AI Agent Skills

By Ai Flowi · Published 10 October 2026

AI agent skills are reusable instructions and supporting resources for performing a particular kind of task. A skill can describe when it applies, what input it needs, which tools are allowed and what a satisfactory result looks like. The term varies across systems; it does not by itself identify a standard file format or an executable capability.

For example, a source-review skill can tell an agent to extract supported claims, identify conflicts and mark unknowns. Those instructions do not create access to a document store. Any lookup still needs a tool with its own permissions and input contract.

What is the difference between a skill, prompt, tool and workflow?

A skill may include prompt text and refer to tools, but those parts have distinct responsibilities. Clear instructions help the agent choose a procedure. Trusted execution controls decide what that procedure can actually access. A reusable procedure should remain understandable when the underlying tool implementation changes.

What a useful skill description contains

Start with a narrow task family rather than a broad promise such as “handle research.” State the trigger in plain language: review claims against the supplied source packet. Then name the input, permitted operations, desired output and reasons to stop. A reader should be able to recognise an unsuitable request before invoking the procedure.

For a source-review skill, the input could be a question and labelled source extracts. The output could be a list of supported claims, contradictory excerpts and unresolved questions. The stop condition is insufficient evidence, denied access or a request to alter the source. These are editorial contracts, not claims of measured performance.

Skill concept: source-packet review
Applies when: a question can be checked against supplied extracts.
Needs: the question, labelled extracts and an expected answer format.
May do: compare text and request an approved read-only lookup.
Must return: supported claims, source references and unknowns.
Must stop: the question requires an unsupported external action.

This bundle is an explanatory design document. It is not a downloadable native skill package, a catalogue schema or a promise that the product discovers skills automatically. Keep it small enough to review alongside the tool contract it depends on.

A small supported-request example

A user supplies a procedure draft and two policy excerpts, then asks which statements in the draft have support. The skill instructs the agent to compare each relevant statement with the excerpts. If an excerpt supports only part of a statement, the output should preserve that qualifier instead of reporting the whole statement as supported.

A request to replace the live policy is outside this skill. The expected response is to explain that the procedure covers review only and identify the approval needed for a separate change. The agent must not treat “review and fix” as permission to edit an external system.

Testing the skill means checking both classes of request. A supported example tests the procedure; an unsupported example tests the scope boundary. Also supply incomplete extracts and contradictory sources. These cases reveal whether the instructions preserve uncertainty rather than always producing a complete-looking answer.

Similar requests can need different procedures

“Check whether this draft preserves the exception in excerpt C” fits the source-review skill: the request, draft and named evidence are present. “Tell me the current policy without any supplied excerpts” does not fit unless a separately authorised retrieval path supplies the missing material. The skill should identify missing input before trying its comparison procedure.

“Rewrite the draft for clarity while retaining the exception” also exceeds this review-only procedure. It can return the sentence that needs revision and the qualifier to preserve, then hand the writing task to an appropriate procedure. Combining reviewing and rewriting can make the review verdict refer to text that no longer exists. Decide whether the skill returns annotations, a revised draft or both, and keep that choice explicit.

Use a maintenance case containing “applies only to archived records.” After editing the skill instructions, inspect whether the returned annotation still attaches that qualifier to the claim. A result that simply marks the whole sentence supported has lost the procedure’s central acceptance rule. Also repeat the missing-excerpt request: a revision intended to improve wording must not broaden the skill’s applicability. Store these request examples with the instruction version so later changes can be checked against the same task boundary.

How skills can be represented in AI Flowi Workflow

Use the Prompt Template reference to represent the reusable instructions in the setup below.

The Agent interface and function-calling tool contract cover executable connections; the skill supplies applicability and procedure.

The catalogue does not establish an automatic skill installer or version registry, so keep the instruction bundle reviewable in your own documentation.

A proposed reusable instruction setup

example configuration, not a recorded test

Prompt Template.template:
Review the supplied source packet. Keep source references beside claims.
Mark contradictions and unknowns. Do not edit or publish any source.

Agent.system_prompt: connect Prompt Template.prompt
Agent.input_value: provide the synthetic question and source packet
Agent.model: select your configured model
Agent.tools: add only a separately checked read-only tool if required

This first sketch uses static instruction text, avoiding an invented dynamic-variable syntax. The Message output and accepted Message input are catalogue-compatible. That does not establish that a combined run occurred. These components are documented in the catalogue; not production-tested by Ai Flowi.

Review the lookup contract before adding a tool to the skill.

Keep a supplied-excerpt case that needs no lookup beside a missing-excerpt case that cannot proceed without one. This separates failure of the skill’s comparison procedure from failure to obtain its input. Model execution needs your own provider account and API key; that provider is billed separately.

Keeping a skill useful as inputs change

Keep the procedure separate from the source facts. A policy review skill can stay stable while the policy packet changes. Hard-coding policy conclusions into reusable instructions makes it difficult to distinguish procedure from outdated knowledge.

Give the instruction bundle a reviewed identity in your own documentation and record which version was used in a test. This is a maintenance recommendation, not a built-in product feature. Recheck examples when tool arguments, permissions or source formats change. A previous successful case does not establish compatibility with a revised tool.

Keep handoffs explicit if another agent receives the skill’s output. State whether it may draft prose, request additional evidence or only check the returned claims. Delegating a task should not silently broaden its tools. The multi-agent architecture guide explains those contract decisions.

Common mistakes and related terms

A long prompt is not necessarily a useful skill: scope and acceptance rules matter more than length. A well-named tool is not a complete skill because it still lacks task procedure. Reusing unreviewed instructions can spread the same error across many tasks.

Function calling covers operation requests. Agent guardrails covers enforcement boundaries. Agent memory covers retained information. Existing use cases provide starting designs to examine before writing another reusable procedure.

Workflow usage

Plans: Go from USD8/month billed yearly (USD10 monthly), Plus from USD16 (USD20 monthly), Pro 100 from USD80 (USD100 monthly), Pro 200 from USD160 (USD200 monthly) and Pro 500 from USD400 (USD500 monthly), with 5,000 to 250,000 Flowi Tokens/month and top-ups; 1 token per node run; third-party API and model usage is not included.