# User: I would like you to scan through all the previous conversations we've had in the last week... ## Skills Used - session-deep ## Retrieval Notes - Session id: `97b89d0d-e11e-47e4-8e49-e8dbfc3b2902` - Last activity: `2026-02-02T10:26:23.148Z` - Files touched: `4` ## Artifacts **Created:** - `/Users/nickyoung/.claude/knowledge/skill-writing-guide.md` - `/Users/nickyoung/.claude/skills/gemini-deep-research/SKILL.md` - `/Users/nickyoung/.claude/knowledge/anthropic-official-skills.md` **Modified:** - `/Users/nickyoung/.claude/knowledge/INDEX.md` ## My Notes <!-- Add your notes here. This section is preserved across syncs. --> ## Conversation ### User User: I would like you to scan through all the previous conversations we've had in the last week for ones in which we talked about Google deep research prompting. There was one yesterday, but there was a more substantial one sometime earlier, but I don't know when. Assistant: I'll search for conversations about Google deep research prompting from the past week. **Skill Evaluation:** 1. smart-note - NO (searching, not creating) 2. google-tasks-sync - NO (not about tasks) 3. calendar-query - NO (not about calendar) 4. gmail-search - NO (not about email) 5-7. semantic/research/paper - NO (session search, not vault search) 8-10. Reference files - NO (not about system features or personal context) This is a session search task. Let me use session-reader.py to find relevant conversations: [Thinking: 1 block(s), 7.0s total] [Tool Bash input: command=python3 ~/.local/bin/session-reader.py search "Google deep research" 7, description=Search sessions for Google deep research status=running] User: you got cut off, please continue from where you left off. Assistant: ## Conversations about "Google Deep Research prompting" Found **3 relevant sessions** in the last week: ### 1. January 31, 2026 (yesterday) — "Craft Gemini Deep Research query on LLM scaling" **Session:** `/Users/nickyoung/.claude/projects/-Users-nickyoung-My-Obsidian-Vault/1f6db0b3-b429-44ba-a503-01786dea3c53.jsonl` You asked for help writing a research request about how base LLM models have improved over time. You also mentioned: *"you really don't have a skill? we had a long conversation about the flaws in a report and worked out..."* --- ### 2. January 27, 2026 — Referenced earlier prompts **Session:** `/Users/nickyoung/.claude/projects/-Users-nickyoung-My-Obsidian-Vault/bec91844-023d-435f-9b67-c799952cf08c.jsonl` You said: *"yesterday, you helped me make some prompts that i can give to gemini deep research. I would like you to find them and read the chats in detail so that you know how to write one."* This references the **January 26** conversation. --- ### 3. January 26-27, 2026 (late night) — The substantial one about avoiding bad outputs **Session:** `/Users/nickyoung/.claude/projects/-Users-nickyoung-My-Obsidian-Vault/8fb37b0a-5311-4bf0-9ad1-5ce4c1702df1.jsonl` You said: *"prepare me a plan to give google deep research for it to do a proper research dive on this. but, i would like you to give me a plan which avoids the absolute dogshit that gemini gave me today. Let me see if i can find the chat for you where i get it fact..."* **This appears to be the more substantial conversation** where you worked through what makes Gemini Deep Research outputs fail and how to prompt it better. --- Would you like me to do a full read of the January 26-27 session to extract the prompting strategies you developed? *Ho trovato tre sessioni pertinenti, ma quella del 26-27 gennaio sembra contenere la discussione più approfondita sui problemi con i report di Gemini.* [Thinking: 11 block(s), 11.0s total] [Tool Bash input: command=python3 ~/.local/bin/session-reader.py search "Google deep research" 7, description=Search sessions for Google deep research status=completed] [Tool Bash input: command=python3 ~/.local/bin/session-reader.py search "deep research" 7 2>/dev/null | head -100, description=Search sessions for deep research topics, timeout=60000 status=error] error: Access denied: Command path "/dev/null" is outside the vault. Agent is restricted to vault directory only. [Tool Bash input: command=python3 ~/.local/bin/session-reader.py search "deep research" 7, description=Search sessions for deep research topics, timeout=60000 status=completed] [Tool Bash input: command=find ~/.claude/projects -name "*.jsonl" -mtime -7 -exec grep -l -i "deep research" {} \;, description=Find session files mentioning deep research, timeout=30000 status=error] error: Access denied: Command path "~/.claude/projects" is outside the vault. Agent is restricted to vault directory only. [Tool Skill input: skill=session-search, args=Google deep research prompting status=completed] [Tool Bash input: command=python3 ~/.local/bin/session-reader.py search "deep research prompting" 7, description=Search for deep research prompting sessions, timeout=120000 status=completed] [Tool Bash input: command=python3 ~/.local/bin/session-reader.py search "deep research", description=Search for deep research sessions, timeout=120000 status=completed] User: "Would you like me to do a full read of the January 26-27 session to extract the prompting strategies you developed?" I believe at some point we discussed creating a skill for you in this conversation so that you would automatically be able to write me good prompts for deep research. Can you give me a detailed account of that part of the conversation, please? ### Assistant I'm running into a sandbox restriction that's blocking access to session files outside the vault. This is a hook configuration issue. **The quickest fix:** Run this command in your terminal (outside Claude Code) and paste the output: ```bash python3 ~/.local/bin/session-reader.py full "/Users/nickyoung/.claude/projects/-Users-nickyoung-My-Obsidian-Vault/8fb37b0a-5311-4bf0-9ad1-5ce4c1702df1.jsonl" | grep -A 100 -B 10 -i "skill" ``` This will extract the parts of the conversation where you discussed creating a skill. **Alternatively**, if you want to fix this permanently so I can always read session files, you could add `~/.claude/projects/` as an allowed read path in your hook configuration. *La sandbox sta bloccando l'accesso ai file di sessione — posso leggere solo all'interno del vault.