An AI skill packages a repeatable organizational method—inputs, instructions, evidence rules, guardrails, and output—so teams can reuse it across workflows.
In short: An AI skill is a reusable, governed method for completing a specific task—not merely a clever prompt. It defines the inputs, ordered instructions, evidence requirements, guardrails, and expected output so a team can apply the same organizational method across programs and workflows. Skills improve consistency and auditability, but people still verify evidence and own consequential decisions.
In short: An AI skill packages organizational know-how into a method that can be reused. Examples include building a data dictionary, reviewing a theory of change, mapping indicators to standards, analyzing open-ended evidence, or preparing a cited reporting brief.
A prompt asks for an answer once. A skill preserves how the work should be done: what information is acceptable, what steps must occur, how uncertainty is treated, what the output must contain, and what a person must verify.
Use five parts: Definition → Instruction → Evidence → Guardrails → Output.
| Part | Question it answers | Example |
|---|---|---|
| Definition | What exact job does this skill perform? | Define one outcome indicator for use across three programs. |
| Instruction | What ordered method must it follow? | Identify construct, unit, population, calculation, cadence, and disaggregation. |
| Evidence | Which sources may support the output? | Approved program documents and the governed data dictionary. |
| Guardrails | What must it never infer or decide? | Do not invent a definition; flag conflicts for the data owner. |
| Output | What reusable artifact should it produce? | A dictionary entry with definition, formula, owner, source, and version. |
In short: A template standardizes the final shape; a prompt gives an instruction in one interaction; a skill governs the complete method. A useful skill can use prompts and templates, but also carries evidence rules, quality checks, ownership, and version history.
Illustrative example. Three workforce programs all report “employed,” but one counts any paid work, another requires 20 weekly hours, and a third counts only employment retained for 90 days. The data-dictionary skill should not silently choose one definition. It should surface the conflict, show which reports use each version, and ask the measurement owner to approve a governed definition—or retain clearly named variants when funder contracts require them.
The resulting entry records the construct, precise definition, allowed values, effective date, data owner, collection source, reporting uses, and every exception. The skill can then apply that approved entry elsewhere without reinterpreting it.
Skills support the five courses without becoming a sixth course. Case Intelligence can reuse an evidence-only case-note skill. Grant Intelligence can reuse eligibility and rubric-review skills. Portfolio Intelligence can reuse standards-mapping skills. Connected Data Intelligence can reuse identity, cleaning, and multilingual-analysis skills. Impact Measurement can reuse data-dictionary, evidence-review, and reporting skills.
In short: Test whether it follows the method across ordinary, incomplete, contradictory, multilingual, and edge-case inputs. Check its citations, missing-information behavior, consistency across repeated runs, and treatment of different participant or applicant groups.
Do not approve a skill because one demonstration looks good. Keep a small evaluation set, document failures, require source citations, and record human overrides. For decisions affecting people or funding, people retain final authority.
Choose one definition, analysis rule, or reporting method your team repeatedly reconstructs. Write its five parts—Definition, Instruction, Evidence, Guardrails, and Output—and test it against two real examples that disagree. The disagreement is where the reusable method becomes valuable.
An AI skill is a reusable method for one task. It tells the system what inputs to use, which steps to follow, what evidence to cite, what it must not assume, and what output to produce.
No. A prompt is one instruction. A governed skill also includes evidence rules, decision criteria, quality checks, guardrails, output structure, ownership, testing, and version control.
They are a reference layer used across the courses. A learner follows Case, Grant, Portfolio, Connected Data, or Impact Measurement, then uses the relevant skills inside that workflow.
No. Generative systems may vary. Fixed instructions, approved definitions, citations, evaluation examples, and human review make outputs more consistent and auditable; they do not create a universal guarantee of identical wording or judgment.
The accountable method or data owner should approve it. Technical staff can implement the skill, but program, measurement, compliance, and subject-matter owners must verify definitions, evidence rules, and decision boundaries.
It should not invent missing evidence, conceal contradictions, make unsupported causal claims, or take final authority for high-stakes decisions affecting applicants, grantees, participants, or investments.
ChatGPT, Claude, and Gemini are fine for a quick test — but not for an answer you'll put in front of a funder or board. When it has to hold up, run it in Sopact Sense.
Try it in Sopact Sense →
