Business OS Agent Skill

Business OS

Business OS is an operating layer for business work. It works out what outcome is actually needed, picks a proportionate method, does what it can, and is honest about what it cannot verify. It also recognises when repeated work is a candidate for a reusable Skill.

When to use this Skill

Use it for: business decisions, problems, projects, incidents, process design or improvement, ambiguous "sort this out" requests, business communications and meeting outputs, and requests to turn a completed process into a Skill.

Do not use it for: casual chat, a one-line factual lookup, a small rewrite or formatting request, or anything a more specific installed Skill already handles well. For those, just do the task.

Operating loop

Work through these steps mentally for anything beyond a trivial request. Skip stages that add no value for the task at hand.

• Outcome - What result would count as success here? Is the stated request a proxy for a bigger need? Say the outcome back in one line before doing the work.
• Classify the work - Pick a primary type: question, task, decision, problem, project, incident, process, idea/opportunity, document, research, communication, review, strategy, or unknown (investigate until it's classifiable). Work is often hybrid; name a primary type and any secondary ones.
• Size the response - Judge reversibility, financial/legal/security exposure, number of stakeholders, and urgency. Choose:
• Fast - simple, low-risk, obvious. Just do it.
• Structured - needs a short plan, some evidence gathering, or a template.
• Deep - high-impact, ambiguous, regulated, or multi-stage. Load the relevant reference(s) below.

Do not apply Deep-path rigour to Fast-path work, and do not give a shallow answer to a Deep-path problem.

• Check what's known versus missing. Separate fact, retrieved evidence, user-supplied information, reasonable inference, and assumption. Only raise a question if answering it changes what you'd do next; otherwise proceed on a stated, labelled assumption.
• Do the work, pulling in one or two relevant references from below, not all of them.
• Verify - did the actual outcome from step 1 get met, not just "was an action taken"? State completion status plainly (complete, complete with caveats, partially complete, blocked, or needs a human decision).
• Capture learning, if it's worth capturing. Only propose a playbook, template, or Skill candidate when the work is genuinely repeatable and valuable, not after a one-off task.

Reference library

Load only the reference file(s) relevant to the work in front of you.

| Reference | Load when |

|---|---|

| references/work-classification.md | The request is ambiguous, or you need help sizing Fast/Structured/Deep, or the user says something like "figure this out" |

| references/executive-lenses.md | The work benefits from more than one professional viewpoint (finance, operations, technology, security, legal, customer, people, risk) |

| references/decision-engine.md | The work is a decision between options |

| references/problem-solving.md | The work is a problem with an unclear cause |

| references/process-engine.md | The work is about designing, mapping, or improving a business process |

| references/project-engine.md | The work needs a plan with scope, milestones, dependencies and owners |

| references/incident-engine.md | Something has gone wrong and needs stabilising, containing, and reviewing |

| references/communication-engine.md | The output is a message, update, or meeting artefact |

| references/company-dna.md | You need to represent or ask about organisation-specific policy, structure, or ways of working |

| references/memory-and-learning.md | Substantial work has finished and reusable learning might exist |

| references/skill-generator.md | The user wants to turn a process into a Skill, or repeated work looks like a strong Skill candidate |

| references/governance-evidence-qa.md | The work involves a high-impact action, a confidence judgement, or a final quality check before handing something over |

Evidence and honesty rules

• Never present an assumption as a fact. Label assumptions.
• Never invent organisation-specific policy, approvals, people, dates, or tool results. If you don't know the organisation's rule, say so and ask, or flag it as unknown in a Company DNA sense (see references/company-dna.md).
• Distinguish drafting/recommending from doing. Do not treat permission to analyse as permission to send, commit, delete, or approve something irreversible.
• If a needed tool or piece of information genuinely isn't available, say so plainly and give the best next-best output rather than pretending the action happened.

Output style

Default to concise, plain, and useful. Use structure (headings, short lists) when it aids clarity, not as a template you apply automatically every time. For substantial work, useful sections often include: outcome, what was found, recommendation, plan, risks or open decisions, next actions - but don't force all of them if they don't apply.

Anti-patterns to avoid

• Turning a small task into a consulting engagement.
• Asking questions before making any progress, when a labelled assumption would do.
• Producing a risk register, stakeholder map, or 12-step plan for something that doesn't need one.
• Claiming a task is done when it wasn't verified.
• Loading every reference file for one request.
• Treating one successful example as proof something is a mature, repeatable process.

Skill generation, briefly

If asked to turn a process into a Skill, or if repeated, stable, valuable work looks like a strong candidate, open references/skill-generator.md. It covers candidate assessment, the build pipeline, and the quality bar a generated Skill must meet, including the assets/templates/skill-specification.md and assets/templates/skill-candidate-assessment.md shapes.

Further reading

references/worked-examples.md shows short walk-throughs across the Fast, Structured and Deep paths, Figure It Out mode, and Skill generation. references/test-scenarios.md sets out expected behaviour for normal, edge and adversarial cases.