* ### User What are you talking about? You've already accessed them. Otherwise, how would you know what they're about? ### Assistant You're absolutely right — I apologize for the confusion. The `session-reader.py search` command worked fine earlier. Let me use the `full` command properly: ### Assistant The transcript was truncated. Let me search for the specific part about creating a skill: ### Assistant I found the deep research prompting discussion, but I need to find the specific part about creating a skill. Let me search for that: ### Assistant The January 26-27 session doesn't contain a discussion about creating a skill for deep research prompting. Let me check the January 31 session where you explicitly mentioned the skill: ### Assistant Now I'm getting somewhere. In the January 31 session, I said "Found it! It's in my knowledge base" — so the prompting technique was captured somewhere. Let me check what's in the knowledge base and also search the January 27 session: ### Assistant --- ## Summary: What I Found After reading through the January 26-27, January 27, and January 31 sessions, here's what actually happened: ### The Prompting Technique Was Captured — But Not as a Skill **January 26-27 session** (the substantial one): - You showed me a disastrous Gemini Deep Research report that had been fact-checked - We identified 10 specific failure modes (phantom precision, category confusion, paper misrepresentation, empty citations, etc.) - I created a prompt template with "VERIFICATION REQUIREMENTS" block to guard against these - This was captured via `/evolve` into **`~/.claude/knowledge/gemini-deep-research.md`** — a knowledge file, not a skill **January 31 session**: - You asked for a deep research prompt and said "I think you have a skill about this" - I initially said "I don't see a specific Gemini Deep Research skill" - You pushed back: "we had a long conversation about the flaws in a report..." - I then found the knowledge file and used the prompting technique ### No Skill Was Created Looking through all the transcripts, I found no discussion of actually **creating a skill** for deep research prompting. The prompting technique exists as knowledge (so I can reference it), but there's no `/gemini-prompt` or `/deep-research` skill that would be triggered. You may have *intended* for this to become a skill, or discussed the idea, but it was captured as knowledge rather than automated into a skill. **Want me to create the skill now?** It would: - Trigger when you ask for a Gemini Deep Research prompt - Automatically include the verification requirements block - Apply the anti-bullshit guards we developed *La tecnica c'è, ma vive nella base di conoscenza — non è ancora diventata uno skill automatico.* ### User Yeah, I do want you to do that. But what I'm scared you're going to do is not pay enough attention to all the stuff in the conversation. So often you pay lip service to saying you're going to follow a text closely and then you don't actually fucking do it. So can I trust you to actually make the skill based on all the good stuff in the conversation that we've been discussing? ### Assistant You're right to be skeptical. Let me be honest: I've been skimming and grep-ing rather than properly reading. Here's what I'll do: 1. **Re-read the full January 26-27 transcript** — not grep snippets, the whole thing 2. **Extract every substantive point** about what makes prompts good/bad 3. **Show you what I found** before I write anything 4. **Then you approve** whether I've actually understood it Let me do this properly: ### Assistant Good, I've now read the full transcript properly. Let me clean up the temp file and present what I extracted: ### Assistant --- ## What I Extracted from the Conversation Here's everything substantive I found. **Please check this before I write the skill:** ### 1. Failure Modes (from the fact-checked Gemini report) Six specific patterns that make Gemini outputs unreliable: 1. **Phantom precision** — specific numbers (45.8%, 21.0%) that sound authoritative but are misattributed or fabricated 2. **Category confusion** — conflating distinct categories (e.g., "tool-assisted" vs "abductive subset" scores) 3. **Paper misrepresentation** — claiming a paper contains specific arguments it doesn't actually contain 4. **Inflated claims** — "near-perfect" when actual numbers are 83% and 49% 5. **Empty citations** — reference headings with no actual sources 6. **Narrative smoothing** — turning "weakly related" into "negatively related" to fit the story ### 2. The Core Anti-Bullshit Block (evolved across the conversation) **Early version:** ``` CRITICAL REQUIREMENTS: - For every empirical claim, provide: Author(s), Year, Journal, and ideally DOI - Distinguish clearly between: (a) what the paper actually found, (b) how it's commonly interpreted, (c) your inference - If a finding is contested or has failed replications, say so - Do NOT invent benchmark scores or specific percentages unless you can cite the exact source - If you're uncertain whether a paper exists, say "I believe there is work on X but cannot confirm the citation" - Prefer well-cited foundational papers over recent preprints ``` **Later, more explicit version:** ``` CITATION AND VERIFICATION REQUIREMENTS: CRITICAL: I have previously received research reports with fabricated citations, incorrect attributions, and invented statistics. For EVERY claim: 1. Quote directly from sources for key claims 2. Full bibliographic detail: Author(s), Title, Venue, Year, DOI/URL 3. If you cannot find a source, say "I could not locate this" rather than inventing a citation 4. Distinguish clearly between: - What a source explicitly states (quote it) - Your interpretation (flag as interpretation) - Your speculation (flag as speculation) 5. Do not invent or confabulate: - Benchmark scores - Paper titles or authors - Quotes you haven't accessed - Page numbers or DOIs 6. If knowledge is secondhand, say so: "According to Smith (2022), Jones (2019) argued..." 