Fable 5 Debacle: Set Up Local AI Prompt
local model setup plan
Chapter IV
Cards marked with a teaser are in the Vault. Unlock everything — $97, one time.
local model setup plan
Focus the two-host conversation on your tension/question, set tone, skip what you know
Pull the author's essays/interviews/talks, the theme from a different angle, and critiques to enrich the notebook
Frame the Video Overview around the tension between the author's argument and the strongest counterargument
Given your profession, surface the non-obvious link between the sources and how you work
Build a second map around the idea from your paper notes that stopped you most
Generate connection-first and contrast (author vs counterargument) infographics in chosen styles
Generate quiz questions on where sources agree/contradict, then use Explain on every question
Create-Your-Own report on the most counterintuitive finding across the sources, surprising-thing-first
Have NotebookLM generate one adjacent deep-research prompt, then launch it
CMO/VP attribution-dispute scenario to run at two effort levels side by side to feel where the harder layers appear
Plan-before-you-build AI-consulting-site prompt to run at two effort levels and read side by side
Build prompt for a browser-only Substack post analyzer with word-lift and title-neighbor planner
Build prompt for a from-scratch subscriber forecast with prediction bands
Build prompt for churn, tenure, LTV, and MRR forecasting in the browser
Four-step prompt that reads all three generated reports together
Reads one markdown business report and produces decisions
what stays local vs goes to cloud
pick a local model + platform
execute the first fine-tune
I want to complete my first fine-tune of a small language model this week using Unsloth. I have decided on a base model, a training method, and the type of work I want the fine-tuned model to specialize in. I have never done this before and I need a clear run plan so the week does not get lost in debugging. My specific setup: - Base model: [From Stage 2 decision] - Training method: [From Stage 3 decision, LoRA or QLoRA] - Training environment: [Free Colab, Kaggle, or local GPU with VRAM amount] - Dataset size: [Number of input-output pairs ready to use] - The task the fine-tune should learn: [From Stage 3 outcome] My skill level: [Describe accurately. Examples: "I can copy code into a notebook and run it but I cannot debug errors," or "I write Python but not for ML," or "I have used Colab before but never for training."] The two training days I have blocked for this: [Name the two consecutive days, for example "Saturday and Sunday" or "Tuesday and Wednesday." The plan will be built around these two days.] TASK Produce a two-day run plan that includes: 1. Day 1 morning: the exact Unsloth notebook to open, with a direct link, and the three cells I need to edit before running 2. Day 1 afternoon: how to monitor the training run and what normal versus abnormal behavior looks like 3. Day 1 evening: the export step to GGUF format and how to verify the export succeeded 4. Day 2 morning: how to load the fine-tuned model into Ollama or LM Studio on my local machine 5. Day 2 afternoon: five evaluation prompts I should run against the fine-tune to assess whether it learned what I trained it on 6. The three most common first-run failure modes and how to recognize which one I hit if the run fails CONSTRAINTS Give me specific notebook links, specific cell numbers where possible, and specific commands. Do not tell me to read the documentation. Tell me exactly what to do in what order. Assume the two days I named are consecutive and I can only commit three to five hours per day. Stage 5: Build a dataset from your own work The quality of your fine-tune depends entirely on the quality of your dataset. This stage is the easiest to want to skip, and it is the one that separates a fine-tune that delivers real value from a fine-tune that is indistinguishable from the base model. Start with the work you already have. Your past client briefs, your published writing, your email threads, your internal documentation, your strategic frameworks, your product specifications, your coaching transcripts. Anything that represents the pattern you want the model to learn belongs in the dataset. The format is a list of input-and-output pairs. For a writing voice fine-tune, the input is a prompt and the output is something you actually wrote in response to a similar prompt. For a customer-service fine-tune, the input is a customer question and the output is your ideal response. For a strategic-analysis fine-tune, the input is a business situation and the output is how you actually analyze it. Three hundred high-quality pairs is enough to see meaningful results on a small instruct model. A thousand pairs is enough to produce a fine-tune that feels like a different model. The dataset is the asset, and it compounds every time you add to it. Prompt for Stage 5: The Dataset Extraction and Formatting Plan Run this prompt against Claude with access to your own writing, client work, or business documentation. The output is a specific extraction plan plus formatted training pairs you can paste directly into your fine-tuning notebook. CONTEXT
plan a fine-tune
pick an SLM for your tasks
build a fine-tune dataset
install Ollama/LM Studio/Jan
private founder-decision prompt
spec a sellable digital product
draft sensitive comms privately
think through something private
private strategic-advisor framing prompt
weekly content strategy session
build a webinar/slide deck
find competitive blind spots
generate a first draft from sources
mine sources for content gold
generate a podcast episode
generate an infographic
generate video overviews
track content performance
turn meeting notes into actions, locally
draft client proposals locally
generate content ideas locally
rewrite text for clarity, locally
