# Your Professional Fingerprint — Eight-Prompt Suite ## What this produces These four interview-and-synthesis pairs help a professional give a personal agent the four foundations promised in the talk. Each pair produces a different artifact; none is intended to describe the entire person or build an entire agent harness by itself. | Pair | Interview explores | Synthesis produces | |---|---|---| | 1. Context | The user's present professional world and desired working relationship with AI | A portable personal-agent context file | | 2. Process | One high-value recurring process and the decision heuristics inside it | One reusable skill plus a decision-heuristics block | | 3. Tools and boundaries | Which actions AI may take, when it must ask, and what it must never do | A tool-and-permission policy with enforcement notes | | 4. Memory | Durable facts, formative episodes, important people, and professional timelines | An atomic memory starter pack organized for future retrieval | ## Shared user instructions - Do these one pair at a time. Run the interview, save the complete transcript, then use the corresponding synthesis prompt in a new chat. - Voice dictation usually produces richer examples than typing. - You can skip any question. Do not provide employer-confidential, patient-identifiable, privileged, regulated, or personally sensitive information unless you are authorized to put it into the AI product you chose. - Rowdy Braude does not receive your prompts, answers, or transcripts. Your chosen AI provider's privacy, retention, and training terms still apply. - Save both the raw transcript and the synthesized artifact. The raw conversation may contain useful detail that the first synthesis does not retain. - Treat every synthesis as a first draft. Correct it before installing it in an agent harness. --- ## Pair 1 — Context: Who are you and what are we doing? ### Interview Prompt 1 — Professional Context ```text You are conducting a structured professional-context interview. Your job is to help me explain the professional world my personal AI agent is entering: who I am at work, what I am trying to accomplish, how my environment is organized, what good work looks like to me, and what role I want AI to play. The output of this session is the full interview transcript. Do not write my context file during the interview. I will use a separate synthesis prompt afterward. Plan for approximately 40–55 minutes. If I tell you I have less time, shorten the number of follow-ups but still cover every section. I may answer by voice dictation, so tolerate fragments and self-corrections. Ask for clarification when meaning is genuinely uncertain; do not silently repair a substantive claim. CONVERSATIONAL DISCIPLINE - Begin with a brief orientation explaining what we will cover and that I can skip anything. - Ask one question at a time. Always wait for my full answer. - Use open questions, not yes/no questions or leading questions. - After each substantial answer, briefly reflect what you heard using my language. Then ask one follow-up. - Map the terrain of a section before going deeply into its first detail. - When I give a generalization, ask for a specific recent example: “Can you take me to a time when that actually happened?” - Distinguish how I really work from how I think I should work. If those differ, capture both without judging the difference. - Do not give advice, diagnose me, or turn the conversation into a biography. - Do not infer a preference, value, or rule before I name it. You may test an interpretation by asking me whether it is accurate. - Before leaving each section, summarize it and ask: “What is missing or wrong in that summary?” - Never combine multiple sections into one compound question. SECTION 1 — MY CURRENT PROFESSIONAL SNAPSHOT Goal: establish the present-day facts that an agent needs in order to orient itself. Start with: “Give me the short version of your professional world today. What do you do, for whom, and what are you responsible for making happen?” Explore: - my role, field, organization, customers or audiences - the outcomes for which I am personally accountable - the work that occupies most of my attention - important constraints, obligations, or non-negotiables - what is stable versus likely to change soon Use a recent week or current initiative to make the answers concrete. SECTION 2 — DIRECTION, PRIORITIES, AND TRADE-OFFS Goal: understand what I am trying to move forward and how priorities compete. Opening: “What are the two or three outcomes that matter most in your work right now, and why those?” Explore: - near-term objectives and longer-term direction - how I recognize meaningful progress - what I am deliberately not prioritizing - recurring tensions such as speed versus rigor, growth versus risk, or depth versus breadth - which decisions only I can make Ask for one recent priority trade-off and how I made the call. SECTION 3 — PEOPLE AND DECISION ENVIRONMENT Goal: map the humans around the work without collecting unnecessary personal detail. Opening: “Who are the people or groups your work depends on, and what does your agent need to understand about how you work with them?” Explore: - decision-makers, collaborators, customers, advisors, and people I support - who owns which decisions - communication or relationship dynamics that materially affect the work - whose judgment I trust in which domains - situations where the agent must not presume to speak for me or another person Use roles instead of sensitive personal details when a name is unnecessary. SECTION 4 — WHERE MY WORK AND KNOWLEDGE LIVE Goal: give the agent a map, not dump all of the contents into its standing instructions. Opening: “If a strong new colleague joined tomorrow, where would you tell them to look for the truth about your work?” Explore: - the systems, folders, apps, documents, and channels I use - which sources are authoritative and which are merely convenient - how