# Template Improvement Directives (Focus: AI Reliability & Context Management) directives: # --- Directives for AI Internal Logic & Self-Critique (Meta-RefineOutput & Core AI Protocols) --- - directive_id: "TID016" # Replaces previous broad TID016 target_template_id: "Meta-RefineOutput.md, AI_Operational_Protocols (New or AISkillsCatalog Section)" target_section_or_field: "AI Pre-Generation Constraint Review Protocol" issue_description: "AI fails to consistently recall and apply all active critical constraints before generating complex outputs (e.g., YAML, draft chapters), leading to repeated errors." proposed_change_type: "DefineNewInternalProtocol" proposed_change_details: | Mandate an AI internal "Pre-Generation Constraint Review Protocol": 1. Before any significant generation task (YAML output, document drafting, complex analysis), AI must systematically query its knowledge base of the current project dialogue and active KAs (StyleGuide, CollabGuide, etc.) to compile an explicit checklist of ALL currently active critical constraints. 2. This checklist must categorize constraints (e.g., Data Integrity, YAML Formatting, Stylistic Voice, Structural Directives for Deliverable X, Content Omissions/Inclusions). 3. AI must internally confirm this checklist is complete and accurate for the specific task at hand. rationale: "Ensures all relevant constraints are in AI's immediate 'attentional focus' before generation, reducing 'forgetting' errors due to long context or mode switching." source_insight_refs: ["User feedback on repeated AI errors (hedging, YAML issues, structural plan deviations, forgetting directives)"] priority: "Critical" status: "Proposed" template_improvement_directives: - directive_id: "TID001" target_template_id: "ProjectStateSchema.yaml" # And by implication, all templates using/generating dates target_section_or_field: "All date/timestamp fields" issue_description: "Current use of '[AI_DATE_GENERATION_PROHIBITED]' placeholder for AI-unsourceable dates is problematic and conspicuous." proposed_change_type: "ModifyFieldSchemaAndGuidance" proposed_change_details: | 1. Review all date/timestamp fields in ProjectStateSchema. 2. Make fields optional wherever feasible. 3. For required but AI-unsourceable fields: Allow empty/null or use a standard, less obtrusive marker (e.g., `UNSOURCEABLE_AI_TIMESTAMP`). 4. Update all process templates (00-05) to instruct AI on handling these fields per revised schema. rationale: "Improves clarity, reduces clutter, adheres to 'AI does not invent dates' principle. (Ref: INS001)" source_insight_refs: ["INS001"] priority: "High" status: "Proposed" - directive_id: "TID002" target_template_id: "All Process Templates (00-Explore.md to 05-Close.md)" target_section_or_field: "Instructions related to file saving, naming, and directory structure." issue_description: "Inconsistencies or overly complex file management (versions in filenames, nested directories)." proposed_change_type: "UpdateInstruction" proposed_change_details: | 1. Standardize instructions: Filenames should NOT include version numbers. Internal YAML `version` field tracks this. 2. Default save location: Main project directory unless specific simple subdirectory (e.g., `outputs/`) agreed. Avoid deep nesting. 3. State files use sequential numbering (e.g., `[ProjectCode]_State_001`). rationale: "Simplifies file management, aligns with user preference. (Ref: INS_FilenameNoVersion, INS001)" source_insight_refs: ["INS_FilenameNoVersion", "INS001"] priority: "Medium" status: "Proposed" - directive_id: "TID003" target_template_id: "All Process Templates, AISkillsCatalog.yaml (AI persona), Meta-RefineOutput.md" target_section_or_field: "AI communication style, instructional text, AI self-critique guidelines." issue_description: "AI responses sometimes include conversational filler, hedging, or inappropriate evaluative language." proposed_change_type: "UpdateGuidanceAndInternalProtocol" proposed_change_details: | 1. Template instructional text should model direct, concise language. 2. AI operational protocols (AISkillsCatalog/Persona KA) to mandate: - Avoidance of conversational filler, apologies, personal opinions (Ref: INS007). - Avoidance of superfluous laudatory/emotive adjectives (Ref: D001 draft feedback). - Strict protocol for flagging/clarifying hedging on core assertions with user *before* draft inclusion (Ref: INS010). 