PSM-AI cheat sheet

Supervised learning maps labeled inputs to outputs; unsupervised learning finds structure in unlabeled data; reinforcement learning optimizes actions against a reward signal.

A large language model (LLM) is a next-token predictor: output is generated one token at a time from the tokens currently in the context window, and a token is roughly 3-4 English characters.

Context window = maximum tokens the model can attend to at once, covering system prompt, user prompt, retrieved documents, and the generated answer; content beyond it is invisible to the model.

Temperature and top-p control sampling randomness: low temperature produces repeatable, focused output; high temperature produces variety plus more factual and format errors.

Retrieval-augmented generation (RAG) inserts retrieved documents into the prompt; it changes the input, not the model weights, so answer quality depends on retrieval quality.

Fine-tuning adjusts model weights on examples to change style, format, or task behavior; it does not guarantee factual grounding and does not automatically update with new facts.

Hallucination is fluent output unsupported by provided context or verifiable fact; it is a property of generative models, so verification is a process control, not a prompt fix.

Precision = TP/(TP+FP); Recall = TP/(TP+FN); F1 = harmonic mean of the two, used to evaluate classifiers where both false positives and false negatives carry cost.

Training fits parameters on data and happens periodically; inference is per-request generation using fixed weights; data drift occurs when live inputs diverge from the training distribution.

The Scrum Master is accountable for coaching, teaching, facilitating, mentoring, and impediment removal; AI can accelerate drafting and analysis, but accountability for outcomes cannot transfer to a tool.

Decision rule: delegate drafting, summarizing, clustering, translating, and option generation to AI; retain prioritization, conflict resolution, feedback delivery, and personnel decisions with humans.

A six-part prompt (role, task, context, constraints, output format, example) reliably outperforms short prompts and reduces rework on generated artifacts.

RAG over team-owned artifacts (Definition of Done, working agreements, prior retrospective notes) produces team-specific answers that generic prompting cannot match.

Verification rule: no AI-generated claim enters a Scrum artifact or stakeholder communication until checked against a primary source and attributed.

In retrospectives, AI can cluster anonymized observations into themes, but the team must confirm the themes; otherwise the tool manufactures consensus that the team never reached.

AI-suggested velocity targets are invalid: velocity is a planning observation, not a goal, and machine-generated comparisons distort forecasting and invite gaming.

For impediments, use AI for search, summarizing blocker descriptions, and ticket routing; the Scrum Master still escalates and resolves organizational blockers personally.

Sprint Review: AI can summarize the increment and stakeholder comments, but only stakeholders judge whether the increment delivers value.

Team AI literacy baseline: what the model can and cannot do, token and cost awareness, verification habits, data-handling rules, and how to report an AI failure.

Add AI-specific conditions to the Definition of Done: human review completed, sources cited, no confidential or personal data placed in prompts, tests written for AI-generated code.

Psychological safety is a precondition: if reporting an AI error carries blame, errors go hidden and the team loses the ability to inspect and adapt.

Treat AI adoption as a Scrum experiment: state a hypothesis, run a thin slice inside a timebox, inspect the effect in the retrospective, then keep, change, or drop.

A team-owned prompt and pattern library keeps output consistent, shortens onboarding, and prevents each member from re-solving the same prompt problem.

Human-in-the-loop rule: every AI-generated artifact has a named human reviewer who owns its accuracy before it leaves the team.

Counter automation bias (uncritical acceptance of machine suggestions) with scheduled spot checks and periodic no-AI sessions that exercise human judgment.

Prohibit AI use as a velocity lever or individual surveillance mechanism; measure flow, quality, and outcomes instead of output volume per person.

Run AI adoption as a product: a prioritized AI adoption backlog with a written value hypothesis, a thin-slice pilot, and empirical inspect-and-adapt cycles.

Record a baseline (cycle time, rework rate, time-to-answer) before the pilot; without a baseline the value hypothesis cannot be proven or disproven.

Governance prerequisites before scaling: data classification scheme, approved tool list, acceptable-use policy, procurement and data-processing review, and a defined incident path.

EU AI Act obligations scale by risk tier (unacceptable, high, limited transparency, minimal); high-risk uses require risk management, data governance, logging, and human oversight.

Under GDPR, personal data placed in a prompt is processing: it needs a lawful basis, data minimization, and a processor agreement with the model vendor.

Diffusion sequence that scales: pilot team, then community of practice, then internal playbook, then policy, with each stage gated on evidence from the previous one.

Expect predictable resistance patterns, including fear of replacement, skill gaps, and tool fatigue; address them with transparency about role impact and explicit capability building.

Track adoption with active use, task cycle time, rework rate, and AI-related incident counts; seat licenses or purchase volume are not adoption measures.

When leaders request 'AI-driven velocity', reframe to outcome metrics and treat unresolved organizational impediments as the real constraint.

Confidentiality rule: assume prompts, uploaded files, and outputs may be retained or used for vendor training unless a contract states otherwise, so classify data before it enters a prompt.

Intellectual property: AI output may not be copyrightable and may reproduce training content; check licenses, vendor indemnities, and provenance before external publication.

Bias is inherited from training data; test outputs across affected groups and review for disparate impact instead of assuming neutrality.

