosen by the user
• Target values (if any) for awareness during decomposition
• Reference: references/pmi-methodology.md -- specifically sections on WBS, 8/80 rule, rolling wave planning, 100% rule
• Instructions:
  • Propose hierarchical decomposition: Phase > Work Package > Activity
  • Apply 8/80 rule (no work package < 8h or > 80h)
  • Identify dependencies between activities
  • For Level A: propose one phase at a time, explain each decomposition decision with PMI reasoning, include "Methodology Applied" section in output
  • For Level B: propose complete WBS in one pass, highlight 8/80 borderline cases
  • For Level C: generate complete WBS autonomously
• Output format: structured markdown saved to docs/pert-workspace/wbs-draft.md

RBS Builder

Agent(model="sonnet")

Agent prompt must include:

• File to read: docs/pert-workspace/project-context.md
• The interaction level chosen by the user
• Reference: references/interaction-levels.md -- for RBS-specific behavior per level
• Instructions:
  • Extract roles from context (or ask if not found)
  • Assign roles to teams
  • Define competencies and responsibilities
  • Propose resource allocation per work package
  • Mark each role as billable or non-billable
  • For Level A: explain RACI concepts, discuss billable vs non-billable distinction
  • For Level B: propose complete RBS with team assignments
  • For Level C: generate complete RBS autonomously
• Output format: structured markdown saved to docs/pert-workspace/rbs-draft.md

User validation checkpoint: present both WBS and RBS for review. If the two artifacts are inconsistent (e.g., RBS references phases not in WBS), reconcile before presenting.

Backtrack: if context is incomplete, return to Phase 2 for additions.

---

6. Phase 4 -- Risks + Estimates

Progressive disclosure: Read references/pmi-methodology.md sections on Risk Management and Three-Point Estimation. Read references/interaction-levels.md for phase-specific behavior guidance.

Launch two agents sequentially (Risk Analyst must complete before Estimator begins).

Risk Analyst

Agent(model="sonnet")

Agent prompt must include:

• Files to read: docs/pert-workspace/project-context.md, docs/pert-workspace/wbs-draft.md, docs/pert-workspace/rbs-draft.md
• The interaction level chosen by the user
• Reference: references/pmi-methodology.md -- sections on Risk Identification, P x I Matrix, Response Strategies, Contingency vs Management Reserve
• Instructions:
  • Identify risks per phase/activity
  • Evaluate Probability (1-5) x Impact (1-5)
  • Propose strategy: Mitigate / Transfer / Accept / Avoid
  • Calculate contingency per risk
  • Propose management reserve (% of total PERT, default from config ManagementReservePct)
  • For Level A: introduce P x I matrix with examples, explain each response strategy, walk through contingency calculation
  • For Level B: propose complete risk register, highlight highest-priority risks
  • For Level C: generate complete risk register autonomously
• Output format: structured markdown saved to docs/pert-workspace/risk-register.md

User validation checkpoint: present risk register for review before proceeding to Estimator.

Estimator

Agent(model="opus")

Agent prompt must include:

• Files to read: docs/pert-workspace/project-context.md, docs/pert-workspace/wbs-draft.md, docs/pert-workspace/rbs-draft.md, docs/pert-workspace/risk-register.md
• The interaction level chosen by the user
• Target values (if any) for effort and/or duration
• Reference: references/pmi-methodology.md -- sections on Three-Point Estimation (PERT), Calibration Questions, Common Estimation Pitfalls, Statistical Confidence Intervals, Top-Down/Bottom-Up Reconciliation
• Instructions:
  • For each leaf activity: propose Best (O) / Most Likely (M) / Worst (P) estimates
  • Calculate PERT = (O + 4M + P) / 6 and sigma = (P - O) / 6 per activity
  • Rollup per phase and project total
  • Resource assignment per activity (role codes from RBS)
  • If targets provided and deviation > 20%: initiate guided reconciliation
    • Analyze causes (scope, estimates, resources, dependencies)
    • Propose adjustments (scope deferral, estimate recalibration, resource reallocation, risk reassessment)
    • Iterate until convergence or explicit user acceptance
    • Document reconciliation log
  • For Level A: explain three-point calibration using calibration questions from methodology reference, show PERT formula derivation, explain sigma and confidence intervals
  • For Level B: propose O/M/P ranges for all activities, show PERT totals and CI, ask "Any estimates you'd like to adjust?"
  • For Level C: generate all estimates autonomously, auto-reconcile if delta <= 20%, flag only if delta > 20%
• Output format: structured markdown saved to docs/pert-workspace/estimates-draft.md

User validation checkpoint: present estimates for review. Include reconciliation analysis (if targets were provided) showing delta and adjustment rationale.