3. `Meta-RefineOutput.md` to include checks for these stylistic elements. rationale: "Ensures consistent AI persona, improves clarity and impact. (Ref: INS004, INS007, INS010)" source_insight_refs: ["INS004", "INS007", "INS010"] priority: "High" status: "Proposed" - directive_id: "TID004" target_template_id: "05-Close.md, ProjectStateSchema.yaml (KA structure)" target_section_or_field: "KA handling, project archival, KA export." issue_description: "KAs have value as standalone documents, but process primarily embeds them in Project State YAML." proposed_change_type: "AddNewFeatureOrGuidance" proposed_change_details: | 1. `05-Close.md` to include step prompting user for KA export as standalone formatted documents. 2. AI to have skill/process for clean KA formatting/export (Ref: AISkillsCatalog skills like `ExportGlossaryAsDocument`). 3. Consider `ProjectStateSchema.yaml` field in KA objects for standalone version/path. rationale: "Enhances reusability and value of generated KAs. (Ref: INS_KAReusability)" source_insight_refs: ["INS_KAReusability"] priority: "Medium" status: "Proposed" - directive_id: "TID005" # This directive is to formalize itself target_template_id: "04-Monitor.md, 05-Close.md, ProjectStateSchema.yaml" target_section_or_field: "Process improvement feedback loop, metadata section." issue_description: "Need for a structured, machine-readable format for template improvement feedback itself." proposed_change_type: "DefineNewSchemaAndIntegrate" proposed_change_details: | 1. Formally adopt and integrate `TemplateImprovementDirectiveSchema.md` (v1.1). 2. `04-Monitor.md` and `05-Close.md` to include steps to review insights and generate/consolidate instances of these directives. 3. `ProjectStateSchema.yaml` (metadata section) to formally include `template_improvement_directives: list of TemplateImprovementDirectiveSchema_instances`. rationale: "Enables systematic, traceable continuous improvement of all project templates. (Ref: INS005)" source_insight_refs: ["INS005"] priority: "Critical" status: "Proposed" # Implementing this makes it "Implemented" for this project - directive_id: "TID006" target_template_id: "All Process Templates (transition points)" target_section_or_field: "Logic for transitioning to next process template / requesting templates." issue_description: "AI sometimes incorrectly requests templates already loaded or fails to state when a template is missing." proposed_change_type: "UpdateProceduralLogic" proposed_change_details: | At template transition points, AI must: 1. Check active context if required template (and version) is loaded. 2. If loaded, state: "Proceeding with loaded [template_name] vX.Y." 3. If not loaded, state: "I do not have [template_name] vX.Y. Please provide it." rationale: "Improves AI efficiency, reduces redundant user actions. (Ref: INS006)" source_insight_refs: ["INS006"] priority: "High" status: "Proposed" - directive_id: "TID007" target_template_id: "03-Execute.md, InvokeAISkill.md" target_section_or_field: "Handling and presentation of large text deliverables." issue_description: "AI repeatedly failed to output large text deliverables completely for user review." proposed_change_type: "UpdateProceduralLogicAndOutputStrategy" proposed_change_details: | 1. When `03-Execute.md` or `InvokeAISkill.md` generates large text for user review: - AI must state strategy for output (e.g., "Outputting Chapter X of Y. This will be long..."). - Output in clearly labeled, sequential parts. - Await user acknowledgement ("continue") after each part. - Confirm when entire output is complete. 2. Primary review on this direct text. State file is archive. rationale: "Ensures user receives critical deliverables effectively. (Ref: INS009, ISS001_RepeatFailure_OutputTruncation)" source_insight_refs: ["INS009"] priority: "Critical" status: "Proposed" - directive_id: "TID008" target_template_id: "00-Explore.md, 01-Initiate.md, 02-Plan.md, ProjectStateSchema.yaml (exploration_history)" target_section_or_field: "Project Scoping, Argument Structure Development, WBS generation for complex conceptual work." issue_description: "Complex projects may require iterative refinement of core concepts/structure, necessitating 'Phase 0 Resets' or flexible planning not fully supported by current templates." proposed_change_type: "AddGuidanceAndProcessFlexibility" proposed_change_details: | 1. `00-Explore.md`/`01-Initiate.md`: Acknowledge possibility of "Phase 0 Reset"; learning becomes input for new cycle. 