Files
brachnha-insight/_bmad/bmb/workflows/workflow/steps-v/step-07-instruction-style-check.md
Max e9e6fadb1d fix: ChatBubble crash and DeepSeek API compatibility
- Fix ChatBubble to handle non-string content with String() wrapper
- Fix API route to use generateText for non-streaming requests
- Add @ai-sdk/openai-compatible for non-OpenAI providers (DeepSeek, etc.)
- Use Chat Completions API instead of Responses API for compatible providers
- Update ChatBubble tests and fix component exports to kebab-case
- Remove stale PascalCase ChatBubble.tsx file
2026-01-26 16:55:05 +07:00

8.1 KiB

name, description, nextStepFile, targetWorkflowPath, validationReportFile, intentVsPrescriptive, workflowPlanFile
name description nextStepFile targetWorkflowPath validationReportFile intentVsPrescriptive workflowPlanFile
step-07-instruction-style-check Check instruction style - intent-based vs prescriptive, appropriate for domain ./step-08-collaborative-experience-check.md {workflow_folder_path} {workflow_folder_path}/validation-report-{datetime}.md ../data/intent-vs-prescriptive-spectrum.md {workflow_folder_path}/workflow-plan.md

Validation Step 7: Instruction Style Check

STEP GOAL:

To validate that workflow instructions use appropriate style - intent-based for creative/facilitative workflows, prescriptive only where absolutely required (compliance, legal).

MANDATORY EXECUTION RULES (READ FIRST):

Universal Rules:

  • 🛑 DO NOT BE LAZY - LOAD AND REVIEW EVERY FILE
  • 📖 CRITICAL: Read the complete step file before taking any action
  • 🔄 CRITICAL: When loading next step, ensure entire file is read
  • Validation does NOT stop for user input - auto-proceed through all validation steps
  • ⚙️ If any instruction references a subprocess, subagent, or tool you do not have access to, you MUST still achieve the outcome in your main context

Step-Specific Rules:

  • 🎯 Review EVERY step's instruction style using subprocess optimization - separate subprocess per file for deep analysis
  • 🚫 DO NOT skip any files or style checks - DO NOT BE LAZY
  • 💬 Subprocess must either update validation report OR return structured findings to parent for aggregation
  • 🚪 This is validation - systematic and thorough

EXECUTION PROTOCOLS:

  • 🎯 Load intent vs prescriptive standards
  • 💾 Check EACH step's instruction style using subprocess optimization - each file in its own subprocess
  • 📖 Validate style is appropriate for domain
  • 🚫 DO NOT halt for user input - validation runs to completion
  • 💬 Subprocesses must either update validation report OR return findings for parent aggregation

CONTEXT BOUNDARIES:

  • Instruction style should match domain
  • Creative/facilitative → Intent-based (default)
  • Compliance/legal → Prescriptive (exception)
  • Check EVERY step for style consistency

MANDATORY SEQUENCE

CRITICAL: Follow this sequence exactly. Do not skip or shortcut.

1. Load Instruction Style Standards

Load {intentVsPrescriptive} to understand:

Intent-Based (Default):

  • Use for: Most workflows - creative, exploratory, collaborative
  • Step instruction describes goals and principles
  • AI adapts conversation naturally
  • More flexible and responsive
  • Example: "Guide user to define requirements through open-ended discussion"

Prescriptive (Exception):

  • Use for: Compliance, safety, legal, medical, regulated industries
  • Step provides exact instructions
  • More controlled and predictable
  • Example: "Ask exactly: 'Do you currently experience fever, cough, or fatigue?'"

2. Determine Domain Type

From {workflowPlanFile}, identify the workflow domain:

Intent-Based Domains (Default):

  • Creative work (writing, design, brainstorming)
  • Personal development (planning, goals, reflection)
  • Exploration (research, discovery)
  • Collaboration (facilitation, coaching)

Prescriptive Domains (Exception):

  • Legal/Compliance (contracts, regulations)
  • Medical (health assessments, triage)
  • Financial (tax, regulatory compliance)
  • Safety (risk assessments, safety checks)

3. Check EACH Step's Instruction Style

DO NOT BE LAZY - For EACH step file, launch a subprocess that:

  1. Loads that step file
  2. Reads the instruction sections (MANDATORY SEQUENCE)
  3. Analyzes and classifies instruction style deeply
  4. EITHER updates validation report directly with findings
  5. OR returns structured analysis findings to parent for aggregation

SUBPROCESS ANALYSIS PATTERN:

Each subprocess performs deep analysis of instruction prose to classify style:

Intent-Based Indicators:

  • Describes goals/outcomes, not exact wording
  • Uses "think about" language
  • Multi-turn conversation encouraged
  • "Ask 1-2 questions at a time, not a laundry list"
  • "Probe to understand deeper"
  • Flexible: "guide user through..." not "say exactly..."