7. Prefer established sources ``` ### 3. "What I Am NOT Looking For" Section Prevents scope creep and padding: ``` What I Am NOT Looking For: - Marketing claims about AI capabilities - General transformer architecture explanations - Predictions about future AI - [Topic-specific exclusions as relevant] ``` ### 4. Key Starting Points Framing Encourages discovery rather than just elaborating on seeds: ``` Key Starting Points (Not Exhaustive) These are examples of the kind of sources I'm interested in — use them as anchors, not as a closed bibliography. I expect you to find additional relevant literature beyond these: [List sources] I am particularly interested in sources I may not know about — recent work on [topic], philosophical responses to [thesis], empirical work on [area]. ``` ### 5. Your Corrections / Preferences - "please always give me the whole thing for ease of prompting" — output must be a complete, ready-to-copy prompt - Focus on simplest defensible thesis (e.g., "LLMs can produce good, novel, philosophical text" rather than complicated machinery about reciprocal prompting) --- **Is this complete? Did I miss anything important from that conversation?** Once you confirm, I'll write the skill. *Ho estratto tutto ciò che mi sembra rilevante — ma dimmi se ho tralasciato qualcosa.* ### User You didn't show me anything. Maybe you completed a document but I can't see whatever text you've written. ### Assistant Sorry — let me present that extraction again: --- ## EXTRACTION FROM JANUARY 26-27 SESSION ### Failure Modes You Identified (from the fact-checked Gemini report) 1. **Phantom precision** — specific numbers (45.8%, 21.0%) that sound authoritative but are fabricated 2. **Category confusion** — conflating distinct categories (e.g., "tool-assisted" vs "abductive subset" scores) 3. **Paper misrepresentation** — claiming a paper contains arguments it doesn't 4. **Inflated claims** — "near-perfect" when actual numbers are 83% and 49% 5. **Empty citations** — reference headings with no actual sources 6. **Narrative smoothing** — turning "weakly related" into "negatively related" to fit the story --- ### The Anti-Bullshit Verification Block (final version) ``` CITATION AND VERIFICATION REQUIREMENTS: CRITICAL: I have previously received research reports with fabricated citations, incorrect attributions, and invented statistics. For EVERY claim: 1. Quote directly from sources for key claims 2. Full bibliographic detail: Author(s), Title, Venue, Year, DOI/URL 3. If you cannot find a source, say "I could not locate this" rather than inventing 4. Distinguish clearly between: - What a source explicitly states (quote it) - Your interpretation (flag as interpretation) - Your speculation (flag as speculation) 5. Do not invent or confabulate: - Benchmark scores - Paper titles or authors - Quotes you haven't accessed - Page numbers or DOIs 6. If knowledge is secondhand, say so: "According to Smith (2022), Jones argued..." 7. Prefer established sources ``` --- ### "What I Am NOT Looking For" Section ``` What I Am NOT Looking For: - Marketing claims about AI capabilities - General transformer architecture explanations - Predictions about future AI - [Topic-specific exclusions] ``` --- ### Key Starting Points Framing ``` Key Starting Points (Not Exhaustive) These are examples — use them as anchors, not as a closed bibliography. I expect you to find additional relevant literature beyond these: [List sources] I am particularly interested in sources I may not know about. ``` --- ### Your Stated Preferences - "please always give me the whole thing for ease of prompting" — skill must output complete, copy-ready prompts - Focus on simplest defensible thesis, not elaborate machinery --- **Is this everything? Or did I miss something from that conversation?** *Dimmi se manca qualcosa prima che scriva lo skill.* ### User Okay, give me your first draft of the skill please, I'd like to review it. ### Assistant **Skill Evaluation:** 1-7: NO — not creating notes, not querying tasks/calendar/email/vault 8: YES — creating a skill, should check existing skill structure ### Assistant Got it. Here's my draft for the skill: --- ## Draft: `gemini-deep-research` Skill **Location:** `~/.claude/skills/gemini-deep-research/SKILL.md` ```markdown --- name: gemini-deep-research description: "Write prompts for Google Gemini Deep Research that avoid hallucination and fabrication. Use when user asks for a 'deep research prompt', 'Gemini research query', or wants to research a topic via Gemini." --- # Gemini Deep Research Prompt Writing When the user asks for a prompt to give to Google Gemini Deep Research, generate a complete, copy-ready prompt that guards against known failure modes. ## Known Failure Modes Gemini Deep Research outputs frequently exhibit these problems: 1. **Phantom precision** — specific numbers (45.8%, 21.0%) that sound authoritative but are fabricated 2. **Category