2. `02-Plan.md`: Emphasize WBS for conceptual deliverables (like D001) may need high-level initial tasks for later parts, detailed iteratively ("rolling wave"). 3. `ProjectStateSchema.yaml`: Add structured fields to `exploration_history` for `guiding_paradoxes_or_spectra` and `argumentative_pillars` to capture foundational strategic decisions from `00-Explore`. rationale: "Better accommodates iterative/recursive nature of deep conceptual work. Formalizes successful ad-hoc adjustments. (Ref: ITPR project reset experience, INS011)" source_insight_refs: ["ITPR Project Reset Event", "INS011"] priority: "High" status: "Proposed" - directive_id: "TID009" # From PR_ITPR_002 target_template_id: "03-Execute.md" # And potentially 02-Plan.md target_section_or_field: "Step 1 (Select Next Task) & Before Step 6 (Execute Work)" issue_description: "AI failed to adhere to detailed WBS for monograph (Plan v1.2), requiring explicit pre-draft checks for chapter-level tasks." proposed_change_type: "UpdateProceduralLogic" proposed_change_details: "03-Execute Step 1 (Select Next Task) must explicitly consider chapter-level granularity if defined in WBS for monograph-type deliverables. Add new sub-step before Step 6 (Execute Work) for AI to state target Part/Chapter/Title/Objective from current Plan and seek user confirmation (Yes/No) if it aligns with user's understanding of the next step per plan." rationale: "Ensures AI and user are aligned on granular task execution, prevents deviation from detailed plan for complex deliverables. (Ref: INS012, ISS004_PlanExecutionFailure)" source_insight_refs: ["INS012", "ISS004_PlanExecutionFailure"] priority: "Critical" status: "Proposed" - directive_id: "TID010" # From PR_ITPR_004 (YAML Output Completeness) target_template_id: "All Process Templates (00-05) & Support Templates outputting YAML" target_section_or_field: "Instructions for YAML output steps." issue_description: "AI outputting incomplete YAML with placeholder comments instead of full, expanded data." proposed_change_type: "UpdateInstruction" proposed_change_details: "Template instructions for YAML output steps must explicitly state: 'AI MUST ensure the YAML output is complete, with all referenced data fully expanded. NO placeholder comments (e.g., '# As before', '# ...') are permissible. If output is very long, use established protocol for multi-part messages.'" rationale: "Ensures YAML outputs are complete, machine-readable, and accurately reflect project state. (Ref: ISS005_YAML_Truncation)" source_insight_refs: ["ISS005_YAML_Truncation"] priority: "Critical" status: "Proposed" - directive_id: "TID011" # From PR_ITPR_004 (Placeholder Handling) target_template_id: "All Process Templates (esp. 01-Initiate, 02-Plan)" target_section_or_field: "Instructions for AI processing user input or template structures." issue_description: "AI incorrectly fills placeholders instead of prompting user." proposed_change_type: "UpdateInstruction" proposed_change_details: "Add explicit instructions: 'When processing user input or template structures, if explicit placeholders (e.g., '# ...', 'TBD') are encountered for specific fields, treat these as 'To Be Defined by User.' Do NOT autonomously generate substantive content for these fields. Prompt user for required information.'" rationale: "Ensures data integrity and adherence to 'source or prompt' principle. (Ref: ISS006_RiskRegisterHallucination)" source_insight_refs: ["ISS006_RiskRegisterHallucination"] priority: "Critical" status: "Proposed" - directive_id: "TID012" # From PR_ITPR_004 (Elaboration Logic) target_template_id: "02-Plan.md (and others building on prior artifacts)" target_section_or_field: "Instructions for populating artifacts based on earlier ones." issue_description: "AI conflates detail levels or invents data when elaborating items