In Lesson 2, you learned the architecture: the SKILL.md file has three sections (Persona, Questions, and Principles) and each section performs a distinct function. Persona defines who the agent is. Questions defines what it handles. Principles defines how it decides. The Agent Skills standard (agentskills.io) defines the SKILL.md format. PQP is our methodology for what goes inside to produce enterprise-grade domain agents. You learned why specificity matters, why identity governs ambiguous situations more reliably than rules, and why the out-of-scope boundary is as important as the in-scope list. What you did not see was all three sections working together in a single, coherent document.
This lesson shows you that. The PQP Framework in Practice means reading a production-ready SKILL.md as a whole: tracing how the sections interact, identifying what makes it trustworthy, and understanding what each design choice prevents. The example is a financial research agent. Financial services is a useful domain for this illustration because the failure modes are concrete (fabricated numbers, misplaced confidence, regulatory exposure), the professional standards are legible, and the range of in-scope and out-of-scope territory is clearly bounded. What you learn here applies equally to legal, clinical, architectural, and operational domains.
The example below is simplified relative to what a production organisation would deploy. A production SKILL.md for a financial research function would be two to four times longer, with more detailed data source specifications, organisation-specific output templates, jurisdiction references, and a fuller escalation taxonomy. But the structure is complete. Every section is present, every quality signal is visible, and the document would produce a reliable, trustworthy agent as written. Read it as an architectural example rather than a sizing guide.
The following is the full SKILL.md for a financial research agent, presented in sections with annotations after each one.
Note: A production SKILL.md includes YAML frontmatter (at minimum name and description fields, as defined by the agentskills.io specification) before the body content below. The frontmatter is how the platform discovers and activates the skill. This example shows only the body content: the PQP sections that encode domain expertise.
Annotation: Persona: This opening establishes three things: professional standing (senior analyst), communication register (precise, economical), and epistemic standard (evidence-grounded). Each one governs how the agent behaves across every interaction.
Notice how the section handles the most important constraint: "You are not an investment adviser." It does not say "Do not give investment recommendations." It states a professional identity. The difference is significant. A rule can be argued around: a user who frames their request cleverly enough may elicit advice that is technically not labelled as a recommendation. An identity constraint cannot be argued around: an agent that is a senior analyst and is not an investment adviser will decline to produce investment recommendations not because it is following a rule but because that is who it is. Identity governs ambiguous situations more reliably than any finite list of instructions.
The sentence "You flag uncertainty with confidence rather than apologetically" is functional rather than stylistic. It tells the agent how to handle incomplete data; not by hedging vaguely or producing caveated analysis, but by stating clearly what is known and what is not. This shapes thousands of individual responses without requiring a specific rule for each scenario.
Annotation: Questions: Each in-scope category is specific enough to govern actual behaviour. "Aggregate and summarise publicly available market data from your configured data sources" is actionable: the agent knows what to pull and where to pull it from. "Be helpful with market research" is not actionable: it gives the agent no guidance on data sources, output format, or how to handle cases where sources disagree.
The out-of-scope section is doing as much work as the in-scope list. Notice two things about how it is written. First, it is explicit and complete: investment recommendations, price predictions, non-public information, employee compensation, and unapproved external distribution are all named. An agent without explicit out-of-scope boundaries will attempt to help in areas where help is harmful, producing confident-sounding output in territory where it has no grounded expertise. Second, the final sentence: "For out-of-scope requests, tell the user clearly why the request is outside your remit and suggest an appropriate alternative where one exists": converts what could be a refusal into positive guidance. The agent does not simply decline; it tells the user what to do instead. This is the difference between a boundary that is a dead end and a boundary that routes users toward the right resource.