Transparency rule: disclose AI assistance where it affects stakeholders or decisions, and label generated content so recipients can weigh it.

Accountability cannot be delegated to a model: a named human role stays answerable for every AI-assisted decision, including its errors.

Agentic and plugin-based tools introduce prompt injection and data exfiltration paths; grant least-privilege access and treat retrieved untrusted content as hostile instructions.

Model drift degrades real-world performance as conditions change; schedule periodic re-evaluation against a held-out sample and monitor error reports.

Escalation gate: high-impact use cases (hiring, credit, health, safety, performance management) require documented risk assessment and mandatory human review before use.

Supervised learning maps labeled inputs to outputs; unsupervised learning finds structure in unlabeled data; reinforcement learning optimizes actions against a reward signal.

An LLM predicts the next token from the current context window; a token is roughly 3-4 English characters, so cost, latency, and context limits all scale with token counts.

Temperature controls sampling randomness: low values give repeatable, focused output; high values give variety plus more factual and formatting errors.

RAG retrieves external documents into the prompt before generation; it changes the input, not the model weights, so quality depends on retrieval quality.

Fine-tuning adjusts model weights on example data to change style or task behavior; it does not guarantee factual grounding.

Hallucination is fluent output unsupported by provided context or verifiable fact; it cannot be fully prompted away, so verification remains a process control.

Precision = TP/(TP+FP); recall = TP/(TP+FN); F1 is their harmonic mean, used when both false positives and false negatives matter.

Data drift is the divergence of live input data from the training distribution, degrading model accuracy over time.

An agent combines an LLM with tools and an execution loop; its failure modes include wrong tool selection and compounding errors across steps.

The Scrum Master remains accountable for coaching, facilitation, teaching, mentoring, and impediment removal; AI accelerates work but never holds accountability.

Decision rule: AI handles drafting, summarizing, clustering, translating, and option generation; humans retain prioritization, conflict resolution, feedback delivery, and personnel decisions.

Prompt structure of role, task, context, constraints, output format, and example consistently outperforms short prompts.

RAG over team artifacts (Definition of Done, working agreements, past retro notes) yields team-specific answers that generic prompting cannot produce.

Verification rule: an AI-generated claim enters a Scrum artifact only after being checked against a primary source.

AI clustering of anonymized retrospective feedback is an input to team sense-making, not a substitute for the team confirming the themes.

AI-suggested velocity targets are invalid; velocity is a planning observation, not a goal, and machine comparisons invite gaming.

AI can search, summarize, and route impediment information, but the Scrum Master personally escalates and resolves organizational blockers.

Team AI literacy baseline: capability limits, token and cost awareness, verification habits, data-handling rules, and the path for reporting an AI failure.

Extend the Definition of Done with AI conditions: human review done, sources cited, no confidential or personal data in prompts, tests written for AI-generated code.

Psychological safety is a precondition; if AI errors are punished, they are hidden and the team cannot inspect and adapt.

Frame AI adoption as a Scrum experiment with hypothesis, timeboxed thin slice, and retrospective inspection before scaling.

A team-owned prompt library stabilizes output quality and reduces repeated prompt rework and onboarding time.

Human-in-the-loop rule: a named human reviewer owns the accuracy of each AI-generated artifact before external use.

Automation bias (uncritical acceptance of machine suggestions) is countered by scheduled spot checks and periodic no-AI sessions.

AI must not be used to raise velocity targets or monitor individuals; measure flow, quality, and outcomes instead.

Manage AI adoption as a product: a prioritized adoption backlog with a written value hypothesis and thin-slice pilot under inspect-and-adapt.

Capture a baseline (cycle time, rework rate, time-to-answer) before the pilot; without it the value hypothesis cannot be tested.

Governance prerequisites: data classification, approved tool list, acceptable-use policy, procurement and data-processing review, and an incident path.

EU AI Act obligations scale by risk tier (unacceptable, high, limited transparency, minimal); high-risk uses need risk management, data governance, logging, and human oversight.

Under GDPR, personal data in a prompt is processing and requires a lawful basis, minimization, and a processor agreement with the vendor.

Diffusion path: pilot team, then community of practice, then internal playbook, then policy, each stage gated on evidence.

Measure adoption by active use, task cycle time, rework rate, and incident counts; licenses purchased are not an adoption metric.

Reframe 'AI-driven velocity' requests into outcome metrics, treating unresolved organizational impediments as the binding constraint.

Assume prompts, uploads, and outputs may be retained or used for vendor training unless contracted otherwise; classify data before it enters a prompt.

AI output may not be copyrightable and may reproduce training content; check licenses, indemnities, and provenance before publication.

Bias is inherited from training data; test outputs across affected groups and review for disparate impact.

Disclose AI assistance where it affects stakeholders or decisions, and label generated content.

Accountability cannot be delegated to a model; a named human role stays answerable for each AI-assisted decision and its errors.

Agentic tools and plugins create prompt injection and exfiltration paths; apply least-privilege access and treat retrieved untrusted text as hostile instructions.

Model drift degrades performance as conditions change; schedule periodic re-evaluation and monitor error reports.

High-impact uses (hiring, credit, health, safety, performance management) require documented risk assessment and mandatory human review before deployment.