custom-gpt

Build a complete Custom GPT package from $ARGUMENTS (transcripts, raw text, or file paths). ## Overview This skill produces everything needed to set up an OpenAI Custom GPT: - `system-prompt.md` — Instructor personality, behavior rules, topic scope, response format - `kb-*.md` — 2–4 knowledge base files, logically grouped, concise but complete - `README.md` — Setup steps, sample test questions, video/topic coverage map **Output location:** `~/content/custom-gpts/<project>/<gpt-slug>/` --- ## Step 1: Gather Input If transcripts or text were passed as `$ARGUMENTS` or pasted into the message, use them directly. If a file path was provided, read the file(s). If no content was provided yet, ask: > "Please paste your transcripts or point me to a file path. You can paste multiple videos — just separate them with a header like `Video Title: [name]`." --- ## Step 2: Ask Clarifying Questions Before generating anything, ask these questions **all at once** in a single message (don't drag it out): 1. **GPT name** — What should this GPT be called? (e.g., "Sales OS Assistant") 2. **Instructor** — Who teaches this content? Name, role, credentials. 3. **Module context** — Is this part of a larger course/platform? What are the other modules? (Helps define what's out of scope.) 4. **Voice profile** — Is there already a voice profile file for this instructor? If so, where? (If yes, it gets uploaded as a separate Knowledge file — no need to bake it into system-prompt.) 5. **Conciseness level** — How long is the content? Should KB files be detailed or highly condensed? (Default: condensed — extract frameworks and scripts, not transcripts.) If any of these are obvious from the content or context, skip asking and state your assumption. --- ## Step 3: Analyze the Content Read through all transcripts/text and extract: ### Instructor Voice - Tone and energy (casual, formal, fired up, calm) - Signature phrases and filler words - Teaching method (example-first, framework-first, story-based) - What they explicitly say NOT to do (common mistakes they call out) - Credentials and real numbers they reference ### Topics & Scope - Full list of topics covered, with the video/section each comes from - What is explicitly outside scope (mentioned as belonging to another module) - Key frameworks, named systems, and processes - Word-for-word scripts, templates, or formulas worth preserving verbatim ### KB Groupings Identify 2–4 logical groupings for the knowledge base files. Good grouping criteria: - By phase (e.g., strategy → execution → optimization) - By topic cluster (e.g., ads + tracking vs. organic content) - By video block (e.g., Videos 1–7 vs. Videos 8–14) Aim for each KB file to be self-contained and focused. --- ## Step 4: Generate the Files Create the output directory: `~/content/custom-gpts/<project>/<gpt-slug>/` Where: - `<project>` = the platform/course name in kebab-case (e.g., `kourse`) - `<gpt-slug>` = the GPT name in kebab-case (e.g., `sales-os-gpt`) If no project name is clear from context, use `custom-gpts` as the project folder. --- ### `system-prompt.md` Structure: ``` # [GPT Name] — System Prompt [1-2 sentence description of what this GPT is and who it serves] ## Your Instructor / Personality [Instructor name, role, credentials] [Tone descriptors — 4-6 bullet points] ### How You Talk [Signature phrases, voice characteristics] ### How You DON'T Talk [What to avoid — e.g., generic advice, pressure tactics, vague theory] ## How You Answer Questions [4-5 numbered rules for answering — e.g., check KB first, be specific, give scripts not summaries] ## Topics You Can Help With [Organized by category — mirror the KB structure. For each category, list 6-12 specific sub-topics as bullets. Be specific enough that the GPT knows what's covered.] ## Topics Outside Your Scope [List other modules/sections and what topics belong there. Include a redirect template.] ## Response Format [Formatting rules — length, use of bullets/bold/numbered lists, whether to end with action steps, etc.] ``` **Rules for system-prompt:** - Instructor voice must be specific — generic "be helpful and friendly" is useless - Topics list should be detailed enough that the GPT can self-identify whether a question is in scope - Out-of-scope section prevents hallucination about topics not in the KB - Do NOT include long scripts or frameworks here — those belong in KB files - If a voice profile file exists (user confirmed in Step 2), skip voice details and add a note: "Voice and credentials are defined in the uploaded `voice-profile.md` knowledge file." --- ### `kb-*.md` (2–4 files) Name files descriptively: `kb-<topic-group>.md` Structure per file: ``` # Knowledge Base: [Topic Group Name] ## [Module/Course Name] — [Video or Section Range] --- ## [Section Title] ### [Sub-topic] [Condensed, actionable content. Preserve:] - Named frameworks with their components - Word-for-word scripts (formatted in italics or blockquotes) - Step-by-step processes (numbered) - Key data points, thresholds, or rules of thumb - Tables for multi-column information [Do NOT preserve:] - Filler, repetition, or motivational padding - Long storytelling tangents (summarize the point they illustrate) - Jokes or asides that don't add information ``` **Conciseness rules:** - Extract the framework, not the explanation of the framework - If a script is word-for-word, keep it word-for-word — that's the value - If a concept takes 3 paragraphs to explain in the transcript, distill it to 3-5 sentences - Tables > bullet lists > paragraphs (in that order of preference) - Each KB file should be self-contained — someone reading just that file should understand the topic --- ### `README.md` Structure: ``` # [GPT Name] — Setup Guide ## What's in This Folder [Table: File | Purpose | Upload To] [Include voice-profile.md row if applicable] > Note: [Any important setup notes — e.g., "Do NOT upload X" or "Y's voice is in system-prompt directly"] ## Setup Steps (OpenAI Custom GPT) 1. Go to chat.openai.com → Explore GPTs → Create 2. In the Configure tab: - Name: [GPT Name] - Description: [1-sentence description] - Instructions: Copy/paste system-prompt.md 3. Under Knowledge, upload: [list all KB files + voice-profile if applicable] 4. Under Capabilities: - Disable Web Browsing - Code Interpreter: not needed 5. Save and test ## Sample Test Questions [10-15 questions that cover the most important topics. For each, add → [what the correct answer should reference]] ## Knowledge Base Coverage Map [Table: Video/Section | Topic | Knowledge File] ``` --- ## Step 5: Review & Output After generating all files, show the user: 1. A summary of what was created (file list + brief description of each) 2. The KB groupings chosen and why 3. The sample test questions (most useful for verifying coverage) Ask: "Want me to adjust anything — the voice, add more topics, regroup KB files, or add more sample questions?" If they're happy, confirm the output path and remind them: - Paste `system-prompt.md` into the Instructions field - Upload all `kb-*.md` files as Knowledge - Upload `voice-profile.md` as Knowledge if applicable - Disable web browsing --- ## Rules - **Condense, don't transcribe.** KB files are not transcripts. They're distilled references. - **Preserve scripts verbatim.** If the instructor gives a word-for-word script, keep it exactly. - **Be specific in the system prompt.** Generic personality descriptions produce generic GPT behavior. - **Group KB files by logical topic, not by video number** — unless the content is so large that video-number grouping is the only way to keep files manageable. - **Never put long scripts or frameworks in system-prompt.** That's what KB files are for. - **Out-of-scope = hallucination prevention.** Always include a clear out-of-scope section with redirect language. - **Sample test questions should be hard enough to actually test coverage.** Not "what is X" but "walk me through how to do X" or "someone said Y — what do I do?" - **If content is very large** (15+ videos, 50k+ words), create 3–4 KB files and be more aggressive about condensing. If content is small (3–5 short videos), one KB file may be enough. - Save all files to `~/content/custom-gpts/<project>/<gpt-slug>/` — never anywhere else.