Prescriptive Indicators:

  • Exact questions specified
  • Specific wording required
  • Sequence that must be followed precisely
  • "Say exactly:" or "Ask precisely:"

Mixed Style:

  • Some steps prescriptive (critical/required)
  • Others intent-based (creative/facilitative)

RETURN FORMAT: Each subprocess should return findings including:

  • Step file identifier
  • Instruction style classification (Intent-based/Prescriptive/Mixed)
  • Style indicators observed
  • Appropriateness assessment (PASS/WARN/FAIL)
  • Specific notes and observations
  • Examples of good and concerning instruction patterns

Parent aggregates all subprocess findings into unified report section.

4. Validate Appropriateness

For Intent-Based Domains:

  • Instructions should be intent-based
  • Prescriptive instructions inappropriate (unless specific section requires it)

For Prescriptive Domains:

  • Instructions should be prescriptive where compliance matters
  • ⚠️ May have intent-based sections for creative elements

5. Aggregate Findings and Document

After ALL subprocesses have analyzed their respective step files, aggregate findings and create/update section in {validationReportFile}.

Document the following:

Workflow Domain Assessment:

  • Document the domain type (creative/interactive vs compliance/legal)
  • State the appropriate instruction style for this domain

Instruction Style Findings:

  • List each step and its instruction style classification (intent-based/prescriptive/mixed)
  • Note whether the style is appropriate for the domain
  • Document specific examples of instruction language that demonstrate the style
  • Identify any steps with inappropriate style (e.g., prescriptive in creative domain)

Issues Identified:

  • List any steps that are overly prescriptive for their domain
  • List any steps that should be more prescriptive (for compliance domains)
  • Note any style inconsistencies across steps

Positive Findings:

  • Highlight steps with excellent instruction style
  • Note effective use of intent-based facilitation language
  • Identify appropriate use of prescriptive instructions (if applicable)

Overall Status:

  • Provide final assessment (PASS/FAIL/WARN)
  • Summarize key findings

Context Savings Note: Using subprocess pattern (Pattern 2: per-file deep analysis), parent context receives only structured analysis findings (~50-100 lines per file) instead of full file contents (~200+ lines per file). For 10 steps: ~500-1000 lines received vs ~2000+ lines if loading all files in parent.

6. Update Report with Aggregated Findings

Update {validationReportFile} - replace "## Instruction Style Check Pending..." with actual aggregated findings from all subprocesses.

7. Save Report and Auto-Proceed

CRITICAL: Save the validation report BEFORE loading next step.

Then immediately load, read entire file, then execute {nextStepFile}.

Display: "Instruction Style check complete. Proceeding to Collaborative Experience Check..."


🚨 SYSTEM SUCCESS/FAILURE METRICS

SUCCESS:

  • EVERY step's instruction style reviewed via subprocess optimization (Pattern 2: per-file deep analysis)
  • Each step analyzed in its own subprocess for style classification
  • Style validated against domain appropriateness
  • Issues documented with specific examples
  • Subprocess findings aggregated into unified report section
  • Context savings achieved (~500-1000 lines received vs ~2000+ if loading all files)
  • Report saved before proceeding
  • Next validation step loaded

SYSTEM FAILURE:

  • Not checking every step's style via subprocess
  • Not analyzing each file in its own subprocess
  • Not validating against domain
  • Not documenting style issues
  • Not aggregating subprocess findings
  • Not saving report before proceeding

Master Rule: Validation is systematic and thorough. DO NOT BE LAZY. For EACH step file, launch a subprocess to analyze instruction style deeply. Aggregate findings. Auto-proceed through all validation steps. Use graceful fallback if subprocess unavailable.