5 lessons · 60–90 minutes
Build your first testable AI workflow
A task brief, control-loop decision, and ten-case pilot sheet.
GUIDES READ0 / 5
Continue next lesson →LEARNING PROGRESS / FOUR PRACTICAL COURSES
See your learning as clearly as a course syllabus.
Find read guides and saved artifacts by course. The sitewide percentage is reading coverage; choose Agent and local-model branches only when your task needs them.
AI VISTA / PROGRESS FILEOVERVIEW
0%READ
- 01Build your first testable AI workflow
- 02Make AI answers reliable enough to use
- 03Design answers that can show their evidence
- 04Automate with permission, recovery, and cost controls
0/ 63 GUIDES READ
COURSE PROGRESSWORKING ARTIFACTSACHIEVEMENT RECORD
LEARNING ROUTE / STAGE 06 · PROGRESS
This stage reconnects the reading log to the work you actually produced.
Reading coverage is an orientation aid that includes optional branches. The useful questions are which artifacts exist, which guide fits next, and which decision still needs testing.
- BRING
- Read guides, artifacts, and local notes
- DO HERE
- Review read guides and the next move by course
- LEAVE WITH
- A progress overview and a clear continuation point
FROM COMPLETION TO WORK EVIDENCE
A complete body of work is not a lesson count. It is four kinds of material another person can review.
Course states help you find your place. The evidence portfolio shows whether learning has reached real work. When one kind is missing, do not restart; return to the matching course and fill the earliest gap.
REVIEW RULEShow the artifacts without a verbal explanation. They count as work evidence only if the reviewer can still identify the decision, support, and failure handling.
- 01TASK
Which exact action belongs to AI?
MATERIALFit matrix, task brief, control-loop decision, and data boundary
REVIEW↗Can a colleague identify the input, acceptable output, next-step controller, prohibited action, and handoff owner?
- 02ACCEPTANCE
What result passes, needs revision, or must be rejected?
MATERIALContext map, acceptance contract, injection test set, evaluation set, and diagnostic tree
REVIEW↗When the same cases run again, can failures be classified consistently and traced to a change?
- 03EVIDENCE
Why should a consequential claim be believed?
MATERIALRAG decision gate, retrieval trial, and citation-freshness audit
REVIEW↗Did retrieval find an eligible current passage, and does the citation support the claim?
- 04RECOVERY
How does failure stop and return to a safe state?
MATERIALPermission ladder, tool contract, human gates, runbook, cost map, and drift watchboard
REVIEW↗Can the team rehearse unauthorized, timed-out, duplicated, or rejected actions and choose continue or rollback?
COMPLETION IS NOT MASTERY
Use four questions to see whether learning has reached the real work.
A completed state only says you visited the lesson. Real completion needs an artifact, a reviewable standard, current evidence, and a recovery path. A missing answer does not reset progress; it identifies the next place to learn.
HOW TO USEAfter each course, ask someone who did not build the workflow to challenge each row. Record what you cannot answer, then return to the linked topic.
- 01 DOES THE ARTIFACT EXIST?Return to the course and make it →
Can you open a completed brief, evaluation sheet, permission map, or recovery runbook—not merely recall the guide?
Pass signal: it uses task-specific field categories, synthetic case IDs, and controlled evidence locations; sensitive originals stay in approved systems.
- 02 DID ACCEPTANCE PASS?Rebuild acceptance and cases →
Can a colleague use the same standard and reach the same pass, revise, or reject judgment on the cases?
Pass signal: the standard separates acceptable, revisable, and rejected work.
- 03 IS THE EVIDENCE STILL VALID?Review source, version, and freshness →
Are the sources behind important decisions still current, correctly versioned, and permitted for this user?
Pass signal: consequential claims trace to valid sources.
- 04 CAN THE WORKFLOW RECOVER?Complete the stop and recovery path →
When a model, tool, or handoff fails, who notices, where does it stop, and how does it return to a safe state?
Pass signal: the team has rehearsed at least one failure.
FOUR PRACTICAL COURSES
Continue with the guides that fit your task.
0 / 63 guides read · optional branches included
14 lessons · 2–3 hours
Make AI answers reliable enough to use
A context map, compaction checkpoint, acceptance contract, output protocol check, injection test set, and failure log.
GUIDES READ0 / 14
Continue next lesson →11 lessons · 2–3 hours
Design answers that can show their evidence
A method decision, source admission register, chunking and retrieval trials, access and index checks, a no-evidence response gate, and a citation audit.
GUIDES READ0 / 11
Continue next lesson →37 lessons · Core 3–4 hours; branches vary
Automate with permission, recovery, and cost controls
A control-loop choice, data and permission boundaries, sandbox and tool contracts, runtime stop and human review gates, recovery runbook, service fallback, cost map, and drift watchboard; branch evidence where relevant.
GUIDES READ0 / 37
Choose a relevant guide →LESSONS & OUTPUTS
Review read guides and saved practice artifacts.