Annotation: Principles: The Source Integrity principle addresses the most dangerous failure mode for a financial agent. A model trained on large amounts of financial data can produce plausible-looking figures: revenue numbers, market capitalisations, deal valuations: that are drawn from training memory rather than connected, current data sources. In a general-purpose assistant, this is an inconvenience. In a financial research context where decisions are made on the basis of these numbers, it is a serious risk. The principle does not say "be accurate." It says: if you cannot ground this figure in a configured connector or a document the user provided, tell them you do not have a grounded source. The instruction is specific enough that the agent knows exactly what to do in the problematic case.
Uncertainty Calibration is infrastructure for trust. Financial professionals who work regularly with this agent will learn what each phrase means. "The data indicates" tells them this is directly grounded: they can act on it with confidence. "It is worth considering whether" tells them this is a hypothesis the data raises but does not resolve: they should seek additional evidence before acting. A shared vocabulary between agent and professional user means outputs are auditable: the professional can read the agent's language and know immediately what degree of reliance is appropriate. This is not achievable with either confident declarative statements (which obscure uncertainty) or uniform hedging (which renders outputs useless). The four levels are a calibrated middle ground.
The Escalation principle defines human handoff conditions precisely. "Board/executive presentations", "transactions above £50M", and "regulatory compliance claims" are specific enough that the agent can recognise them: and specific enough that a professional user reviewing the agent's behaviour can verify whether the threshold was applied correctly. Compare this to a generic escalation principle: "escalate complex matters to the appropriate team." The generic version gives the agent no guidance and gives the professional no way to audit whether escalation was applied correctly.
Reading the three sections together reveals something that reading them separately does not: the SKILL.md is a coherent professional specification, not a list of settings. Every design choice connects to a failure mode it prevents.
The Persona prevents the agent from behaving as a general-purpose assistant in a professional context where general-purpose helpfulness is harmful. A financial research agent that tries to be maximally helpful, answering questions beyond its data, producing investment-flavoured analysis because the user seems to want it, creates liability. The Persona's identity constraints close that failure mode.
The Questions section prevents scope creep in both directions. Without an explicit in-scope list, users do not know what the agent is for and will underuse it. Without an explicit out-of-scope list, users will ask it questions it cannot answer well and receive confident-sounding output that is unreliable. The section defines the envelope of reliable performance.
The Principles section prevents the most dangerous operational failures: fabricated numbers (Source Integrity), outdated analysis presented as current (Recency Transparency), confident statements for inferential conclusions (Uncertainty Calibration), and autonomous action in situations that require human judgment (Escalation). None of these is a generic quality standard. Each one addresses a specific failure mode that professionals in this domain will recognise from experience.
The production version of this SKILL.md would be longer. It would specify data sources by name and API endpoint, include organisation-specific research brief templates as appendices, list jurisdictions in scope, define a fuller escalation taxonomy with specific roles and queues, and address edge cases discovered during the shadow mode evaluation period. But the structure would be identical. Persona, Questions, Principles: in that order, with that level of specificity in each section.
Use these prompts in Anthropic Cowork or your preferred AI assistant to apply what you have learned.
What you're learning: How to read a Persona as a professional specification rather than a description. The analysis of what is missing is often more instructive than the analysis of what is present: it forces you to think about failure modes that the current text does not address.
What you're learning: The difference between a rule and an identity constraint becomes clear when you attempt to write both for the same constraint. The exercise of converting a rule into an identity statement, "do not give medical advice" into "you are not a clinician and do not substitute for clinical judgment", makes the functional difference concrete.
What you're learning: Writing an out-of-scope section with positive redirection is significantly harder than writing a refusal list. The exercise forces you to think about your domain's boundary conditions (where reliable expertise ends and where professional liability begins) and to design a response for each one that is useful rather than merely defensive.
This lesson presents a complete, annotated SKILL.md for a financial research agent: showing all three sections (Persona, Questions, Principles) working together as a coherent professional specification. The annotated example reveals the quality signals that distinguish a production-ready SKILL.md from a minimal or generic one. The SKILL.md is not a list of settings; it is a document where every design choice prevents a specific failure mode.
📋Quick Reference
Access condensed key takeaways and quick reference notes for efficient review.
Free forever. No credit card required.
Ask