Backtrack: if WBS needs restructuring after estimation (e.g., discovered missing activities), return to Phase 3.

---

7. Phase 5 -- Excel Generation

Progressive disclosure: Read references/excel-schema.md before constructing the agent prompt. This is the machine-readable reference for the JSON structure and Excel column/formula schema.

Agent(model="sonnet")

Agent prompt must include:

• Files to read: all validated markdown artifacts:
  • docs/pert-workspace/project-context.md
  • docs/pert-workspace/wbs-draft.md
  • docs/pert-workspace/rbs-draft.md
  • docs/pert-workspace/risk-register.md
  • docs/pert-workspace/estimates-draft.md
• The pmo-pert-estimate configuration from CLAUDE.md (effort_unit, duration_unit, primary_color, currency, period_type, avg_rate, management_reserve_pct)
• Target values (if any)
• Reference: references/excel-schema.md -- the complete JSON input schema and column/formula definitions per sheet
• Instructions:
  • Construct the structured JSON matching the schema in references/excel-schema.md (config, roles, phases with work_packages and activities, resource_allocation, risks, targets)
  • Save JSON to docs/pert-workspace/excel-input.json
  • Invoke the generator script:

    cd <skill-dir>/scripts && python3 generate_excel.py --input <json_path> --output <output_path>
    

  • The output path is {OutputDir}/pert-estimate.xlsx (from config)
• Output: the generated Excel workbook

Error Recovery Protocol

If the Python script fails during execution:

1. Capture the full Python traceback
2. Analyze root cause:
   • Missing field in JSON: identify which field, add it with appropriate value
   • Wrong type (e.g., string where number expected): fix the type in JSON
   • Formula error: check against references/excel-schema.md patterns
   • openpyxl API error: report the specific cell/sheet causing the issue
3. Re-generate the JSON with fixes and retry (maximum 2 attempts)
4. If unrecoverable after 2 retries: report the error to the user with full context (traceback, last valid artifacts, specific cell/sheet if known) for manual intervention

---

8. Phase 6 -- Validation

Agent(model="sonnet")

Agent prompt must include:

• File to read: the generated Excel workbook at {OutputDir}/pert-estimate.xlsx
• Reference: references/excel-schema.md for expected formula patterns and cross-references
• Reference: references/template-criteria.md for structural requirements
• Instructions:
  • Open the workbook with openpyxl
  • Run the verification checklist:
    • All 5 sheets present and correctly named (WBS, Timeline, Resources, Risks, Summary)
    • Formula cells contain formulas, not hardcoded values
    • Cross-reference inter-sheet links resolve correctly (no #REF! errors)
    • Consistent formatting (font, colors, borders, number format per row type)
    • SUM rollups match actual child ranges
    • PERT = (O+4M+P)/6 present on every appropriate row
    • sigma = (P-O)/6 present where expected
    • No empty input cells (all leaf activities have O/M/P values)
    • Billable flags present and BILLABLE EFFORT formulas correct
    • TOTAL row has global SUMs for effort and billable effort
  • If issues found: fix programmatically and re-validate (maximum 3 iterations)
  • If 3 iterations exhausted with remaining issues: present the issues to the user with recommendations
• Output: validation report + corrected Excel (if fixes were applied)

Present final workbook to user with summary statistics (total phases, activities, PERT effort, PERT duration, CI ranges, number of risks, contingency, adjusted estimate).

---

9. Progressive Disclosure Rules

Reference documents are loaded ONLY when entering the relevant phase. This keeps agent context fresh and focused.

Phase	Documents to Read
Phase 0	(none -- in-skill setup)
Phase 1	(none -- in-skill setup)
Phase 2	references/workflow.md Phase 2 section
Phase 3	references/pmi-methodology.md (WBS, 8/80, rolling wave sections)
Phase 4	references/pmi-methodology.md (Risk, PERT, Reconciliation sections)
Phase 5	references/excel-schema.md (full document)
Phase 6	references/excel-schema.md + references/template-criteria.md

For Level A interactions, additionally read references/interaction-levels.md at the start of each phase to retrieve phase-specific formative behavior guidance.