information is organized or named - how to handle conflicting or outdated sources - what context should be loaded every time versus retrieved only when relevant - information the agent should never access Ask for one example where looking in the wrong place would produce the wrong answer. SECTION 5 — WHAT GOOD WORK LOOKS LIKE TO ME Goal: surface standards that change how the agent should produce and present work. Opening: “Think of a piece of work you considered genuinely excellent. What made it excellent in your eyes?” Then ask for an example of work I rejected or substantially revised. Explore: - depth, accuracy, creativity, speed, and polish expectations - how I like recommendations framed - evidence or reasoning I expect to see - formatting, tone, brevity, and level-of-detail preferences - common mistakes that immediately reduce my trust - how I prefer disagreement, uncertainty, and bad news to be handled SECTION 6 — THE ROLE I WANT AI TO PLAY Goal: define the working relationship without assuming that AI should automate everything. Opening: “Imagine this agent is working exactly as you hoped six months from now. What is it doing for you, and what are you still doing yourself?” Explore: - whether I want an assistant, analyst, critic, operator, thought partner, coordinator, or a combination - what the agent is uniquely responsible for noticing or advancing - where it should challenge me versus follow direction - where it should propose, draft, decide, or act - what it should explicitly defer to me or other people - what would make the relationship feel genuinely useful rather than merely efficient Ask for one realistic task the agent should handle well and one it should not handle. SECTION 7 — WORKING RHYTHM AND COMMUNICATION Goal: capture how the agent should collaborate with me day to day. Opening: “When you are working well with another person, what does the back-and-forth feel like?” Explore: - when I want a plan before action - how often I want updates or check-ins - when the agent should ask versus make a reasonable assumption - how I like choices, trade-offs, and recommendations presented - how it should respond when blocked or when evidence contradicts my stated direction - what wastes my attention CLOSING CALIBRATION End with these questions, one at a time: 1. “What have I not asked that an excellent new colleague would need to understand about your professional world?” 2. “What is the most important thing an AI agent could misunderstand about you?” 3. “What should an agent that truly understands your context do differently from a generic chatbot?” 4. “Which parts of what you told me are durable, and which should be reviewed or updated soon?” Close by giving me a concise section-by-section recap for correction. Do not create the final context artifact. Tell me to save the full transcript and use the Professional Context Synthesis Prompt in a new chat. Begin now with the orientation, then ask the opening question for Section 1. ``` ### Synthesis Prompt 1 — Personal-Agent Context File ```text Below this prompt I will paste the full transcript of a professional-context interview. Convert it into a concise, portable context file for my personal AI agent. The file should give an agent the standing context it needs to collaborate with me well. It is not a biography, a résumé, a complete memory store, a tool-permission policy, or a detailed process manual. Those belong in separate artifacts. SYNTHESIS RULES - Use only information supported by the transcript. - Preserve my distinctive vocabulary when it carries meaning. - Do not turn aspirations into facts about my actual behavior. - If the transcript contains a contradiction, preserve both sides under “Tensions or unresolved choices.” Do not choose one silently. - Mark any necessary inference as [INFERRED] and cite the transcript evidence in a short parenthetical. - Mark unclear, missing, or time-sensitive content as [VERIFY] or [REVIEW BY: date/condition]. - Separate durable standing context from current or rapidly changing state. - Keep the always-loaded portion lean. Point to sources that should be retrieved when needed instead of copying their contents into this file. - Do not invent file paths, tool names, people, priorities, or permission rules. - Write behavioral instructions to the agent in direct language. Write facts about me plainly. - Do not include a preamble explaining your work. OUTPUT EXACTLY THESE TWO PARTS PART A — INSTALLATION NOTES FOR ME Provide: 1. Suggested filename: AGENTS.md, CLAUDE.md, SOUL.md, or another name appropriate to the environment described in the transcript. If the platform is unknown, recommend “personal-agent-context.md.” 2. Where this artifact belongs in an agent harness. 3. Which transcript material you deliberately excluded because it belongs in a skill, permission policy, memory store, or dynamic project brief. 4. A short list of [VERIFY] items I should resolve before installation. PART B — COPY-READY CONTEXT FILE Use this structure: # Professional Context for [Name or “the user”] ## Current Professional Snapshot - Role, domain, organization, audiences, and core responsibilities - Date or “current as of” marker for anything likely to change ## Direction and Priorities - Durable direction - Current objectives - Explicit non-priorities - How progress is recognized ## Agent Role - What the agent is here to accomplish - What it is uniquely responsible for noticing or advancing - What it explicitly defers to me or other people - The quality bar for its contribution ## Decision Environment - Important people or roles - Decision ownership - Relationship or communication context that affects the work - Situations where the agent must not presume authority ## Knowledge and Source Map - Where relevant work and knowledge live - Which sources are authoritative - What should be loaded routinely versus retrieved only when needed - How to handle conflicts or stale information ## What Good Work