← The Agentic AI Series

Companion to the Series  ·  Stage 1 of 3

Agentic AI PM Skill Package

The lifecycle, as working aids a PM can invoke at the moment of need.

For Claude Cowork  ·  by Yoram Friedman

Seven skills covering the agentic PM lifecycle, distilled from the four books in the series. Install in Claude Cowork and invoke at the moment of need. Each skill walks the PM through the work for one phase and produces the artifacts the team will defend in the room.

The skills are working aids. The books are where the reasoning lives. When a skill needs depth, it points back to the chapter.

Download the seven skills → View on GitHub

What is in Stage 1

Stage 1 covers the PM lifecycle: one front-door router, five phase skills, and one portfolio-level strategic skill.

The five phases, at a glance
1
Decide
Decide whether an agent is the right solution before engineering begins. Most expensive failures are prevented here.
Deliverables
  • Suitability record (six criteria)
  • Context sufficiency map
  • Distribution gap analysis
  • Cost model with break-even
  • Go-or-no-go memo
Questions answered
  • Is the problem bounded with tolerable error?
  • Will the agent receive the relationships and semantics it needs?
  • Does the math work at projected volume?
  • Should we actually build this?
2
Design
Design the two products: the agent and the supervisory system. Runtime behavior engineered deliberately, not by default.
Deliverables
  • Declared system type
  • Four runtime artifacts
  • Consequence classification
  • Trust scaffold
  • Channel 2 composition
  • Adversarial defense plan
Questions answered
  • What can the agent do unilaterally, what needs approval?
  • When a human approves, what information do they see?
  • How do we reverse the agent when it is wrong?
3
Eval
Prove readiness through evidence. Replaces the single pass-fail gate with distribution, projection, and state-validation.
Deliverables
  • Pass@K with worst-slice variance
  • Compound probability projection
  • Semantic-vs-state validation
  • Coverage statement
  • Adversarial suite results
  • All-owner readiness memo
Questions answered
  • When we run the eval ten times, how often does it pass?
  • What is the end-to-end success rate, not per-step?
  • Did the target system actually change?
  • Are all four owners willing to sign?
4
Observe
Measure what the agent is actually doing in production. Continuous front end of Operate.
Deliverables
  • Six observation instruments
  • Data-layer observation
  • Operational guardrails
  • Supervised-vs-bypass report
  • Supervisory engagement metrics
  • Stratified performance report
Questions answered
  • Did the agent do what the user intended?
  • Did it act outside its declared boundary?
  • Are humans still supervising or nodding along?
  • Is it calibrated, or confidently wrong?
5
Operate
Act on what Observe is showing. Governance runs at runtime. Drift, silent degradation, and retirement are named stages.
Deliverables
  • Drift detection (six vectors)
  • Sealed Decision Artifact policy
  • Constitutional runtime rules
  • Adaptive governance rule set
  • Instrument half-life policy
  • Affected-person audit view
Questions answered
  • Is the agent still doing the thing it did at launch?
  • Has the model underneath shifted?
  • Are rules running at runtime or in review meetings?
  • Can we produce the full decision history?
Router  ·  agentic-pm-lifecycle  ·  Diagnoses where the PM is and dispatches to the right phase skill.
Cross-phase  ·  agentic-pm-behavior-governance  ·  Portfolio strategy across multiple agents.
Router
agentic-pm-lifecycle
Diagnoses the PM’s situation and dispatches to the right specialized skill. Use when you are not sure which phase you are in.
Read
Phase 1  ·  Discover & Decide
agentic-pm-discover-decide
Decide whether to build the agent at all. Suitability record (six criteria), context sufficiency map, distribution gap analysis, cost model with break-even, go-or-no-go memo. Includes a fully worked example for a refund agent.
Read
Phase 2  ·  Design
agentic-pm-design
Specify runtime behavior before engineering builds. The four runtime artifacts, the two briefs, Channel 2 composition, consequence classification, trust scaffold, adversarial defense plan, Sealed Decision Artifact spec.
Read
Phase 3  ·  Eval
agentic-pm-eval
Prove readiness through evidence. Pass@K with worst-slice variance, compound probability projection, semantic-vs-state validation, coverage statement, adversarial suite, LLM-as-judge calibration, model-version policy, all-owner readiness memo.
Read
Phase 4  ·  Observe
agentic-pm-observe
Measure what the agent is actually doing in production. Six observation instruments, data-layer observation, operational guardrails (ceilings, kill switch, circuit breaker, burn-rate), supervised-vs-bypass report, supervisory engagement metrics, stratified performance report.
Read
Phase 5  ·  Operate
agentic-pm-operate
Long-horizon governance. Drift detection (six vectors), Sealed Decision Artifact policy, instrument half-life, constitutional runtime rules, affected-person audit view, adaptive governance rule set, currency question, external audit on held-out sample.
Read
Cross-phase  ·  Strategic posture
agentic-pm-behavior-governance
Portfolio strategy across multiple agents. Maturity assessment, consolidation roadmap, three-stage governance arc (hand-built artifacts → control plane → platform primitive).
Read

Who this is for

A senior product manager (10+ years) who is fluent in classical product management and is now building, evaluating, or operating an agentic AI product. The skills assume PM fluency on the conventional craft. Each phase skill includes a “first-timer foundations” section that teaches the constructs new to agentic work (golden datasets, Pass@K, the four runtime artifacts, drift vectors, the Sealed Decision Artifact, the constitutional runtime layer).

If you are an enterprise PM team adopting agentic AI and looking for a shared discipline, the package is designed to be invoked phase by phase as the team’s work moves through the lifecycle.

How to install

Each skill is packaged as a .skill zip. To install all seven:

1. Download the latest release from GitHub.
2. In Claude Cowork, go to Settings → Capabilities → Skills.
3. Drag each .skill file into the install area.
4. The skills appear in your available skills list and can be invoked by name or by natural-language trigger.

To install one skill at a time, download only the .skill zip for that skill.

The four books behind the skills

The skills condense doctrine that is fully developed in the four books in the series. When a skill needs depth, it points back to a specific chapter. The books are free to read online.

Book 1  ·  The PM
Agentic AI for Busy Product Managers
Book 2  ·  The Failures
Why Agentic AI Products Fail
Book 3  ·  The Team
The Agentic AI Team
Book 4  ·  The Practitioner
The Agentic AI Practitioner

What is coming next

Stage 1 covers the lifecycle. Two more stages will follow:

Stage 2: Team Collaboration. Five skills built from The Agentic AI Team: the collaboration grid, seam audit, agent-as-team-member onboarding, fleet governance, team skill protection.

Stage 3: Practitioner. Five skills built from The Agentic AI Practitioner: proficiency check, model dossier, configuration interview, deliberate-practice loops, first-month on-ramp.

Stage 1 covers the PM’s work. Stage 2 covers the team that ships with the PM. Stage 3 covers the individual who has to stay competent operating the agent over years. Each stage will be released as a separate skill bundle.

The skills are working aids, not condensed books. They condense doctrine into structured prompts and artifact templates the PM can use at the moment of need. The books are where the reasoning lives. The skills are how the reasoning shows up in the room when there are forty minutes to write the memo.