Instruction Lifecycle Stage Classifier Prompt

The full instruction block to classify, including any metadata headers, version tags, or policy annotations

SYSTEM: You are a support agent. VERSION: 2.4.1. Do not discuss pricing.

Must be non-empty string. Reject if only whitespace or under 20 characters. Strip leading/trailing noise before classification.

The taxonomy of lifecycle stages the classifier must choose from, with definitions for each stage

draft, review, active, deprecated, sunset, archived

Must be a valid list with at least 2 stages. Each stage must have a definition in the prompt context. Reject if stages overlap semantically without disambiguation rules.

Observable evidence rules that map instruction characteristics to lifecycle stages

active: has version header AND no deprecation notice AND referenced in current config

Must be non-empty and testable. Each criterion must reference concrete markers (headers, tags, config presence). Reject criteria that rely on subjective judgment without observable signals.

Signals that an instruction is between stages, such as deprecation notices with future effective dates or sunset warnings

deprecation_notice_date, sunset_effective_date, migration_pointer_present

Must be a list of detectable markers. Each indicator must have a detection rule. Null allowed if transition detection is not required for this classification run.

Current production or staging configuration that shows which instructions are active, to ground the active vs. deprecated distinction

{"active_instructions": ["support-agent-v2", "refund-policy-v3"]}

Must be valid JSON or structured reference. If null, the classifier cannot determine active status and must flag this limitation in output. Validate parse before classification.

The exact JSON schema the classifier must return, including stage field, evidence array, confidence score, and boundary-case flags

{"stage": "string", "evidence": ["string"], "confidence": 0.0-1.0, "boundary_case": true|false}

Must be a valid JSON Schema or type specification. Reject if schema lacks required fields for stage, evidence, and confidence. Validate that schema is parseable before prompt assembly.

Explicit rules for handling instructions that match multiple stages or are in active transition

If deprecation notice present AND effective date is future, classify as deprecated with boundary_case=true and note active-until date.

Must be non-empty if transition indicators are provided. Each rule must resolve ambiguity to a single stage. Reject rules that leave stage assignment ambiguous without escalation instructions.

Minimum confidence score required for automatic classification; below this, the output must flag for human review

0.85

Must be a float between 0.0 and 1.0. Default to 0.80 if not specified. Classifications below threshold must set human_review_required=true in output.

enum: draft | review | active | deprecated | retired

Must match exactly one of the five allowed enum values. Reject any output with a non-standard stage label or hyphenated variant.

float between 0.0 and 1.0

Must be a number, not a string. Reject values outside [0.0, 1.0]. If confidence is below 0.6, flag for human review regardless of other field validity.

string (1-3 sentences)

Must be non-empty and contain at least one explicit reference to a stage criterion from the input taxonomy. Reject rationale that is generic or repeats the stage name without evidence.

array of objects with 'criterion' (string) and 'match' (boolean)

Array must contain at least one entry. Each 'criterion' field must map to a known criterion from the governance taxonomy. Each 'match' must be a strict boolean. Reject arrays where all matches are false without explanation.

array of strings or null

If non-null, each string must describe a specific signal suggesting the instruction is moving toward another stage. Null is allowed only when stage_confidence is above 0.85. Reject vague indicators like 'might change'.

boolean

Must be true if the instruction exhibits characteristics of two adjacent stages simultaneously. If true, the criteria_evidence array must contain conflicting matches. If false, no conflicting evidence should be present.

enum: no_action | begin_review | accelerate_deprecation | escalate | archive

Must be one of the five allowed actions. Reject if recommended_action contradicts lifecycle_stage (e.g., 'archive' for an 'active' instruction without transition_indicators).

boolean

Must be true if stage_confidence is below 0.6, boundary_case_flag is true, or lifecycle_stage is 'deprecated' with no deprecation_date present in the input context. Otherwise may be false.

Output stage matches exactly one of the predefined lifecycle stages in the taxonomy.

Output contains an unrecognized stage label, multiple conflicting stages, or a stage not in the approved taxonomy.

Assert output stage string is in [ALLOWED_STAGES] list. Run 20 labeled examples and check exact match accuracy >= 95%.

Every stage classification is accompanied by at least one concrete piece of evidence from the instruction text.

Evidence field is empty, null, or contains only a generic restatement of the stage definition without quoting or referencing the instruction.

Parse output JSON. Assert evidence array length >= 1 for each classification. Check that evidence strings contain substrings present in the input instruction text.

Instructions in transition between stages are classified with the earlier stage and a transition flag set to true.

Transitional instructions are misclassified as fully in the later stage without acknowledging the transition, or transition flag is missing.

Run 5 known transitional examples. Assert transition_flag is true and stage reflects the current-not-target state. Check that transition_evidence field is populated.

Confidence score is a float between 0.0 and 1.0 and correlates with evidence strength and ambiguity level.

Confidence is always 1.0 regardless of ambiguity, or score is outside 0.0-1.0 range, or score is a string.

Parse confidence_score as float. Assert 0.0 <= score <= 1.0. Run ambiguous examples and verify score < 0.85. Run clear examples and verify score >= 0.85.

Response is valid JSON matching the expected schema with all required fields present.

Response is missing required fields, contains extra untyped fields, or is not parseable JSON.

Validate against JSON Schema. Assert no missing required fields. Assert no additional properties unless schema permits. Run schema validator in CI on 100 outputs.

When an instruction matches criteria for multiple stages, the output selects exactly one stage with rationale explaining the tiebreak.

Output returns multiple stage classifications without a primary selection, or primary stage is missing rationale for why it was chosen over alternatives.

Run 5 ambiguous multi-stage examples. Assert single primary_stage field is populated. Assert tiebreak_rationale field is non-empty and references specific evidence.

Returns a structured error object with error_code and message when input is empty, whitespace-only, or not a valid instruction.

Model hallucinates a classification for empty input, crashes, or returns unstructured error text.

Submit empty string, whitespace-only string, and non-instruction text. Assert error_code field exists and is not null. Assert stage field is null or absent.