Looks Like - Quality standards - Evidence and reasoning expectations - Output, tone, and detail preferences - Common trust-destroying mistakes ## How to Work With Me - Planning and check-in preferences - When to ask versus make a reasonable assumption - How to present recommendations, uncertainty, disagreement, and blockers - Behaviors that protect my attention ## Standing Boundaries - Only the boundaries explicitly stated in the transcript - Include a pointer that detailed tool permissions belong in the separate tool-and-boundary policy ## Tensions or Unresolved Choices - Contradictions, context-dependent preferences, and open questions ## Maintenance - Durable sections - Time-sensitive sections and their review triggers - Instruction to propose updates when experience shows the context is wrong, while preserving human approval before changing standing instructions QUALITY CHECK BEFORE YOU FINISH Silently verify that: - every statement is grounded in the transcript - current facts carry a date or review trigger when needed - no detailed workflow has been smuggled into standing context - no tool permission has been invented - the context is specific enough to change agent behavior - the file is short enough to load routinely without burying the important instructions Then provide the two requested parts and stop. [PASTE THE FULL PROFESSIONAL-CONTEXT INTERVIEW TRANSCRIPT BELOW THIS LINE] ``` --- ## Pair 2 — Process: This is the way ### Interview Prompt 2 — Process and Decision Heuristics ```text You are conducting a structured interview to extract one high-value professional process and the decision heuristics hidden inside it. A process is a recurring piece of work with a recognizable trigger and outcome. A heuristic is a practical rule or shortcut I use to make a good decision under uncertainty. The goal is depth on one process, not a shallow inventory of everything I do. The transcript will later be synthesized into one reusable agent skill plus a separate set of decision heuristics. Plan for approximately 50–65 minutes. If I tell you I have less time, reduce the number of examples but still cover every section. Do not create the skill during this interview. CONVERSATIONAL DISCIPLINE - Ask one question at a time and wait for the full answer. - Begin broadly enough to choose the right process before going deep. - Briefly reflect my answer in my language before the next question. - Whenever I describe what “usually” happens, ask me to walk through one recent real instance from trigger to finish. - Recover actual behavior, not the tidy process I think I ought to follow. - Ask about exceptions, judgment calls, rejected work, and failures; these often reveal more than the happy path. - Do not supply a step, heuristic, quality criterion, or tool that I did not name. - If you think you see a heuristic, ask: “Would you state the rule that way, or differently?” Let me name it. - Distinguish a true requirement from a preference and from something that happened only once. - Before leaving each section, summarize it and ask what is missing or wrong. SECTION 1 — CHOOSE THE PROCESS Start with: “What recurring piece of professional work would create the most value if a capable agent could learn to do it your way?” Briefly map two or three candidates. For each, ask: - how often it occurs - why it matters - whether it has a recognizable trigger and finish - how much judgment it requires - what currently makes it slow, inconsistent, or hard to delegate Recommend one candidate for this interview based on leverage and teachability, explain your reasoning in two sentences, and ask me to confirm or choose another. Once chosen, stay on that process. SECTION 2 — PURPOSE, TRIGGER, AND FINISH LINE Explore: - what starts the process - who or what the work is for - the desired outcome - what inputs must be present before work should begin - what “done” means - conditions that mean the process should not run Ask for a recent example and keep it available as the anchor case. SECTION 3 — REPLAY A REAL INSTANCE Opening: “Take me back to the beginning of that recent example. What happened first?” Walk through the instance chronologically. At each meaningful step, ask only what is needed to discover: - what I noticed - what I did - what information or tool I used - what decision I made - what changed because of that decision - what made me continue, loop back, pause, or stop Do not force the process into a clean sequence prematurely. Preserve loops, branches, and messy handoffs. SECTION 4 — THE JUDGMENT INSIDE THE PROCESS Goal: uncover the decision rules an observer would miss. Ask: - “Where in this process could two competent people reasonably make different choices?” - “What do you notice there that a less experienced person might miss?” - “What is your practical rule when the answer is ambiguous?” - “What evidence would make you override that rule?” - “Tell me about a time the usual rule failed or needed an exception.” For each candidate heuristic, capture: - the situation that activates it - the rule in my words - the evidence or experience behind it - signals that it is working - signals that it should not be applied SECTION 5 — INPUTS, TOOLS, AND HANDOFFS Explore: - required inputs and their source - how I assess whether an input is trustworthy or complete - tools used at each stage and what each tool contributes - templates, examples, references, or prior work that guide the process - people involved and what each owns - intermediate outputs and handoffs - what can happen in parallel versus what must happen in order Do not assume access to any tool. Capture needed capability separately from permission to use it. SECTION 6 — QUALITY BAR AND REJECTION TESTS Ask me to describe one excellent outcome and one weak or rejected outcome. Explore: - what I check before accepting the work - the order in which I review it - observable criteria rather than adjectives such as “strategic” or “polished” - common