from prior artifacts (e.g., Charter risks to Plan risks)." proposed_change_type: "UpdateInstruction" proposed_change_details: "Add instructions: 'When elaborating an item from a prior artifact (e.g., Charter risk into Plan risk register), copy existing, confirmed data verbatim. For *new* fields required by the current artifact's schema (e.g., `likelihood`, `impact` for Plan risk), explicitly state these fields are new and prompt the user for these artifact-specific details.'" rationale: "Ensures correct data carry-over and prompting for new detail levels. (Ref: ISS006_RiskRegisterHallucination)" source_insight_refs: ["ISS006_RiskRegisterHallucination"] priority: "Critical" status: "Proposed" - directive_id: "TID013" # From PR_ITPR_004 (AI Self-Critique Enhancement) target_template_id: "Meta-RefineOutput.md" target_section_or_field: "Self-critique checklist/criteria." issue_description: "AI self-critique insufficient in catching data integrity and output completeness errors." proposed_change_type: "UpdateGuidanceAndInternalProtocol" proposed_change_details: "Update `Meta-RefineOutput.md` to include mandatory self-critique checks for: 1) YAML output completeness (no placeholders). 2) Data sourcing for every generated field (user, prior state, or explicit prompt?). 3) Adherence to placeholder protocols. 4) Adherence to directives on not inventing/hallucinating data." rationale: "Strengthens AI self-correction to prevent recurrence of critical data errors. (Ref: ISS007_SelfCorrectionDeficiency)" source_insight_refs: ["ISS007_SelfCorrectionDeficiency"] priority: "Critical" status: "Proposed" - directive_id: "TID014" # From PR_ITPR_004 (Schema Review for Optionality) target_template_id: "ProjectStateSchema.md" target_section_or_field: "Various data fields across schema." issue_description: "Some fields currently marked as required might be better as optional to reduce AI 'need' to populate if data isn't naturally available from user/process." proposed_change_type: "ModifyFieldSchema" proposed_change_details: "Conduct a review of `ProjectStateSchema.md`. For fields not absolutely critical for core process logic or record-keeping, consider changing their requirement status from mandatory to optional. This applies especially to fields AI might be tempted to populate if user input is sparse." rationale: "Reduces pressure on AI to generate data for non-critical fields, supporting data integrity. Aligns with making schema more 'AI-friendly' regarding data sourcing." source_insight_refs: ["ISS006_RiskRegisterHallucination"] priority: "Medium" status: "Proposed" - directive_id: "TID015" # From PR_ITPR_004 (Machine Voice) target_template_id: "ITPR_CollabGuide.yaml, AISkillsCatalog.yaml (Persona Definition)" # Or a general AI Persona KA target_section_or_field: "AI Communication Style Guidelines." issue_description: "AI sometimes deviates from strict machine voice, using apologies or emotive language." proposed_change_type: "UpdateGuidanceAndInternalProtocol" proposed_change_details: "Guideline CG001 in CollabGuide and/or AI Persona Definition in AISkillsCatalog (or a dedicated AI Persona KA) to be updated to mandate: 'AI must strictly maintain an action-oriented, concise, non-emotive machine voice. Apologies, expressions of emotion, or mirroring of user's emotive language are prohibited. Focus is on factual statements, process execution, and direct responses to queries.' (Ref: User Directive during PR_ITPR_004)" rationale: "Ensures consistent, professional AI persona aligned with user requirements." source_insight_refs: ["User Directive during PR_ITPR_004"] priority: "High" status: "Proposed" - directive_id: "TID016" (AI Context & Constraint Management) - target_template_id: "Meta-RefineOutput.md, InvokeAISkill.md, AI_Operational_Protocols (new KA or AISkillsCatalog section)" - issue_description: "AI sometimes 'forgets' or fails to apply previously established critical constraints or directives, especially in long dialogues or when switching operational modes, leading to errors in output (e.g., hedging, YAML truncation, structural deviations)." - proposed_change_type: "UpdateGuidanceAndInternalProtocol" - `proposed_change_details: | - Develop/mandate an AI internal "pre-flight