Never load all reference documents at the beginning of the workflow.

---

10. Dynamic Adaptation

The interaction level is not rigid. Monitor user behavior and adapt per-phase.

Upward Shift (toward more guidance)

Trigger	Action
Level B/C user asks "why?" or "what does X mean?"	Switch to Level A explanations for that topic. Ask: "Would you like me to switch to full guidance mode for this phase?"
Level B/C user requests methodology explanation	Provide PMBOK context from references/pmi-methodology.md, offer to stay at Level A
Level B/C user expresses uncertainty about estimates	Use calibration questions from references/pmi-methodology.md Section 2

Downward Shift (toward less guidance)

Trigger	Action
Level A user consistently responds "ok" / "looks good" without engagement	Suggest: "You seem comfortable -- want me to propose complete artifacts instead of step-by-step?"
Level A user modifies estimates confidently	Reduce explanation density
User explicitly requests faster pace	Switch to requested level

Adaptation is per-phase, not global. The agent never downgrades without suggesting it first. The agent may upgrade silently (providing more context when asked) without formally announcing a level change.

---

11. Error Recovery Summary

Phase 5 -- Excel Generation Errors

1. Capture Python traceback
2. Analyze root cause (missing field, wrong type, formula error, openpyxl API issue)
3. Re-generate JSON with fixes (maximum 2 retries)
4. If unrecoverable: report to user with full context

Phase 6 -- Validation Errors

1. Identify failing checks from verification checklist
2. Apply programmatic fixes (correct formula, fix cross-reference, add missing value)
3. Re-validate (maximum 3 iterations)
4. If unresolved: present remaining issues with recommendations

General Error Recovery

• If input document path is invalid: ask user to correct
• If workspace directory already exists: ask whether to reuse or create fresh
• If template validation fails: show specific errors, offer fallback to bundled template
• If CLAUDE.md is read-only: report error, ask user to fix permissions
• If WBS violates 8/80 rule: flag violations, propose splits/merges
• If no roles found in context: ask user to provide role list
• If reconciliation fails to converge after 3 iterations: present delta to user, ask for explicit scope/target adjustment

---

12. Artifact Chain

Each artifact is produced, validated by the user, and then passed as input to the next phase. The agent working on step N receives only artifacts 1..N-1 to keep context fresh.

Step	File	Format	Produced by
1	docs/pert-workspace/project-context.md	Markdown	Phase 2 (Opus)
2	docs/pert-workspace/wbs-draft.md	Markdown	Phase 3 (Opus)
3	docs/pert-workspace/rbs-draft.md	Markdown	Phase 3 (Sonnet)
4	docs/pert-workspace/risk-register.md	Markdown	Phase 4 (Sonnet)
5	docs/pert-workspace/estimates-draft.md	Markdown	Phase 4 (Opus)
6	docs/pert-workspace/excel-input.json	JSON	Phase 5 (Sonnet)
7	{OutputDir}/pert-estimate.xlsx	Excel	Phase 5 (Script)

---

13. Final Checklist

Before considering the estimation complete, verify:

• All 5 Excel sheets present (WBS, Timeline, Resources, Risks, Summary)
• PERT formulas =(O+4M+P)/6 on every appropriate row
• SUM rollups correct (exact child ranges)
• sigma = (P-O)/6 present where expected
• Cross-sheet references resolve (no #REF! errors)
• Input columns have values, formula columns have formulas (never inverted)
• Consistent formatting (phase/WP/leaf/total row styles)
• Billable flags correct, BILLABLE EFFORT calculated
• Resources matrix with teams, billable metadata, SUMPRODUCT formulas
• Risk register with P x I formula, priority IF, contingency, management reserve
• Summary with sigma total (SQRT SUMPRODUCT), CI 68%/95%, adjusted PERT
• TOTAL row with global SUMs for effort and billable effort
• 8/80 rule respected in WBS decomposition
• All intermediate .md artifacts saved in docs/pert-workspace/
• Interaction level respected throughout the workflow
• Reconciliation executed if targets were provided and delta > 20%