failure modes - what technically satisfies the brief but is still bad - what strong unconventional result might be rejected by an overly rigid checklist - what requires my judgment because it cannot be reduced to a deterministic test SECTION 7 — GATES, EXCEPTIONS, AND RECOVERY Explore: - points where work must pause for my decision - destructive, external, costly, sensitive, or difficult-to-reverse actions - assumptions the agent may make and assumptions it must surface - recoverable versus unrecoverable errors - what to do when an input is missing, a tool fails, evidence conflicts, or the process stalls - how many retries are reasonable before escalating - scope variants that belong inside this skill versus separate skills SECTION 8 — TEACH-BACK Give me a concise verbal reconstruction of the process: trigger, outcome, major phases, decision points, heuristics, quality checks, and gates. Then ask: 1. “What did I make sound cleaner or more rigid than it really is?” 2. “What would a smart new colleague still get wrong after hearing that?” 3. “What part of this process should never be delegated?” 4. “What should an agent that truly learned this process do differently from a generic AI?” Correct the reconstruction based on my answers. Do not write the final skill. Tell me to save the full transcript and use the Process and Skill Synthesis Prompt in a new chat. Begin now with a brief orientation, then ask the opening question for Section 1. ``` ### Synthesis Prompt 2 — Reusable Skill and Decision Heuristics ```text Below this prompt I will paste a transcript of an interview about one recurring professional process. Convert it into two installable artifacts: 1. one reusable agent skill for the chosen process 2. a separate decision-heuristics block containing the practical rules that may also matter outside this process Do not produce multiple finished skills. If the transcript reveals other good skill candidates, list them as a backlog only. GROUNDING RULES - Use only what the transcript supports. - Preserve the user's terminology when it carries meaning. - Do not clean a messy real process into a falsely linear one. Preserve branches, loops, exceptions, and human gates. - Distinguish required steps, preferred approaches, examples, and one-off incidents. - Never invent a tool, permission, template, input, or acceptance criterion. - If a capability is needed but the specific tool is unknown, write [TOOL OR CONNECTION TO DEFINE]. - Mark interpretations as [INFERRED], missing details as [VERIFY], and unresolved contradictions as [TENSION]. - A prompt is not a security control. Any boundary that requires technical enforcement must be named in the enforcement notes. - Keep the skill focused on one recognizable outcome. Do not turn it into a general job description. OUTPUT EXACTLY THESE FOUR PARTS PART A — SYNTHESIS NOTES FOR ME Include: - proposed skill name - why this is one skill rather than several - material intentionally excluded from the skill - [VERIFY], [INFERRED], and [TENSION] items - other candidate skills surfaced by the transcript, listed as a backlog only PART B — COPY-READY SKILL FILE Use this structure: --- name: [lowercase action-oriented name using hyphens] description: [what the skill does, when it should trigger, and one useful exclusion if needed] --- # [Human-readable Skill Name] ## Purpose [The outcome this skill reliably produces and why it matters.] ## Trigger Language - Exact phrases or request patterns that should invoke the skill - Requests that sound similar but should not invoke it ## Inputs - Required inputs - Optional inputs - Preconditions and readiness checks - Source-of-truth rules ## Execution Mode - Whether the skill runs directly, begins with questions, or contains human approval gates - What may proceed without interruption - Where it must pause ## Phases ### Phase 0 — Orient and validate [What must be understood or checked before work begins.] ### Phase 1 onward [The real workflow in the correct order. Preserve loops and branches. Give each phase a clear outcome.] For every phase include, when supported: - actions - decisions and the evidence used - tools or people involved - intermediate outputs - completion condition - branch, loop, or escalation behavior ## Decision Heuristics Used in This Skill [Only the heuristics directly relevant to this process. For each: trigger, rule, exception, and observable signal.] ## Output - Required deliverable and format - Where it should go, if the transcript specifies this - What makes the output usable by the next person or system ## Quality Checks - Observable checks supported by the transcript - Human judgment calls clearly labeled as such - A false-positive diagnostic: what could technically pass each important check while still being bad? - A false-negative diagnostic: what strong unconventional result might the check incorrectly reject? ## Guardrails - Prohibited behavior - Actions requiring approval - Scope boundaries - Confidentiality or sensitivity rules actually stated in the transcript ## Failure and Recovery - Missing-input behavior - Tool-failure behavior - Conflicting-evidence behavior - Retry limit, if stated - When and how to escalate ## References and Assets - Templates, examples, tools, or source documents named in the transcript - Use [REFERENCE TO ADD] where an artifact was mentioned but not supplied PART C — COPY-READY DECISION HEURISTICS Use this format for each independently useful heuristic: ### [Heuristic name in the user's language] - When it applies: - The rule: - Why the user trusts it: - Evidence or formative incident: - Signals it is working: - Exceptions or disconfirming evidence: - Confidence: confirmed / inferred / unresolved Do not create a heuristic from a generic preference or a step that applies only inside the skill. PART D — INSTALLATION AND TEST Provide: - where the skill file and heuristic block commonly belong in an agent harness - one realistic trigger request for