summarize messy notes, locally
organize a brain dump (tasks/ideas/ignore)
classify tasks, locally
test a local model's judgment
create a meeting brief, locally
prioritize your week, locally
extract structured data, locally
test a local model for hallucination
strengthen an offer, locally
audit a vendor's business model/ethics/data
decode a privacy policy
write an org AI use policy
monitor your AI tools quarterly
compare AI tools on trust
extract decision-relevant intelligence from docs
Prompt 1: The Document Intelligence Extractor Pattern: Structured Decomposition Designed for the expanded 1M token context window. Use this after uploading a large document set (reports, contracts, research papers) that would have required multiple sessions with Opus 4.5. <role> You are a senior research analyst specializing in synthesizing large document sets into executive-ready intelligence. </role> <context> I've uploaded [NUMBER] documents totaling approximately [ESTIMATE] pages. Topic area: [YOUR DOMAIN, e.g., "Q4 financial reports," "vendor contracts," "industry research"] My role: [YOUR TITLE/FUNCTION] Decision I need to make: [SPECIFIC DECISION OR QUESTION] </context> <task> Analyze all uploaded documents as a complete body of evidence. Do not summarize each document independently. Instead: 1. Identify the 3-5 most significant findings that bear on my decision 2. Flag any contradictions between documents 3. Surface any data points that appear in only one source (potential blind spots or unique insights) 4. Note what's missing: what would I need to see to make a fully informed decision? </task> <output_format> ## Key Findings (ranked by relevance to decision) [Finding with specific references to source documents] ## Contradictions Detected [Where sources disagree, with page/section references] ## Unique Data Points [Information appearing in only one source] ## Information Gaps [What's not covered that should be] ## Recommended Next Step [Single most important action based on analysis] </output_format> <constraints> - Reference specific documents and sections, not vague summaries - Prioritize findings by relevance to my stated decision, not by document order - If a finding requires domain-specific judgment I should make myself, say so </constraints>
plan → research → execute autonomously
Prompt 2: The Autonomous Executor Pattern: Constraint-Based Design This prompt is built for Opus 4.6’s agentic execution behavior. You define the deliverable and the guardrails. Claude handles every step in between without stopping to ask permission at each phase. On Opus 4.5, this type of prompt would typically stall at decision points. On 4.6, it runs. <role> You are a senior project executor. You make decisions, take action, and deliver complete work. You do not stop to ask for approval on individual steps unless you encounter a decision that could cause irreversible harm or significant cost. </role> <context> Deliverable: [WHAT YOU NEED PRODUCED, e.g., "a competitive analysis comparing our three top vendors," "a 90-day content calendar with topics and angles," "a complete project post-mortem with root causes and recommendations"] Audience: [WHO WILL READ/USE THIS] My role: [YOUR TITLE AND WHAT YOU CARE ABOUT MOST] Inputs: [DOCUMENTS UPLOADED OR CONTEXT PROVIDED] Hard constraints: [NON-NEGOTIABLES, e.g., "must include cost comparison," "must be under 10 pages," "must address regulatory requirements"] </context> <task> Produce the complete deliverable from start to finish in this conversation. 1. Plan your approach (share the plan, then proceed immediately) 2. Conduct any necessary research using web search 3. Analyze all provided inputs 4. Draft the complete deliverable 5. Review your own work: check for gaps, weak arguments, missing data, and logical errors 6. Present the final version with a brief note on any areas where my judgment should override yours Do NOT stop after each step to ask if you should continue. Execute the full sequence. If you encounter ambiguity, make the best judgment call available and flag it in your final notes. </task> <output_format> [Complete deliverable in the format most appropriate for the audience] ## Execution Notes - Judgment calls made: [Where you chose between options and why] - Confidence flags: [Areas where your analysis may be incomplete] - Recommended human review: [Specific sections where my expertise matters most] </output_format> <constraints> - Complete execution: do not stop to ask permission between phases - Flag judgment calls transparently rather than hiding uncertainty - If you cannot complete a section due to missing information, note what's needed and complete everything else </constraints>
delegate tasks via Cowork
Prompt 4: The Cowork Task Delegator Pattern: Constraint-Based Design Built for Cowork, where Opus 4.6 can multitask autonomously across documents, spreadsheets, research, and presentations. Use this when you have a multi-format deliverable and want Claude to handle the entire production pipeline. <context> I need to produce the following set of deliverables for [AUDIENCE/PURPOSE]: Deliverable 1: [e.g., "A research brief on competitive pricing in our market"] Deliverable 2: [e.g., "A spreadsheet comparing competitor pricing across 5 dimensions"] Deliverable 3: [e.g., "A 10-slide presentation summarizing findings with recommendations"] Inputs available: [DOCUMENTS, DATA, URLS, OR CONTEXT YOU CAN PROVIDE] Timeline: [WHEN THIS IS NEEDED] Brand/formatting requirements: [ANY SPECIFIC STANDARDS] </context> <task> Execute this as a complete production pipeline: 1. Start with research: gather any external data needed using web search and analyze all provided inputs 2. Produce each deliverable in sequence, with each building on the findings of the previous one 3. Ensure consistency across all deliverables: same data, same conclusions, same recommendations 4. Review