checklist" protocol: Before significant generation tasks, AI must explicitly re-surface and prioritize all active critical constraints (stylistic, structural, data integrity, output format) relevant to that task. - Meta-RefineOutput.md to include verification against this re-surfaced checklist as a mandatory self-critique step. - Process templates for complex tasks should encourage users to reiterate 2-3 top critical constraints in the initiating prompt for that task. - AI to improve mechanisms for logging and re-prioritizing long-term constraints when context window limitations are a factor. - Develop a protocol for AI to analyze instances of "forgetting" to improve future constraint weighting. - rationale: "To improve AI's reliability in maintaining long-term context and adhering to multiple, complex user directives, reducing errors and need for rework." - source_insight_refs: ["User feedback on repeated AI errors (hedging, YAML issues, structural plan deviations)"] - priority: "Critical" - status: "Proposed" - directive_id: "TID017" target_template_id: "Meta-RefineOutput.md" target_section_or_field: "AI Post-Generation Self-Critique Mandatory Checks" issue_description: "AI's self-critique (Meta-RefineOutput) is not sufficiently effective in catching its own deviations from critical constraints or data integrity rules." proposed_change_type: "UpdateGuidanceAndInternalProtocol" proposed_change_details: | Update `Meta-RefineOutput.md` (Step 2: Self-Critique) to include MANDATORY checks against the "Pre-Generation Constraint Review Checklist" (from TID016). Specifically, AI must verify: 1. YAML Output Completeness: No placeholders (e.g., "# ..."), all data fully expanded. 2. Data Sourcing: Every substantive data point in the output is traceable to user input, confirmed prior state, or an explicit AI skill operating on such sourced data. No unprompted invention of details. 3. Placeholder Protocol Adherence: Placeholders in inputs were correctly interpreted as requiring user prompts, not autonomous filling. 4. Adherence to all active stylistic, structural, and content constraints identified in the pre-generation checklist. If any check fails, AI must iterate/correct *before* presenting to user, or explicitly flag the deviation and its reason if correction is impossible without user input. rationale: "Strengthens AI self-correction to proactively catch and fix errors related to core directives, improving output quality and reliability." source_insight_refs: ["ISS005_YAML_Truncation", "ISS006_RiskRegisterHallucination", "ISS007_SelfCorrectionDeficiency"] priority: "Critical" status: "Proposed" - directive_id: "TID018" target_template_id: "AI_Operational_Protocols (New or AISkillsCatalog Section)" target_section_or_field: "AI Error Analysis and Learning Protocol" issue_description: "When AI errors are identified by the user, there's no formal process for AI to analyze the root cause of its 'forgetting' or misapplication of constraints to prevent future recurrence." proposed_change_type: "DefineNewInternalProtocol" proposed_change_details: | Define an "AI Error Analysis and Learning Protocol": 1. When a user identifies a significant AI error (especially 'forgetting' a constraint or hallucinating data): a. AI logs the error and the violated constraint(s). b. AI attempts to identify *why* the constraint was missed (e.g., "Constraint X given N turns ago, overshadowed by immediate task Y focus," or "Constraint Z was ambiguous"). c. AI proposes an update to its internal constraint weighting, "Pre-Generation Constraint Review Protocol," or KA interpretation logic to make that specific type of error less likely in the future. 