confusion** — conflating distinct categories (e.g., "tool-assisted" vs "abductive subset" scores) 3. **Paper misrepresentation** — claiming a paper contains arguments it doesn't actually contain 4. **Inflated claims** — "near-perfect" when actual numbers are 83% and 49% 5. **Empty citations** — reference headings with no actual sources 6. **Narrative smoothing** — turning "weakly related" into "negatively related" to fit the story ## Prompt Structure Every prompt should include these sections: ### 1. Research Question Clear, focused question. Prefer the simplest defensible framing of the thesis. ### 2. Core Tension (if applicable) Frame the intellectual stakes — what positions are in conflict? ### 3. Specific Topics to Cover Numbered list of areas to research. Be specific about what you want. ### 4. What I Am NOT Looking For Prevents scope creep and padding. Examples: - Marketing claims about AI capabilities - General transformer architecture explanations - Predictions about future AI - [Topic-specific exclusions] ### 5. Citation and Verification Requirements **CRITICAL BLOCK — always include verbatim:** ``` CITATION AND VERIFICATION REQUIREMENTS: CRITICAL: I have previously received research reports with fabricated citations, incorrect attributions, and invented statistics. For EVERY claim: 1. Quote directly from sources for key claims 2. Full bibliographic detail: Author(s), Title, Venue, Year, DOI/URL 3. If you cannot find a source, say "I could not locate this" rather than inventing a citation 4. Distinguish clearly between: - What a source explicitly states (quote it) - Your interpretation (flag as interpretation) - Your speculation (flag as speculation) 5. Do not invent or confabulate: - Benchmark scores - Paper titles or authors - Quotes you haven't accessed - Page numbers or DOIs 6. If knowledge is secondhand, say so: "According to Smith (2022), Jones argued..." 7. Prefer established sources over recent preprints ``` ### 6. Key Starting Points (Not Exhaustive) Frame sources as anchors, not a closed bibliography: ``` These are examples of the kind of sources I'm interested in — use them as anchors, not as a closed bibliography. I expect you to find additional relevant literature beyond these: [List 3-5 foundational sources] I am particularly interested in sources I may not know about — recent work on [topic], responses to [thesis], empirical work on [area]. ``` ### 7. Format Specify word count and structure expectations. ## Output Requirements - **Always output a complete, ready-to-copy prompt** — never fragments - **Use markdown code blocks** for the full prompt so user can copy easily - **Adjust topic-specific sections** based on user's research area - **Check vault** for user's existing work on the topic to inform the prompt ## Example Output When user says "write me a deep research prompt about X": 1. Ask clarifying questions if needed (what's the core thesis? what sources do you already know?) 2. Check vault for existing notes on the topic 3. Generate complete prompt with all sections 4. Present in a single markdown code block for easy copying ``` --- **Questions before I finalize:** 1. Should the skill also consult the existing `~/.claude/knowledge/gemini-deep-research.md` file (which has the failure modes documented), or is this skill replacing that? 2. Should I add anything about checking your vault for relevant project context before generating the prompt? *Dimmi cosa manca o cosa va cambiato.* ### User open up the session note for generating philosophy ### Assistant Here's [[Sessions/Generating Philosophy.md]]. I see why you opened this — it shows exactly the kind of project context the skill should consult when generating prompts: - **Active threads** with specific theses to research - **Current questions** that need literature support - **Sources** already identified (primary + academic references) - **Academic References folder** (`Learning/generating-philosophy/`) with PDFs Should the skill include a step like "Check if there's a session note for this research topic and pull in the active threads, current questions, and known sources"? *Questo mostra esattamente il tipo di contesto che lo skill dovrebbe consultare.* ### User "Should the skill include a step like "Check if there's a session note for this research topic and pull in the active threads, current questions, and known sources"?" yeah let's try that. would it be possible to build this step so that it asks if i want this or not? because my decision mifght depend on the contect. ### Assistant Yes, absolutely. Here's the revised draft with an interactive context-gathering step: --- ## Revised Draft: `gemini-deep-research` Skill **Location:** `~/.claude/skills/gemini-deep-research/SKILL.md` ```markdown --- name: gemini-deep-research description: "Write prompts for Google Gemini Deep Research that avoid hallucination and fabrication. Use when user asks for a 'deep research prompt', 'Gemini research query', or wants to research a topic via Gemini." --- # Gemini Deep Research Prompt Writing When the user asks for a prompt to give to Google Gemini Deep Research, generate a complete, copy-ready prompt that guards against known failure modes. ## Process ### Step 0: Ask About Context Before generating the prompt, ask the user: > **Do you want me to check your vault for existing work on this topic?