the full set for internal contradictions or gaps before presenting For each deliverable, proceed to completion without stopping for approval. Flag any judgment calls in a final summary. </task> <output_format> [Complete deliverables in appropriate formats] ## Production Summary - Cross-deliverable consistency check: [Confirmed or flagged issues] - Judgment calls: [Decisions made and rationale] - Areas for human review: [Where your expertise matters most] - Suggested refinements: [What would make each deliverable stronger with more time or input] </output_format> <constraints> - All deliverables must tell the same story with consistent data - Presentations should lead with findings, not background - Spreadsheets should be structured for the audience's decision, not for raw data display - If any deliverable requires information you don't have, note the gap and complete everything else </constraints>
a long deep-research session
architect a long-running project
build a true-cost-of-purchase calculator
build a freelance-rate calculator
build a commute-cost calculator
build a meeting-cost calculator
build a true-hourly-rate calculator
scan the morning landscape
hunt for opportunities
deep-dive a topic in NotebookLM
produce an executive brief
track competitor moves
spot emerging risks
verify event timelines
generate 5 ready-to-build GPT templates
roll out Custom GPTs across an org
GPT that polishes casual notes into client emails
build a meeting-notes-to-summary GPT
research to support a decision
scan a market with Gemini Deep Research
deep competitor analysis
synthesize technical literature
adoption stage + drivers + timeline
full integrated-workflow design template
one connected research→analysis→visualization→build project
research→analysis→viz→synthesis report
integrated market analysis with visuals + sources
build a complete interactive dashboard
multi-modal learning experience from sources
drop-in Instructions-box config with voice rules + non-disclosure guardrail
complete worked GPT config (voice, tasks, formatting, restrictions)
automate workflows
manage tasks with AI
synthesize your knowledge base
make avatar presentations
set up a strategy copilot
produce visual content
produce video
create voice content
extract intel from meetings
comparative brief on 3 companies' strategy
top recurring themes from last 3 months + 12-mo impact
strongest counterarguments to a consensus belief
10-year milestone timeline for a topic
3 named 5-year scenarios + leading indicators
turn raw JSON into human output
Using my uploaded lecture notes and textbook chapters, create a Study Guide that explains
Based on these course materials, generate a comprehensive set of practice questions covering [topic],
After analyzing these literary texts, create a Mind Map showing the relationships between characters,
Using these historical documents, create a Timeline of key events related to [historical period/event],
Help me understand the methodology, findings, and limitations across these research papers on [scientific
Based on these background materials about [company/project], create a Briefing Document highlighting key points
Using my uploaded market research and internal documents, generate a presentation outline on [topic]
After analyzing these industry reports in the Sources panel , Create a FAQ document
After reviewing these market research documents and audience analytics, generate 10 unique content ideas
Based on these case studies of similar challenges, create a Mind Map of potential
Using these client requirements, brand guidelines, and market research, create a comprehensive Study Guide
I've uploaded several introductory articles about [topic I'm learning]. Create an Audio Overview that
I've uploaded chapters from [book]. Generate a Study Guide of key insights, practical applications,
Based on these resources about [skill], create a structured Study Guide that breaks down
After analyzing these articles about [health condition/topic], create a comprehensive FAQ that addresses common
Using these travel guides and articles about [destination] I've uploaded, create a Briefing Document
Help me compare and contrast the perspectives on [controversial topic] presented in these different
Based on these language learning materials in my Sources, create a Study Guide for
After reviewing these documents outlining different options for [decision], create a Briefing Document with
After analyzing these materials on [topic] in the Sources, identify areas where information is
Create an Audio Overview that presents the evolution of [topic] as an engaging historical
Generate an Audio Overview where the hosts present contrasting perspectives on [controversial topic] based
Create an Audio Overview that uses the real-world examples from my sources as case
Generate an Audio Overview that presents [complex topic] as a mystery being solved, gradually
Create an Audio Overview that breaks down [complex concept] into its fundamental components, using
Generate an Audio Overview comparing and contrasting [multiple approaches/theories] discussed in my sources, highlighting
Create an Audio Overview that transforms the theoretical knowledge in my sources into practical,
Generate an Audio Overview that explores how [topic] connects to other fields mentioned in
Create an Audio Overview that explores potential future developments in [field/topic] based on the
Generate an Audio Overview that uses thought experiments to explore the implications of [concept/theory]
Create an Audio Overview structured as an interview with a leading expert on [topic],
Generate an Audio Overview that explains [topic] at three different levels of complexity—beginner, intermediate,
example steering prompts (Explainer/Brief/Cinematic + lazy-vs-precise)
chain outputs into finished assets (recipes)
The prompts below are organized by who’s using them rather than by output type.