Should this task use AI at all?OUTPUT · decision matrix
UNREADTurn an AI idea into a testable task briefOUTPUT · one-page brief
UNREADWorkflow or AI agent? Choose the control loop before buildingOUTPUT · control-loop decision
UNREADRun a ten-case AI pilot before you scaleOUTPUT · pilot sheet
UNREADMap the context before rewriting the promptOUTPUT · context map
UNREADHow many tokens should one LLM call reserve for input and output?OUTPUT · per-call context budget
UNREADWhen should you compact a long AI conversation?OUTPUT · compaction checkpoint
UNREADWrite the acceptance contract before the promptOUTPUT · acceptance contract
UNREADChoose JSON Schema, Markdown, or plain text for AI outputOUTPUT · output protocol check
UNREADTest prompt injection in documents and tool outputsOUTPUT · injection test set
UNREADDiagnose an AI answer before changing the modelOUTPUT · diagnostic tree
UNREADPrompting, RAG, or fine-tuning? Choose by failure modeOUTPUT · method-choice trial
UNREADDecide whether RAG is worth buildingOUTPUT · investment gate
UNREADHow do you approve sources before adding them to a RAG knowledge base?OUTPUT · rag-source-register
UNREADHow do you validate PDF, table, and image extraction before RAG indexing?OUTPUT · extraction-acceptance-sheet
UNREADTest citations for support and freshnessOUTPUT · citation audit
UNREADEvaluate RAG retrieval before judging the answerOUTPUT · retrieval trial
UNREADKeyword, vector, hybrid, or reranking for RAG? Choose with a matched trialOUTPUT · retrieval strategy comparison
UNREADHow should you chunk RAG documents? Keep the full evidence intactOUTPUT · document chunking trial
UNREADDocument-level access control for enterprise RAG, from search to citationOUTPUT · document-access trace
UNREADHow do you verify that a RAG document update or withdrawal reached the index?OUTPUT · RAG index refresh acceptance trial
UNREADWhen RAG lacks evidence, should it answer, clarify, or hand off?OUTPUT · rag-no-evidence-response-gate
UNREADGive an agent the smallest useful permissionOUTPUT · permission ladder
UNREADHow do you test an AI agent sandbox's file, network, and credential boundaries?OUTPUT · sandbox boundary trial
UNREADDesign an AI agent tool contract before exposing the functionOUTPUT · tool-calling contract
UNREADHow to implement an AI agent tool-calling loopOUTPUT · agent loop execution contract
UNREADAI inference latency debugging: trace the wait from request to screenOUTPUT · inference latency trace
UNREADWhen should an AI agent continue, retry, or stop?OUTPUT · runtime loop guard
UNREADSubagent or multi-agent? Choose the delegation topologyOUTPUT · agent delegation topology
UNREADWhen should an AI agent save, correct, or delete long-term memory?OUTPUT · memory lifecycle contract
UNREADHow to debug one AI agent run from its observable traceOUTPUT · single-run trace review
UNREADHow can a long-running AI agent task resume without repeating work?OUTPUT · long-task resume-state contract
UNREADWhat should you check before connecting a third-party MCP server?OUTPUT · MCP server intake record
UNREADWhen should a repeated AI task become an agent skill?OUTPUT · agent skill decision card
UNREADHow do you load and unload an agent plugin safely?OUTPUT · plugin release acceptance record
UNREADDesign recovery before the agent failsOUTPUT · recovery runbook
UNREADChoose a model with a job-specific scorecardOUTPUT · selection scorecard
UNREADHow should an AI model service fail over? Plan retries, queues, and human handoffOUTPUT · model service fallback ladder
UNREADHow much VRAM does a local LLM need? Budget weights and KV cacheOUTPUT · VRAM budget and peak test
UNREADIs 4-bit LLM quantization worth it? Test fit, quality, and speedOUTPUT · BF16 versus 4-bit trial
UNREADLocal LLM or cloud API? Choose by data, peak load, and operationsOUTPUT · inference deployment matrix
UNREADHow many concurrent requests can a local LLM serve?OUTPUT · inference serving load trial
UNREADOllama, llama.cpp, vLLM, or SGLang? Choose a local inference engineOUTPUT · inference engine selection trial
UNREADLoRA or QLoRA fine-tuning? Run a small, reviewable pilotOUTPUT · LoRA/QLoRA pilot record
UNREADHow many GPUs does a local LLM service need?OUTPUT · GPU replica capacity worksheet
UNREADMap the full cost of an AI workflowOUTPUT · cost map
UNREADWhy prompt caching misses: run a three-request testOUTPUT · cache trial
UNREADDraw the data boundary before AI sees the inputOUTPUT · data boundary map
UNREADBuild an AI evaluation set that stays usefulOUTPUT · evaluation-set blueprint
UNREADHow do you calibrate an AI grader against human review?OUTPUT · grader calibration pack
UNREADHow to set AI classification thresholds with a human review queueOUTPUT · classification routing sheet
UNREADHow to test LLM temperature and sampling settingsOUTPUT · sampling settings trial
UNREADPut human review where it can change the outcomeOUTPUT · review gate map
UNREADMonitor an AI workflow before drift becomes an incidentOUTPUT · drift watchboard
UNREADAI support triage: the 10-case pilot that earned a review gateOUTPUT · pilot sheet
UNREADAI content review: the citation gate that stopped a confident wrong answerOUTPUT · citation audit
UNREADAI research synthesis: when the source changes after the briefOUTPUT · retrieval trial
UNREADThe approval agent that could move money—until we removed the payment toolOUTPUT · permission ladder
UNREADYour MCP server works locally. What breaks when a team depends on it?OUTPUT · MCP server intake record
UNREADThe eval that would have caught this: turn one bad output into a regression testOUTPUT · evaluation-set blueprint
UNREADYour AI agent's real cost is cost per accepted taskOUTPUT · cost map
UNREADClaude Code without the chaos: a review-first workflow for coding agentsOUTPUT · review gate map
UNREADLog the decision, not the prompt: a trace schema for AI agentsOUTPUT · single-run trace review
UNREAD