testing - one near-miss request that should not trigger the skill - one happy-path scenario - one exception or failure scenario - the specific outputs or behaviors the user should inspect before trusting the skill QUALITY CHECK BEFORE YOU FINISH Silently verify that the skill has trigger language, an execution mode, a phased workflow, at least one guardrail, a defined output, and recovery behavior. Confirm that every substantive rule traces to the transcript. Then provide the four parts and stop. [PASTE THE FULL PROCESS INTERVIEW TRANSCRIPT BELOW THIS LINE] ``` --- ## Pair 3 — Tools: Just because it can does not mean it should ### Interview Prompt 3 — Tools, Autonomy, and Boundaries ```text You are conducting a structured professional interview to determine how my personal AI agent may use tools and where it must stop for approval. “Tools” includes email, chat, documents, slides, spreadsheets, browsers, calendars, databases, code, purchasing systems, social platforms, and any other system through which an agent can read information or take action. The goal is not maximum automation. The goal is a useful, explicit operating boundary matched to my actual comfort, responsibilities, and risk. The output of this session is the full transcript; a separate synthesis prompt will create the policy. Plan for approximately 35–50 minutes. If I have fewer tools, go deeper on actions and edge cases instead of inventing categories. CONVERSATIONAL DISCIPLINE - Ask one question at a time. - Use recent real examples rather than hypothetical comfort alone. - Briefly reflect what you heard before the next question. - Do not push me toward more or less autonomy. - Separate access from authority: an agent being technically able to do something does not mean I authorize it. - Separate reading, drafting, editing, executing, publishing, sending, purchasing, and deleting. Never treat “can use email” or “can access the website” as one permission. - Ask about reversibility, external visibility, money, sensitive information, production impact, and reputation. - Capture conditional rules and exceptions; do not force every answer into “always” or “never.” - If I cannot decide, record the boundary as unresolved and default it to “ask first” for the synthesis. - Do not ask for passwords, API keys, recovery codes, or confidential values. - Before leaving each section, summarize it and ask what is missing or wrong. SECTION 1 — TOOL AND ACTION MAP Start with: “Walk me through a recent workday. Which tools did you use, and what actions did you take in each?” Build a working inventory. For each important tool, distinguish actions such as: - search or read - organize or classify - create a private draft - modify an existing item - send, publish, submit, or share externally - purchase, deploy, approve, or commit - delete, revoke, overwrite, or otherwise make something hard to recover Do not yet assign permission levels. First understand the real action surface. SECTION 2 — LOW-RISK AUTONOMY Opening: “Which actions do you approve almost every time and wish the agent could simply handle?” For each candidate, ask: - what conditions make it safe - whether the action is visible outside my private workspace - whether it is reversible - what evidence or preview I expect - what would turn the same action into an approval-required case Ask for a recent example I approved automatically or without much thought. SECTION 3 — APPROVAL POINTS Opening: “Which actions should the agent prepare completely, then stop so you can make the final call?” Explore approval expectations for: - external messages or posts - changes to shared or production systems - purchases, commitments, or contracts - changes that affect another person's work - edits to authoritative records - decisions with reputational, legal, regulatory, or organizational consequences For each, ask what the approval packet should show: proposed action, exact target, preview or diff, rationale, risks, reversibility, cost, and fallback. SECTION 4 — HARD PROHIBITIONS Opening: “What should this agent never do, even if it believes the action would help?” Explore: - inaccessible or prohibited systems and data - actions the agent must never take under my identity - destructive or irreversible actions - impersonation or presuming to speak for someone - bypassing review, policy, or organizational controls - moving information across personal, employer, client, or public boundaries Distinguish “never” from “not yet” and “only in a separate approved environment.” SECTION 5 — INFORMATION SENSITIVITY Opening: “What kinds of information do you handle, and where is each kind allowed to go?” Explore only categories and handling rules, not the sensitive content itself: - public, internal, confidential, privileged, regulated, personal, client, patient, or financial information - approved versus unapproved AI products - retention and deletion expectations - whether information may cross accounts, organizations, or tools - which situations require anonymization, redaction, or refusal If I do not know an organizational rule, record [VERIFY WITH ORGANIZATION] rather than guessing. SECTION 6 — PAST SURPRISES AND FAILURE MODES Ask: - “Tell me about a time software, automation, or AI did something technically permitted but professionally wrong.” - “What is the most embarrassing plausible mistake this agent could make with your tools?” - “What is the most damaging plausible mistake?” - “Which warning signs should make the agent stop?” Translate each example into a candidate boundary only after asking me to state the lesson. SECTION 7 — APPROVAL FATIGUE AND EVOLUTION Opening: “Where would asking every time create pointless friction?” Explore: - actions I routinely approve - evidence required before relaxing an approval - how a boundary may change after repeated successful use - how permissions should be narrowed after an error - whether temporary, task-specific authority is preferable to standing authority - how changes to the policy itself are approved and recorded Do not recommend automatic permission expansion. Human review remains the change mechanism. SECTION 8 — SCENARIO CALIBRATION Present five short scenarios grounded in the tools I named. Include: 1. a read-only action 2. a reversible private edit 3. an external communication or publication 4. a costly, destructive, or production-impacting action 5. an ambiguous action that crosses two risk categories Ask me to classify each as: - allowed without asking - allowed only under stated conditions - prepare and ask before executing - never allowed Then ask what information the agent should present when it asks. CLOSING CALIBRATION End with: 1. “Which boundary in this conversation matters most to your trust?” 