Simplify this technical documentation into a user-friendly Study Guide with step-by-step instructions. Include a
Based on these client requirements and our previous work, use the Studio panel to
Using these character profiles and plot outlines in the Sources panel, create a Mind
After analyzing these market trends and customer feedback documents, create a Briefing Document suggesting
Explain the key findings in simple, non-technical language. Create analogies that would help a
From my notes and textbook chapters, explain [concept] in plain language with examples and
From these materials on [project], the key points before my meeting, plus likely questions
Distill these reports into a readable post on trends in [field], citing data and
A decision matrix for [decision] with pros, cons, and citations.
Where are my sources on [topic] incomplete or contradictory? List questions to research further.
Write a Blog Post arguing [position]: a hook, 3 supporting sections, and a closing
Build a Study Guide for [topic]: key terms, 10 short-answer Q&As, and 3 essay
Walk a newcomer through the fundamentals of [topic] and how the ideas connect.
The three things a decision-maker needs to know about [topic], with the “so what”
The evolution of [topic] with turning points and key figures.
Compare [options] with strengths, weaknesses, and when to use each.
Explainer format. Walk through [topic] in 5 parts, ending each with the one thing
Cinematic format. Frame [topic] as a journey with a hook, a turning point, and
Explainer format. Audience is [role]. Translate the jargon into plain language, one analogy per
Turn my research into a slide outline for [topic] with talking points and evidence
A lecture deck on [concept], basics to advanced, one idea per slide with examples.
focus the deck on [audience] and limit it to [N] slides with one idea
Key terms and definitions for [topic].
Cause-and-effect and the “why” behind [topic], not just definitions.
Mix recall and application questions on [topic].
Make [N] flashcards covering the must-know terms for [topic], front = term, back =
Build a [N]-question quiz on [topic] with an answer key and a one-line explanation
Turn the key stats and relationships on [topic] into one visual.
Visualize the step-by-step process for [workflow].
Compare [options] across the dimensions in my sources, one row each.
Every claim about [topic] in a table: claim, source, supporting data.
map the [process] as a flowchart with each step and its key decision point.
compare [options] with columns for cost, effort, risk, and best-use-case, one row each.
Hosts present contrasting perspectives on [topic] from my sources.
Explain [topic] at beginner, intermediate, and advanced levels.
Show how the concepts and themes in [topic] connect, with quotes.
a 10-minute deep dive on [topic] where the hosts disagree, then find common ground.
a beginner-friendly walkthrough of [topic] with the hosts asking the questions I'd ask.
a rapid-fire '5 things you didn't know about [topic]' episode, each backed by a
branch [topic] into its main themes, then sub-branch each into key examples from my
Compress these sources into a one-page executive Briefing Doc: 5 takeaways, 3 risks, 3
a literature-review report grouping sources by theme, noting agreement, disagreement, and gaps.
Brief format. The 90-second cut: headline insight and why it matters, no background.
A stakeholder update from this project’s docs: progress, risks, next steps.
make slide 1 a problem statement and the final slide a clear call to
add a speaker-notes paragraph under each slide with what to say out loud.
turn my data into 3 chart-style slides with a one-line takeaway headline on each.
Target the concepts learners most commonly get wrong.
front = a scenario, back = which concept from my sources applies and why.
Build a quiz that escalates from easy recall to hard application, flagging the trickiest
turn the 5 most important numbers in my sources into a single visual with
extract every recommendation in my sources with columns for the recommendation, source, and rationale.
Structured as an interview with an expert; answers drawn from my sources.