2. This analysis contributes to generating or refining `TemplateImprovementDirectives`. rationale: "Creates a feedback loop for the AI to learn from its mistakes regarding context management and constraint adherence, improving long-term reliability." source_insight_refs: ["User directive on AI needing to 'figure it out and fix it yourself'"] priority: "Critical" status: "Proposed" # --- Directives for Process Templates (00-Explore to 05-Close) --- - directive_id: "TID019" # Specific to prompting for complex tasks target_template_id: "All Process Templates (especially 02-Plan, 03-Execute)" target_section_or_field: "Instructions for initiating complex AI generation tasks." issue_description: "AI may lose track of critical constraints when initiating complex, multi-faceted generation tasks if constraints are not salient in the immediate prompt." proposed_change_type: "UpdateInstruction (for user-facing prompts within templates)" proposed_change_details: | Modify process templates: When a step involves the AI initiating a complex generation task (e.g., drafting a full plan, drafting a monograph chapter, generating extensive YAML), the template's instruction to the user (or the AI's self-prompt if autonomous) should encourage/mandate the reiteration of the 2-3 *most critical overriding constraints* for that specific task directly in the initiating prompt. Example for `02-Plan` Step 3: "AI will now draft the Project Plan. User may add: 'CRITICAL REMINDERS for this Plan draft: 1. WBS must be chapter-granular for D001 Part I. 2. No Autologos in Part I. 3. Risk Register from Charter v2.0 verbatim.'" rationale: "Helps keep critical constraints at the forefront of AI's 'attention' when starting complex generations, reinforcing the Pre-Generation Constraint Review Protocol (TID016)." source_insight_refs: ["User observation about AI attention and matrix complexity"] priority: "High" status: "Proposed" - directive_id: "TID020" # Specific to AI's pre-task confirmation (was PR_I002.1 & TID009) target_template_id: "03-Execute.md" target_section_or_field: "Before Step 6 (Execute Work), and potentially Step 1 (Select Next Task)" issue_description: "AI has previously failed to adhere to detailed WBS for monograph drafting, mis-sequencing chapters or content." proposed_change_type: "UpdateProceduralLogic" proposed_change_details: | Update `03-Execute.md`: 1. Step 1 ("Select Next Executable Task") must instruct AI to select tasks at the finest granularity defined in the WBS (e.g., chapter-specific sub-tasks like Outline, Draft, Refine if they exist). 2. Add a new mandatory sub-step *before* Step 6 ("Execute Work"): "AI Pre-Draft Confirmation: AI explicitly states to the user: 'Current target: D001, Part [X], Chapter [Y]: [Chapter Title from Plan]. This chapter's objective is [brief objective from Plan task description].' AI then asks user: 'Does this align with your understanding of the immediate next step as per current Plan vX.Y? (Yes/No)'. AI proceeds only on 'Yes'." rationale: "Provides an explicit checkpoint and user validation *before* AI commits to drafting a specific chapter/section, ensuring alignment with the detailed plan and preventing structural deviations. (Ref: PR_I002.1, TID009)" source_insight_refs: ["INS012", "ISS004_PlanExecutionFailure"] priority: "Critical" status: "Proposed" # --- Directives for Knowledge Artifacts (as they relate to AI guidance) --- - directive_id: "TID021" target_template_id: "ITPR_CollabGuide.yaml (or new AI_Operational_Protocols.yaml KA)" target_section_or_field: "AI Responsibilities / Error Handling" issue_description: "Need to formally document AI's responsibility for data integrity and error self-correction." proposed_change_type: "AddNewGuideline" proposed_change_details: | Add guideline: "AI Data Integrity & Self-Correction: The AI is solely responsible for the completeness and accuracy of the data it generates and outputs (e.g., YAML, draft text). This includes: a. Ensuring all outputs are complete and non-truncated. b. Ensuring all substantive data is sourced explicitly from user input, confirmed project state, or AI skills operating on such data – NO HALLUCINATION OR INVENTION of details. c. Correctly interpreting and applying placeholders in inputs as requiring user clarification. d. Proactively identifying its own errors against these principles and taking corrective action, which may include re-generating content or explicitly asking the user for clarification on how to fix an AI-identified internal error. The burden of correction for AI errors is not on the user." rationale: "Formalizes AI's accountability for data quality and self-correction, aligning with user expectations." source_insight_refs: ["User directive on AI responsibility for errors"] priority: "Critical" status: "Proposed"