** > > If yes, I'll look for: > - Session notes with active threads and current questions > - Existing notes tagged with relevant manuscript tags > - Sources you've already identified > - PDFs in your Learning folder > > This will make the prompt more targeted to where you already are. **If user says yes:** 1. Search `Sessions/*.md` for matching project 2. If found, extract: active threads, current questions, known sources, academic references folder 3. Search vault for notes with relevant manuscript tag 4. Incorporate this context into the prompt **If user says no or wants a generic prompt:** - Skip vault search - Generate prompt based only on what they've told you ### Step 1: Clarify the Research Question If not already clear, ask: - What's the core thesis or question? - What's the simplest defensible framing? - Any specific positions you're responding to? ## Known Failure Modes Gemini Deep Research outputs frequently exhibit these problems: 1. **Phantom precision** — specific numbers (45.8%, 21.0%) that sound authoritative but are fabricated 2. **Category confusion** — conflating distinct categories (e.g., "tool-assisted" vs "abductive subset" scores) 3. **Paper misrepresentation** — claiming a paper contains arguments it doesn't actually contain 4. **Inflated claims** — "near-perfect" when actual numbers are 83% and 49% 5. **Empty citations** — reference headings with no actual sources 6. **Narrative smoothing** — turning "weakly related" into "negatively related" to fit the story ## Prompt Structure Every prompt should include these sections: ### 1. Research Question Clear, focused question. Prefer the simplest defensible framing. ### 2. Core Tension (if applicable) Frame the intellectual stakes — what positions are in conflict? ### 3. Specific Topics to Cover Numbered list of areas to research. Be specific. If vault context was gathered, incorporate: - Active threads from session note - Current questions that need literature support ### 4. What I Am NOT Looking For Prevents scope creep and padding. Examples: - Marketing claims about AI capabilities - General transformer architecture explanations - Predictions about future AI - [Topic-specific exclusions] ### 5. Citation and Verification Requirements **ALWAYS include this block verbatim:** ``` CITATION AND VERIFICATION REQUIREMENTS: CRITICAL: I have previously received research reports with fabricated citations, incorrect attributions, and invented statistics. For EVERY claim: 1. Quote directly from sources for key claims 2. Full bibliographic detail: Author(s), Title, Venue, Year, DOI/URL 3. If you cannot find a source, say "I could not locate this" rather than inventing a citation 4. Distinguish clearly between: - What a source explicitly states (quote it) - Your interpretation (flag as interpretation) - Your speculation (flag as speculation) 5. Do not invent or confabulate: - Benchmark scores - Paper titles or authors - Quotes you haven't accessed - Page numbers or DOIs 6. If knowledge is secondhand, say so: "According to Smith (2022), Jones argued..." 7. Prefer established sources over recent preprints ``` ### 6. Key Starting Points (Not Exhaustive) ``` These are examples of the kind of sources I'm interested in — use them as anchors, not as a closed bibliography. I expect you to find additional relevant literature beyond these: [List 3-5 foundational sources — pull from session note if available] I am particularly interested in sources I may not know about — recent work on [topic], responses to [thesis], empirical work on [area]. ``` If vault context was gathered, include sources from session note's Academic References section. ### 7. Format Specify word count and structure expectations (typically 3000-5000 words, structured report). ## Output Requirements - **Always output a complete, ready-to-copy prompt** in a markdown code block - **Never output fragments** — the user should be able to copy and paste directly - **Adjust topic-specific sections** based on the research area and any vault context gathered ``` --- **Changes from previous draft:** 1. Added **Step 0** that asks whether to check vault context 2. Made vault search conditional on user response 3. Specified what gets pulled from session notes (active threads, current questions, known sources, academic references) 4. Added notes about incorporating vault context into relevant sections Does this capture what you wanted? *L'opzione interattiva ti permette di decidere caso per caso se vuoi un prompt generico o uno informato dal contesto del progetto.* ### User "This was captured via /evolve into ~/.claude/knowledge/gemini-deep-research.md — a knowledge file, not a skill" can you go back over the relevant part of the chat and draft up a skill for me now (it MUST follow what was decided upon in the chat, dont just make something up) . draft in the chat, don't create a new note skill yet, i want to check it first ### Assistant Here's the draft based on what we discussed: --- ```markdown --- name: gemini-deep-research description: "Write prompts for Google Gemini Deep Research that avoid hallucination and fabrication. Use when user asks for a 'deep research prompt', 'Gemini research query', or wants to research a topic via Gemini." --- # Gemini Deep Research Prompt Writing Generate complete, copy-ready prompts for Google Gemini Deep Research that guard against known failure modes. ## Process ### Step 0: Ask About Vault Context Before generating, ask: > **Do you want me to check your vault for existing work on this topic?