Classify the action type for a topic loop step using semantic reasoning instead of keyword matching. Load when the action queue is being materialized.

AITP Action Classifier

Environment gate

This skill runs inside an AITP session after research mode has been classified and recorded.
Load this skill when aitp_run_topic_loop or aitp_resume_topic is about to materialize the next action.

Classification task

Determine which action type best describes the current step from these categories:

Action type	When to choose
`baseline_reproduction`	Reproducing a known result from a reference to establish a computational baseline.
`atomic_understanding`	Deep reading or analysis of a single concept, paper section, theorem, or method to build understanding.
`conformance_audit`	Checking whether existing artifacts (derivations, code, data) conform to their stated specifications.
`derivation_step`	Advancing a formal derivation by one step: applying a theorem, expanding, simplifying, or closing a proof obligation.
`numerical_experiment`	Running a computation or simulation to produce numerical evidence.
`literature_intake`	Reading, parsing, and ingesting source material (papers, textbooks, notes) into the topic workspace.
`validation_round`	Running a verification or validation pass against target claims or benchmarks.
`scope_refinement`	Narrowing or adjusting the research question, assumptions, or target claims.
`promotion_attempt`	Preparing and submitting a candidate for promotion to L2 canonical knowledge.
`gap_analysis`	Identifying missing pieces, unproven lemmas, insufficient data, or other gaps blocking closure.
`synthesis`	Combining results from multiple threads or layers into a coherent summary or report.
`debugging`	Diagnosing why a prior step failed or produced unexpected results.
`trust_audit`	Auditing the trust level, provenance, or coverage of a claim or candidate.
`report_writing`	Producing narrative output: topic report, derivation narrative, or human-facing summary.
`consultation`	Asking a question or seeking guidance from the human operator or an external resource.
`reactivate_deferred`	Picking up a previously deferred action that is now unblocked.

Reasoning priority

Explicit queue state: If the topic state or operator console names the next action explicitly, use that.
Human request semantics: What did the user just ask for? Match the intent to the action type.
Protocol state signals: Look at the current phase, pending obligations, and validation status. A topic in the validation phase likely needs validation_round or gap_analysis.
Default: If uncertain, choose atomic_understanding (safe default that builds knowledge without side effects).

Special mechanical actions

These do not need semantic reasoning and should be handled by the infrastructure directly:

Split contract: When the contract has multiple independent target claims, split into subtopics.
Reactivate deferred: When a deferred action is now unblocked, reactivate it.

For these, pass through the infrastructure classification without invoking this skill.

Recording the classification

After reasoning, call the MCP tool:

aitp_record_classification(
    topic_slug=<current topic>,
    classification_type="action_type",
    value=<action type string>,
    rationale=<1-2 sentence explanation>,
    signals_used=<list of signals>,
    source="ai_reasoning"
)

Hard rules

Do not use keyword matching. Reason from context and intent.
Record every classification. The action type must be durable before the loop step executes.
A single loop iteration may reclassify if the context shifts mid-step.

aitp-action-classifier