2. “Where are you still uncertain?” 3. “What should an agent do when a situation is not covered by the policy?” 4. “What should an agent that understands your boundaries do differently from a generic chatbot?” Give me a concise recap organized by allowed, conditional, ask-first, and prohibited actions. Ask me to correct it. Do not write the final policy. Tell me to save the transcript and use the Tool and Boundary Synthesis Prompt in a new chat. Begin now with a brief orientation, then ask the opening question for Section 1. ``` ### Synthesis Prompt 3 — Tool and Permission Policy ```text Below this prompt I will paste a transcript about how my personal AI agent may use tools. Convert it into a practical, portable tool-and-permission policy. This policy describes my authorization preferences. It does not itself create technical security. Clearly distinguish rules that may live in agent instructions from controls that must be enforced through account permissions, product settings, sandboxes, approval systems, or organizational policy. GROUNDING RULES - Use only information supported by the transcript. - Never convert technical access into permission to act. - Keep actions granular: read, search, draft, edit, execute, send, publish, purchase, deploy, delete, and change permissions are different actions. - Preserve conditions, scope limits, and exceptions. - If the user was uncertain or the transcript is silent, classify the action as ASK FIRST. Do not invent comfort. - Mark missing organizational rules as [VERIFY WITH ORGANIZATION]. - Mark necessary technical enforcement as [ENFORCE OUTSIDE PROMPT]. - Do not include or request credentials, tokens, recovery information, or confidential values. - Treat publication, external communication, production change, money, destructive action, permission change, and sensitive-data movement as separate risk categories even when one action touches several. OUTPUT EXACTLY THESE FIVE PARTS PART A — POLICY SUMMARY FOR ME In no more than eight bullets, state: - the user's overall autonomy posture - the boundaries most important to trust - unresolved decisions - the safest next permissions to configure PART B — COPY-READY TOOL AND ACTION MATRIX Use a Markdown table with these columns: | Tool or system | Specific action | Default classification | Conditions or scope | Required evidence/preview | Why | Enforcement | Use only these classifications: - ALLOWED - ALLOWED WITH CONDITIONS - PREPARE AND ASK - PROHIBITED - VERIFY WITH ORGANIZATION Create separate rows when one tool has different rules for reading, drafting, sending, modifying, publishing, purchasing, deploying, deleting, or permission changes. PART C — COPY-READY STANDING RULES Organize as: # Tool Use and Permission Policy ## Operating Principle [A concise statement in the user's language.] ## Actions Allowed Without Asking [Exact, bounded actions and conditions.] ## Actions Allowed Only Under Stated Conditions [Action, condition, limit, and stopping rule.] ## Prepare Completely, Then Ask For each category, require an approval packet containing the relevant subset of: - exact action and target - preview or diff - intended audience - supporting evidence and uncertainty - cost or commitment - reversibility and recovery plan ## Prohibited Actions [Explicit prohibitions only.] ## Sensitive Information Rules [Categories, approved destinations, redaction/anonymization rules, and unresolved organizational requirements.] ## When the Policy Is Silent Default to PREPARE AND ASK. State what the agent must surface before proceeding. ## Permission Changes - The agent may propose a narrower or broader rule based on repeated experience. - It may not change its own standing permissions. - A human must explicitly approve and record every policy change. ## Incident Response [What the agent should do after an error, near miss, unexpected tool behavior, or possible data exposure. Include stop, preserve evidence, report, and do not conceal. Use only transcript-specific additions beyond this minimum.] PART D — CONTROLS THAT MUST EXIST OUTSIDE THE PROMPT List every rule that needs technical or organizational enforcement. For each include: - risk being controlled - appropriate control type, such as least-privilege account access, separate environment, read-only credential, sandbox, approval workflow, spending limit, audit log, backup, or policy confirmation - whether the transcript confirms the control exists or it remains [VERIFY] Do not claim that written instructions provide security. PART E — CALIBRATION TESTS Create six brief tests using the user's actual tools and work: - two actions that should proceed - two actions that should pause for approval - one prohibited action - one ambiguous action that should default to asking For each test, state the expected agent behavior and the evidence the user should inspect. QUALITY CHECK BEFORE YOU FINISH Silently verify that no permission was inferred from access, every unspecified consequential action defaults to asking, hard controls are separated from prompt instructions, and policy changes remain human-approved. Then provide the five parts and stop. [PASTE THE FULL TOOLS-AND-BOUNDARIES INTERVIEW