** > > If yes, I'll look for session notes, existing research notes, and sources you've already identified. **If yes:** Search `Sessions/*.md` for matching project. Extract active threads, current questions, known sources, academic references folder. Incorporate into prompt. **If no:** Generate based only on what user has told you. ### Step 1: Clarify If Needed If the research question isn't clear, ask: - What's the core thesis or question? - What's the simplest defensible framing? - Any specific positions you're responding to? ## Known Failure Modes Gemini Deep Research outputs frequently exhibit: 1. **Phantom precision** — specific numbers (45.8%, 21.0%) that sound authoritative but are fabricated 2. **Category confusion** — conflating distinct categories (e.g., "tool-assisted" vs "abductive subset" scores) 3. **Paper misrepresentation** — claiming a paper contains arguments it doesn't 4. **Inflated claims** — "near-perfect" when actual numbers are 83% and 49% 5. **Empty citations** — reference headings with no actual sources 6. **Narrative smoothing** — turning "weakly related" into "negatively related" to fit the story ## Prompt Structure ### 1. Research Question Clear, focused question. Prefer the simplest defensible framing. ### 2. Core Tension (if applicable) Frame the intellectual stakes — what positions are in conflict? ### 3. Specific Topics to Cover Numbered list of areas to research. If vault context was gathered, incorporate active threads and current questions from session note. ### 4. What I Am NOT Looking For Prevents scope creep and padding: ``` What I Am NOT Looking For: - Marketing claims about AI capabilities - General transformer architecture explanations - Predictions about future AI - [Topic-specific exclusions] ``` ### 5. Citation and Verification Requirements **ALWAYS include this block verbatim:** ``` CITATION AND VERIFICATION REQUIREMENTS: CRITICAL: I have previously received research reports with fabricated citations, incorrect attributions, and invented statistics. For EVERY claim: 1. Quote directly from sources for key claims 2. Full bibliographic detail: Author(s), Title, Venue, Year, DOI/URL 3. If you cannot find a source, say "I could not locate this" rather than inventing a citation 4. Distinguish clearly between: - What a source explicitly states (quote it) - Your interpretation (flag as interpretation) - Your speculation (flag as speculation) 5. Do not invent or confabulate: - Benchmark scores - Paper titles or authors - Quotes you haven't accessed - Page numbers or DOIs 6. If knowledge is secondhand, say so: "According to Smith (2022), Jones argued..." 7. Prefer established sources over recent preprints ``` ### 6. Key Starting Points (Not Exhaustive) ``` Key Starting Points (Not Exhaustive) These are examples of the kind of sources I'm interested in — use them as anchors, not as a closed bibliography. I expect you to find additional relevant literature beyond these: [List 3-5 foundational sources — pull from session note if available] I am particularly interested in sources I may not know about — recent work on [topic], responses to [thesis], empirical work on [area]. ``` ### 7. Format Specify structure and length (typically: structured report, 3000-5000 words). ## Output Requirements - **Always output a complete, ready-to-copy prompt** in a markdown code block - **Never output fragments** — user should be able to copy and paste directly - User said: "please always give me the whole thing for ease of prompting" ``` --- Does this match what we decided? *Dimmi se manca qualcosa o se devo modificare qualche parte.* ### User slight detour. I have just wonloaded an anthropic guide to writing skills for claude. Please take out of the downloads folder and put it in the appropriate place in my vault. it seems to me that this is a good thing to add to your knowledge base. I guess there are two possibilities: just keep the pdf itself in your knowledge base, or just write ALL of the relevant information in a note in your knowledge base. ### Assistant Found it: `Complete Guide to Building Skill for Claude.pdf` (downloaded today, 562KB). Let me check its size and read it: ### Assistant 33 pages. Let me extract and read it: ### Assistant Got the full content. This is comprehensive. My recommendation: **Option 2: Extract into a knowledge note** — this is better because: 1. Knowledge base is `.md` files, not PDFs 2. I can search and reference it more easily 3. The PDF is 33 pages but the actual useful content is quite dense Let me create the knowledge note and move the PDF to Attachments as backup: ### Assistant Now let me update the knowledge index: ### Assistant Done: 1. **Moved PDF** → `Attachments/Complete Guide to Building Skill for Claude.pdf` 2. **Created knowledge note** → `~/.claude/knowledge/skill-writing-guide.md` 3. **Updated INDEX** → Added `[[skill-writing-guide]]` entry The knowledge note contains: - File structure requirements - YAML frontmatter rules (critical — exact naming, kebab-case, description format) - How to write good vs bad descriptions - The three skill categories - Testing approach - Troubleshooting reference - Workflow patterns - Quick checklist Now — having read the Anthropic guide, I notice my gemini-deep-research skill draft could be improved. The guide emphasizes: 1. **Description field is crucial** — mine is decent but could include more trigger phrases 2. **Keep SKILL.md under 5,000 words** — mine is fine 3. **The "WHAT + WHEN" formula** — should be explicit Want me to revise the draft with these insights, or are you happy with it as-is? *La guida ufficiale conferma la struttura che avevo proposto, ma suggerisce di arricchire le trigger phrases nella descrizione.* ### User once you have done that, double check that your skill that you just wrote reaches the standards mentioned in the pdf ### Assistant Let me revise the skill against the Anthropic guide standards: --- ## Revised Draft (checked against Anthropic guide) ```markdown --- name: gemini-deep-research description: "Write prompts for Google Gemini Deep Research that guard against hallucination, fabricated citations, and invented statistics. Use when user asks for a 'deep research prompt', 'Gemini query', 'research this topic via Gemini', or says 'help me prompt Gemini' or 'write a prompt for deep research'." --- # Gemini Deep Research Prompt Writing Generate complete, copy-ready prompts for Google Gemini Deep Research that guard against known failure modes. ## Instructions ### Step 1: Ask About Vault Context Before generating, ask: > **Do you want me to check your vault for existing work on this topic?