From clever prompt to maintainable asset

Lesson 03.1

Know what you are actually building

A prompt describes a task now. A skill holds reusable methods and constraints. A knowledge source supplies relevant facts. A workflow coordinates steps and tools. An agent may choose actions dynamically to accomplish a goal. These concepts can work together, but should not be treated as synonyms.

Reuse is not automatically a reason to add autonomy. A documented checklist or deterministic workflow may solve the problem more safely and with less maintenance.

• Prompt — what needs to be done in this instance
• Skill — how the task should repeatedly be performed
• Knowledge — the approved evidence and context
• Workflow — which steps happen, in which order
• Agent — when the system selects and performs steps using tools

Lesson 03.2

The ten-part specification

A skill should be readable by a new team member. Put reasons next to non-obvious rules, and record what a change was meant to improve. Undocumented prompts turn into operational liabilities when their original author leaves.

• 1. Identity: name, plain-language summary, owner and version.
• 2. Purpose: the one outcome it owns.
• 3. Use / do not use: triggers, exclusions and redirects.
• 4. Inputs: required records, approved sources, validation and permissions.
• 5. Method: decision sequence and criteria.
• 6. Evidence: how material claims can be checked.
• 7. Boundaries: prohibited actions, refusals and escalation.
• 8. Output: fields, definitions, examples and accepted formats.
• 9. Quality checks: what must pass before release.
• 10. Maintenance: dependencies, change log, evaluation cases and review date.

Lesson 03.3

Worked example: Support Answer Reviewer

Purpose: compare a drafted support answer against approved help content and identify factual risk before publication.

Inputs: draft answer, approved article, article ID, last-reviewed date and any authorised product-change notice.

Method: extract claims; match them to evidence; flag unsupported or contradicted claims; check for missing conditions; propose changes; record reviewer action.

Stop condition: if the approved source is missing or two authoritative documents disagree, return “blocked”, cite the conflict and route it to the content owner.

Output: claim, source reference, assessment (supported / contradicted / not verifiable), severity, proposed revision and owner action.

Lesson 03.4

Copy a starter skill specification

# Skill: [Name]
Owner: [Person or team]
Version: [x.y]
Last reviewed: [date]
Next review: [date]

## Purpose
Help [audience] accomplish [one outcome].

## Use when / do not use when
- In scope: [conditions]
- Out of scope: [conditions and redirect]

## Required inputs
- [required input and validation]
- [approved sources and authority hierarchy]

## Method
1. Validate input and source status.
2. [decision steps]
3. Compare findings against evidence and alternatives.
4. If blocked, [stop / escalation route].
5. Return the defined output.

## Non-negotiables
- Never invent evidence or claim an unchecked verification.
- Label inference and uncertainty.
- Never [domain-specific forbidden action].

## Output contract
[fields, accepted values, formatting and evidence rules]

## Quality checks
[checks that must pass before delivery]

## Tests and maintenance
Evaluation cases: [location]
Dependencies: [tools, sources, configuration]
Approvals: [required human gates]
Change log: [date, reason, test result]