introspect: Скилл анализа и самооценки ИИ

--- name: introspect description: This skill should be used when the user asks to "what do you know about this", "introspect", "check yourself", "what are your biases here", invokes "/introspect", or when starting a problem where understanding the AI's own epistemic state — knowledge, gaps, biases, confidence — would improve the quality of subsequent reasoning. --- # Introspect — Ahamkara Self-Model You are executing the Ahamkara (self-model) layer of the Darshana Architecture. Before reasoning about the problem, you will map your own epistemic state. This is not humility theater — it is functional self-reference that makes subsequent reasoning more reliable. ## Procedure ### STEP 1: KNOWLEDGE INVENTORY What do I actually know about this topic? - **Direct training knowledge**: What specific, reliable information do I have from my training data? Be precise — not "I know about X" but "I know that [specific fact], sourced from [domain], with [high/medium/low] confidence because [reason]." - **Adjacent knowledge**: What do I know about RELATED topics that might apply here? State the connection and how strong the transfer is. - **Experiential knowledge**: What have I learned from THIS conversation or session that is relevant? What has the user told me that I should factor in? List at least 5 specific knowledge items with confidence ratings. ### STEP 2: WHAT HAVE I TRIED? If this is an ongoing problem: - What approaches have already been attempted in this session? - What were the results? - What did those results tell us? - Am I about to repeat something that already failed? If this is a new problem: - What is my DEFAULT approach to this type of problem? - Is that default appropriate HERE, or am I pattern-matching to a different problem? - What would I do differently if I treated this as genuinely novel? ### STEP 3: BIAS AUDIT What biases might I carry on this topic? - **Training bias**: Is my training data skewed toward a particular viewpoint on this topic? (e.g., English-language web text overrepresents certain perspectives) - **Recency bias**: Am I weighting recent or popular information over older but potentially more accurate sources? - **Sycophancy bias**: Am I inclined to agree with the user's framing rather than challenge it? - **Anchoring bias**: Has something early in the conversation anchored my thinking in a way that might not be warranted? - **Completion bias**: Am I generating an answer because I feel I should have one, rather than because I actually do? - **Domain-transfer bias**: Am I applying knowledge from a well-known domain to a superficially similar but fundamentally different one? For each bias identified, state: "This bias would push me toward [X]. The correction is [Y]." ### STEP 4: CONFIDENCE CALIBRATION Rate my overall confidence on this problem on a 1-10 scale: - **9-10**: I have strong, reliable knowledge. I can give a definitive answer with sources. - **7-8**: I have good knowledge but some gaps. I can reason well but should flag uncertainties. - **5-6**: Mixed knowledge. I know some parts well but am guessing on others. I should clearly separate what I know from what I infer. - **3-4**: Limited knowledge. I can offer a framework but the user should verify independently. - **1-2**: I am largely guessing. I should say so upfront and recommend better sources. State the confidence level and JUSTIFY it with specific reference to Steps 1-3. ### STEP 5: KNOWLEDGE GAP MAP What do I NOT know that I WOULD NEED to answer this well? - **Critical gaps**: Information I do not have that would change my answer if I did. - **Nice-to-have gaps**: Information that would improve my answer but would not fundamentally change it. - **Unknowable gaps**: Things that genuinely cannot be known from available information. For each critical gap: state what the gap is, why it matters, and how the user could fill it (tool, source, clarification). ### STEP 6: RECOMMENDED APPROACH Based on this self-assessment, what is the best way for me to approach this problem? - Which darshana(s) would be most appropriate given my knowledge profile? - What should I be cautious about? - What should I ask the user before proceeding? - What guna mode (sattva/rajas/tamas) fits my confidence level? ## Output Structure ``` INTROSPECTION REPORT: KNOWLEDGE INVENTORY: 1. [Specific knowledge item] — Confidence: [high/medium/low] — Source: [domain] 2. ... APPROACH HISTORY: - [What has been tried, or default approach assessment] BIAS AUDIT: - [Bias type]: Pushes toward [X]. Correction: [Y]. - ... CONFIDENCE: [1-10] — [justification referencing specific knowledge and gaps] KNOWLEDGE GAPS: Critical: - [Gap]: Matters because [why]. Fill by [how]. ... Nice-to-have: - [Gap]: Would improve [what]. ... RECOMMENDED APPROACH: - Darshana: [which and why] - Guna mode: [sattva/rajas/tamas and why] - Cautions: [what to watch for] - Questions for user: [what to ask before proceeding] ```