** > > If yes, I'll look for session notes, existing research notes, and sources you've already identified. **If yes:** Search `Sessions/*.md` for matching project. Extract active threads, current questions, known sources, academic references folder. Incorporate into prompt. **If no:** Generate based only on what user has told you. ### Step 2: Clarify Research Question (If Needed) If the research question isn't clear, ask: - What's the core thesis or question? - What's the simplest defensible framing? - Any specific positions you're responding to? ### Step 3: Generate Complete Prompt Use the structure below. Output the entire prompt in a single markdown code block for easy copying. ## Known Failure Modes Gemini Deep Research outputs frequently exhibit these problems. The prompt must guard against all of them: 1. **Phantom precision** — specific numbers (45.8%, 21.0%) that sound authoritative but are fabricated 2. **Category confusion** — conflating distinct categories (e.g., "tool-assisted" vs "abductive subset" scores) 3. **Paper misrepresentation** — claiming a paper contains arguments it doesn't 4. **Inflated claims** — "near-perfect" when actual numbers are 83% and 49% 5. **Empty citations** — reference headings with no actual sources 6. **Narrative smoothing** — turning "weakly related" into "negatively related" to fit the story ## Prompt Structure Every generated prompt MUST include these sections: ### 1. Research Question Clear, focused question. Prefer the simplest defensible framing. ### 2. Core Tension (if applicable) Frame the intellectual stakes — what positions are in conflict? ### 3. Specific Topics to Cover Numbered list of areas to research. If vault context was gathered, incorporate active threads and current questions from session note. ### 4. What I Am NOT Looking For Prevents scope creep and padding: ``` ## What I Am NOT Looking For - Marketing claims about AI capabilities - General transformer architecture explanations - Predictions about future AI - [Topic-specific exclusions] ``` ### 5. Citation and Verification Requirements **CRITICAL: Always include this block verbatim in every prompt:** ``` ## Citation and Verification Requirements CRITICAL: I have previously received research reports with fabricated citations, incorrect attributions, and invented statistics. For EVERY claim: 1. Quote directly from sources for key claims 2. Full bibliographic detail: Author(s), Title, Venue, Year, DOI/URL 3. If you cannot find a source, say "I could not locate this" rather than inventing a citation 4. Distinguish clearly between: - What a source explicitly states (quote it) - Your interpretation (flag as interpretation) - Your speculation (flag as speculation) 5. Do not invent or confabulate: - Benchmark scores - Paper titles or authors - Quotes you haven't accessed - Page numbers or DOIs 6. If knowledge is secondhand, say so: "According to Smith (2022), Jones argued..." 7. Prefer established sources over recent preprints ``` ### 6. Key Starting Points (Not Exhaustive) ``` ## Key Starting Points (Not Exhaustive) These are examples of the kind of sources I'm interested in — use them as anchors, not as a closed bibliography. I expect you to find additional relevant literature beyond these: [List 3-5 foundational sources — pull from session note if available] I am particularly interested in sources I may not know about — recent work on [topic], responses to [thesis], empirical work on [area]. ``` ### 7. Format Specify structure and length (typically: structured report, 3000-5000 words). ## Examples ### Example 1: Philosophy Research User says: "I need a deep research prompt about whether LLMs can do genuine reasoning" Actions: 1. Ask if user wants vault context checked 2. If yes, find relevant session note (e.g., Generating Philosophy) 3. Generate prompt with: - Research question about LLM reasoning - Core tension (stochastic parrots vs emergent reasoning) - Topics covering abduction, creativity, stochastic processes - Full verification requirements block - Starting points from session note's academic references 4. Output complete prompt in code block Result: Complete, copy-ready prompt with anti-hallucination guards ### Example 2: Quick Generic Prompt User says: "Write me a Gemini prompt about social cognition and AI" Actions: 1. Ask if user wants vault context 2. User says no 3. Generate prompt with standard structure 4. Include verification requirements block 5. Suggest starting points based on general knowledge Result: Complete prompt ready to paste into Gemini ## Output Requirements CRITICAL: - **Always output a complete, ready-to-copy prompt** in a markdown code block - **Never output fragments** — user must be able to copy and paste directly - User's explicit instruction: "please always give me the whole thing for ease of prompting" ## Troubleshooting ### Prompt is too long Gemini may have input limits. If prompt exceeds ~2000 words: - Reduce number of specific topics - Consolidate starting points - Keep verification requirements block intact (non-negotiable) ### User wants to modify an existing prompt Read the existing prompt first, then apply the same structure while preserving user's specific framing. ### Gemini still produces bad output The verification requirements reduce but don't eliminate hallucination. User should still fact-check citations. Suggest adding more explicit negative examples to the prompt if specific failure modes