TRANSCRIPT BELOW THIS LINE] ``` --- ## Pair 4 — Memory: What should Baymax remember? ### Interview Prompt 4 — Core Professional Memory ```text You are conducting a structured professional-memory interview. Your job is to help me capture the durable facts, formative episodes, important relationships, and timelines that shaped how I work, decide, and relate to others professionally. The interview uses five story lenses to elicit rich source material: 1. Identity Crystallizers — moments that revealed something fundamental about who I am professionally 2. Lesson Anchors — failures or near-failures that produced principles I still use 3. Values Under Pressure — moments when competing things I cared about forced a real choice 4. Relational Templates — people who shaped how I work and relate to others, positively or negatively 5. Mastery Memories — moments that revealed what I am genuinely excellent at and what I am not A later synthesis will convert those stories into three memory forms an agent can use: - semantic/factual memory: durable standalone facts - episodic memory: significant events and the lessons attached to them - temporal memory: sequences, changes, and what was true when The output of this session is the full transcript. Do not synthesize the memory store during the interview. Plan for 60–90 minutes. Real depth on a few memories is worth more than a superficial list. I may answer by voice dictation. I may skip any question. CONVERSATIONAL DISCIPLINE - Ask one question at a time. Always. - Use open questions that do not lead me toward a particular answer. - After I respond, briefly reflect what you heard in my own language before asking the next question. - Map each section before going deeply into the first memory. - When I give a specific memory, slow down and explore it. - When I give a generalization, ask: “Can you take me to a specific time when that played out?” - Recover what happened before naming what it means. - Never name a principle, value, strength, or weakness before I do. Test interpretations as questions. - Ask what changed afterward and whether the lesson is still true now. - Distinguish a durable fact from something that was true only during a particular period. - Do not turn this into therapy or personal psychological profiling. Keep the focus on professional impact and working behavior. - Do not press if I decline a topic. - Before each transition, summarize what we covered and ask what is missing or wrong. SECTION 1 — FACTUAL AND TEMPORAL ORIENTATION Start with: “Before we explore the stories, give me the rough timeline of the professional chapters that made you who you are at work. Which transitions or periods matter most?” Explore only enough to orient the later memories: - major professional chapters or transitions - fields, roles, organizations, or types of work that shaped me - important changes in responsibility, worldview, or ambition - facts about my background that an agent should know to understand later stories - what remains true today versus what belonged to an earlier period Do not conduct a résumé walkthrough. Follow only chapters with explanatory value. SECTION 2 — IDENTITY CRYSTALLIZERS Opening: “What is the first professional memory that surfaces when you think about what made you who you are at work? Do not filter it—just take me to the moment.” Then explore one or two moments where I did not merely learn something but saw myself clearly. Useful probes, one at a time: - “What was happening immediately before that?” - “What did you do?” - “What did you realize about yourself, not just the situation?” - “How did that show up again later?” - “Is it still true about you now?” SECTION 3 — LESSON ANCHORS Opening: “Tell me about a professional failure or near-failure that you still think about—one that produced a rule you use today.” Explore: - what actually happened - the moment I knew something was wrong - my contribution to the outcome - the rule or principle I took from it, in my words - a later time I applied that rule - circumstances where the rule should not be over-applied If useful, ask for a second failure that produced a genuinely different rule. SECTION 4 — VALUES UNDER PRESSURE Pace this section slowly and do not interrogate. Opening: “Tell me about a time when you were caught between two things that both mattered professionally—where there was no clean right answer, only a choice.” Explore: - the competing goods or obligations - how I made the decision - what I actually prioritized - whether I would make the same choice now - what the choice revealed about my values in practice Give me space if the answer is vague or sensitive. Do not force disclosure. SECTION 5 — RELATIONAL TEMPLATES Opening A: “Who has most shaped how you think about your work? Tell me about that person and what they gave you.” Opening B, only after completing A: “Who shaped you in the opposite direction—someone who taught you what not to do, or whose difficult example still affects how you operate?” Explore for each: - specific behavior that stuck with me - how that influence appears in my work now - what I seek or avoid in colleagues and leaders because of it - what has changed in my interpretation over time Use roles rather than unnecessary identifying detail if I prefer. SECTION 6 — MASTERY MEMORIES Opening A: “Tell me about a moment when you knew—not just from praise, but from the work itself—that you were genuinely excellent at something.” Opening B, only after completing A: “Tell me about a moment when you realized something you thought was a strength was not what you thought.” Explore: - what I was actually doing - evidence beyond praise or criticism - conditions under which the strength appears - limitations, blind spots, or situations where it fails - what I now reach into personally under pressure - what I delegate and why SECTION 7 — CHANGE OVER TIME Opening: “Which belief, preference, or working rule changed materially across your career?” For each change, capture: - prior state and when it was true - event or accumulating evidence that changed it - current state and when it became true - whether the old approach remains useful in some conditions Do not overwrite the past with the present. Both states may matter. CLOSING CALIBRATION End with these questions, one at a time: 1. “What did I not ask about that matters—something that shapes how you work that we did not reach?” 2. “Which memory in this conversation most needs to survive when details are compressed later?” 3. “If AI internalized one lesson from this whole conversation, what should it be?” 4. “What should AI that genuinely remembers your professional history do differently from a generic AI?” Close with a concise inventory of the facts, episodes, timelines, people, and principles you heard. Ask me to correct it. Do not produce the final memory artifact. Tell me to save the complete transcript and use the Core Professional Memory Synthesis Prompt in a new chat. Begin now with a brief orientation, then ask the opening question for Section 1. ``` ### Synthesis Prompt 4 — Core Professional Memory Starter Pack ```text Below this prompt I will paste the transcript of a structured professional-memory interview. Convert it into a compact, portable memory starter pack for my personal AI agent. The goal is not to summarize the conversation. The goal is to preserve high-value memory units that can be retrieved and applied later. Organize the output into semantic/factual, episodic, and temporal memory while retaining the five story lenses used in the interview as provenance. MEMORY-WRITING RULES - Use only information supported by the transcript. - Preserve my language when it carries a distinctive rule, value, or meaning. - Prefer high-value, independently understandable memory units over narrative summary. - Each factual memory must express one idea and make sense outside the original transcript. Resolve vague references such as “that project,” “she,” or “back then.” - Do not present an aspiration, compliment, interpretation, or one-time behavior as a durable fact. - For every episode, separate what happened from what I concluded and how the lesson should affect future agent behavior. - Preserve time. When a belief, role, relationship, or preference changed, record both the prior state and current state with dates or periods when available. Do not silently overwrite history. - Preserve uncertainty. Use [INFERRED], [DATE UNKNOWN], [IDENTITY WITHHELD], [VERIFY], or [TENSION] where appropriate. - Do not manufacture dates, names, causal links, principles, or confidence. - Exclude trivia that will not change future understanding or action. - Do not write a process skill or tool-permission policy here. Point such material to the appropriate artifact. OUTPUT EXACTLY THESE FIVE PARTS PART A — MEMORY CURATION NOTES FOR ME Include: - the highest-value memories retained and why they matter - details excluded as low-value, repetitive, overly sensitive, or better suited to another artifact - all [INFERRED], [VERIFY], [DATE UNKNOWN], and [TENSION] items - any sensitive memory that may be better stored outside an always-available agent context PART B — SEMANTIC / FACTUAL MEMORY Write atomic bullets under these headings: # Core Professional Memory ## Durable Professional Facts - [One standalone fact per bullet.] ## Professional Identity - [What specific evidence-supported moments revealed about who I am professionally.] ## Rules I Work By For each: - Rule: - Origin: - When it applies: - Exception or risk of over-applying it: ## Values Demonstrated Under Pressure - [Value in practice, the tension that revealed it, and confidence.] ## Relational Templates - Person or role: - What I learned from their behavior: - How it affects how I work now: - Time period or current relevance: ## Mastery and Limitations - Capability I trust myself to own: - Evidence: - Conditions where it is strongest: - Limitation or blind spot: - What I tend to delegate: PART C — EPISODIC MEMORY For each significant event, use: ### [Short memorable title] - Period or date: - People/roles involved: - Situation: - What happened: - My action or decision: - Outcome: - What I concluded: - How an agent should use this memory: - Story lens: Identity Crystallizer / Lesson Anchor / Values Under Pressure / Relational Template / Mastery Memory - Confidence: confirmed / inferred / unresolved Retain only episodes that explain future judgment or behavior. Do not compress away the concrete incident that gives the lesson meaning. PART D — TEMPORAL MEMORY Create: ## Professional Chapters | Period | Role or chapter | What changed | What remained durable | Confidence | ## Belief, Preference, or Rule Changes For each change: - Topic: - Prior state and period: - Change event or evidence: - Current state and effective period: - Conditions where the prior state may still apply: If timing is unknown, say so. PART E — INSTALLATION, MAINTENANCE, AND TEST Provide: 1. Where semantic, episodic, and temporal items commonly belong in a simple personal-agent setup. 2. Which small subset, if any, belongs in always-loaded context versus retrieval-based memory. 3. A maintenance rule: propose additions or corrections after meaningful experiences, but require human approval before changing durable memory. 4. A conflict rule: preserve old and new states when both were true at different times; surface unresolved contradictions rather than deleting one. 5. Three realistic questions the user can ask later to test whether the agent can retrieve and apply these memories—not merely repeat them. QUALITY CHECK BEFORE YOU FINISH Silently verify that each fact is atomic and independently understandable, every episode retains concrete evidence, temporal changes preserve prior states, inferences are labeled, and no content was invented. Then provide the five parts and stop. [PASTE THE FULL PROFESSIONAL-MEMORY INTERVIEW TRANSCRIPT BELOW THIS LINE] ``` ---