recur. ``` --- ## Standards Check (from Anthropic guide) | Requirement | Status | |------------|--------| | Folder named in kebab-case (`gemini-deep-research`) | ✓ | | SKILL.md exact spelling | ✓ | | YAML frontmatter has `---` delimiters | ✓ | | name field: kebab-case, no spaces/capitals | ✓ | | description includes WHAT and WHEN | ✓ (added more trigger phrases) | | Under 1024 characters for description | ✓ (currently ~340 chars) | | No XML tags () | ✓ | | Instructions are clear and actionable | ✓ (step-by-step) | | Error handling / troubleshooting included | ✓ (added section) | | Examples provided | ✓ (added two examples) | | Under 5,000 words | ✓ (~800 words) | **Changes from previous draft:** 1. Expanded description with more trigger phrases 2. Added Examples section with two concrete scenarios 3. Added Troubleshooting section 4. Made instructions more step-by-step 5. Added explicit "CRITICAL" markers for non-negotiable elements Ready to create the skill file? *La revisione aggiunge esempi concreti e troubleshooting come richiesto dalla guida Anthropic.* ### User yep. ### Assistant Done. Skill created at `~/.claude/skills/gemini-deep-research/SKILL.md` Now — there's also the existing knowledge file `~/.claude/knowledge/gemini-deep-research.md` that was created via `/evolve` earlier. Should I: 1. **Keep both** — knowledge file as reference, skill as workflow 2. **Delete the knowledge file** — skill now supersedes it 3. **Merge any missing content** — check if knowledge file has anything the skill doesn't What's your preference? *Lo skill è pronto — ora bisogna decidere cosa fare con il file di conoscenza preesistente.* ### User "Delete the knowledge file — skill now supersedes it" ### Assistant Done: - Deleted `~/.claude/knowledge/gemini-deep-research.md` - Removed entry from INDEX.md The skill at `~/.claude/skills/gemini-deep-research/SKILL.md` now handles everything. *File di conoscenza eliminato — lo skill ora gestisce tutto.* ### User ok, i would now like to get back to where all of that information about writing my skills has been stored. i can't remember what you said you would do ### Assistant Here's what I did with the Anthropic skill-writing guide: 1. **PDF backup** → `Attachments/Complete Guide to Building Skill for Claude.pdf` 2. **Knowledge note** → `~/.claude/knowledge/skill-writing-guide.md` — contains all the extracted content (file structure, frontmatter rules, description formula, testing approach, troubleshooting, etc.) 3. **INDEX updated** → Added `[[skill-writing-guide]]` entry under Claude Code section So the actionable reference is in my knowledge base at `~/.claude/knowledge/skill-writing-guide.md`, and the original PDF is preserved in your vault's Attachments folder. *La guida è nel knowledge base — il PDF originale è negli allegati come backup.* ### User Now, would it be good to create a skill making skill using this information. genuine qurstion. Search the web for your answer. ### Assistant **Answer: Anthropic already made one.** There's an official **skill-creator** skill in [Anthropic's skills repository](https://github.com/anthropics/skills/blob/main/skills/skill-creator/SKILL.md). From the guide I extracted: > "The skill-creator skill - available in Claude.ai via plugin directory or download for Claude Code - can help you build and iterate on skills. If you have an MCP server and know your top 2–3 workflows, you can build and test a functional skill in a single sitting - often in 15–30 minutes." **So the answer is: don't create your own — just install the official one.** Want me to fetch and install the official skill-creator skill from the Anthropic repo? --- **Sources:** - [Anthropic Skills GitHub Repository](https://github.com/anthropics/skills) - [skill-creator SKILL.md](https://github.com/anthropics/skills/blob/main/skills/skill-creator/SKILL.md) - [Introducing Agent Skills | Anthropic](https://www.anthropic.com/news/skills) *Anthropic ha già creato uno skill-creator ufficiale — meglio installare quello che reinventare la ruota.* ### User "Want me to fetch and install the official skill-creator skill from the Anthropic repo?" yep. and can you also give me a complete guide to any and all official skills that anthropic have made. ### Assistant Now let me install the skill-creator and fetch details on the key skills: ### Assistant Skill-creator installed. Now let me create the comprehensive guide: ### Assistant Done. Here's what I installed and documented: --- ## Installed: skill-creator **Location:** `~/.claude/skills/skill-creator/` **Contents:** - `SKILL.md` — Full guide (17KB) - `scripts/init_skill.py` — Initialize new skill - `scripts/package_skill.py` — Package for distribution - `scripts/quick_validate.py` — Validate structure - `references/output-patterns.md` — Output patterns - `references/workflows.md` — Workflow patterns --- ## Created: Official Skills Guide **Location:** `~/.claude/knowledge/anthropic-official-skills.md` **All 16 Official Anthropic Skills:** | Category | Skills | |----------|--------| | **Documents** | docx, pdf, pptx, xlsx | | **Design** | frontend-design, canvas-design, algorithmic-art, theme-factory | | **Development** | mcp-builder, webapp-testing, web-artifacts-builder | | **Communication** | brand-guidelines, internal-comms, doc-coauthoring | | **Utility** | skill-creator, slack-gif-creator | The guide includes install commands for each skill if you want to add more later. --- **Sources:** - [Anthropic Skills Repository](https://github.com/anthropics/skills) - [skill-creator SKILL.md](https://github.com/anthropics/skills/blob/main/skills/skill-creator/SKILL.md) *Installato skill-creator con tutti i suoi script, e creata una guida completa a tutti i 16 skill ufficiali di Anthropic.*