A little AI.
A lot of
possibility.
Ideas worth exploring. Things worth building.
Follow your curiosity.
Good knowledge.
Lasting impact.
Some people’s work stays with you.
Matt’s has stayed with me.
Matt Murphy has been a wealth of knowledge for me. His explanations have helped me work through problems more times than I can count.
This feature is my way of saying thank you—and introducing you to someone whose work I value. Matt, I appreciate what you share.
— JAB Systems
Original perspective: Matt Murphy · @mattmurphyai
Appreciation and accompanying field note by JAB Systems.
Every word
carries weight.
A little context helps. A lot can tip the balance. Try adding a little more—and watch what happens.
A visual experiment in knowing when to stop.
Read “Your prompt has a budget” →
No field notes match. Try another topic.
AI workflows / Practical field guide
Why is it sideways?
Check EXIF orientation before resizing images for a vision workflow with Pillow.
Read the field guide →
AI workflows / Practical field guide
Where did the note go?
Refresh an AI summary while preserving the human review note with SQLite.
Read the field guide →
AI workflows / Practical field guide
No became False.
Check loaded YAML label types before an AI workflow uses them.
Read the field guide →
AI workflows / Practical field guide
Page 1 of what?
Keep PDF positions and page labels together for AI citation review with pypdf.
Read the field guide →
AI workflows / Practical field guide
Round when?
Compare per-job and batch rounding for AI usage credits with Python Decimal.
Read the field guide →
AI workflows / Practical field guide
Still fits?
Check a stale Git patch without writing files, then update it against the current copy.
Read the field guide →
AI workflows / Practical field guide
Wrong inbox?
Run an offline Python batch lab and keep each summary with the ticket it came from.
Read the field guide →
AI workflows / Practical field guide
Who flipped the switch?
Run a jq fallback lab and keep an explicit false from turning into true.
Read the field guide →
AI evaluation / Practical field guide
Same answer? Look closer.
Run a Python Unicode lab to see why an exact-match check can reject equivalent answers.
Read the field guide →
AI workflows / Practical field guide
Did it queue twice?
Run a SQLite lab that distinguishes a replayed action from a changed payload or a new request.
Read the field guide →
AI workflows / Practical field guide
Who counted this ticket twice?
Run a pandas lab that catches duplicate account lookups and flags unmatched tickets before an AI summary.
Read the field guide →
AI workflows / Practical field guide
Which flag did your agent mean?
See what Python does with a shortened command flag, then require the full spelling in a ten-case offline lab.
Read the field guide →
AI workflows / Practical field guide
Did AI summarize the cookie banner?
Choose a web page’s source region before prompting. Run the Beautiful Soup lab and inspect the correction a selector can miss.
Choose the source
AI workflows / Practical field guide
Your AI regex matched too much.
Give a drafted regex a second label, inspect what it captures and review the accepted format with a ten-case practice lab.
Test the second example
AI workflows / Practical field guide
What did the test click?
Inspect a failed Playwright run, find the button it clicked and repair the test while keeping the assertion.
Trace the click
AI agents / Practical field guide
Your AI agent's path is not permission.
Test file requests against a trusted root. Reproduce a sibling-prefix trap and symlink escape, then separate location checks from access permission.
Check the file scope
AI workflows / Practical field guide
Your AI JSON has two answers.
Catch duplicate JSON keys before a decoder discards them. Run ten authored cases, including nested and escaped names, before schema validation.
Catch the duplicate
AI tools / Practical field guide
Give your AI tool a fake bad day.
Use HTTPX MockTransport to try success, server errors, malformed data and timeouts. Keep a failed lookup distinct from a negative answer.
Test the handoff
AI workflows / Practical field guide
Your prompt has a blank spot.
Check missing context with Jinja StrictUndefined. Try empty and null inputs, then add an explicit required-text policy.
Check the prompt
AI workflows / Practical field guide
AI gave you a number. Where’s the unit?
Check power, energy and temperature with Pint. Run six authored cases and keep source units and assumptions visible.
Check the unit
AI workflows / Practical field guide
A 100 score. Still two records?
Shortlist fictional names with RapidFuzz. Inspect a 100 token-set score, preserve original records and keep the merge decision with a reviewer.
Review the match
AI-assisted coding / Practical field guide
AI wrote a CSV export. Try a comma.
Use Hypothesis to find a counterexample and check a CSV round trip. Run seven authored edge cases and a bounded generated test.
Test the CSV
AI workflows / Practical field guide
Which 1:30 did AI mean?
Check an AI-drafted meeting time for repeated and skipped hours. Try seven local authored cases before any calendar write.
Check the wall time
AI services / Practical field guide
Your AI request failed. Should it retry?
Read wait hints, inspect SDK retry settings and set a budget. Try ten offline authored cases without an API key.
Read the signal
AI-assisted coding / Practical field guide
AI suggested a package. Does it exist?
Check public PyPI metadata before installing an AI-suggested dependency. Try three authored cases and keep package trust as a separate decision.
Check the package
AI research / Practical field guide
AI gave you a citation. Does it exist?
Compare a DOI and title against Crossref, try three authored cases, and keep source verification separate from a successful lookup.
Check the citation
AI-assisted coding / Practical field guide
AI wrote the script. Can you rerun it?
Keep a reviewed script and its dependency lock together. Run a fictional uv example and see what the lockfile can—and cannot—check.
Rerun the script
AI-assisted analysis / Practical field guide
Let AI draft the query. Check the math.
Inspect an AI query draft, run a fictional CSV locally with DuckDB, and see how one missing filter changes a checkable answer.
Check the math
Local AI / Practical field guide
Check the download before the model.
Preview selected Hugging Face files at a pinned revision, read the proposed transfer, and keep download size separate from runtime claims.
Preview before downloading
AI publishing / Practical field guide
Your AI graphic needs alt text.
Draft a text alternative with context, check it against the image, and inspect the same graphic in three HTML roles.
Give the image another way in
Prompt preparation / Practical field guide
Your prompt doesn't need their email.
Replace selected fictional identifiers locally with Presidio, then inspect the missed email and context that survive.
Inspect the local lab
Agent control / Practical field guide
Give your AI agent a stop rule.
Run four scripted scenarios and inspect completion, a decision cap, repeated calls, and a denied action before connecting real tools.
Try the local lab
Local transcription / Practical field guide
Your transcript is a draft.
Run Whisper on a synthetic English clip, inspect the actual transcript and timestamp files, and keep a listen-back review before treating the draft as a record.
Try the local walkthrough
AI-assisted edits / Practical field guide
The AI edit you didn't ask for.
Keep an original draft and compare two fictional revisions with Git. Spot changed dates, owners, and approval language before accepting an edit.
Review the changesKeep exploring · 11 more field guides & checklists
Local retrieval field guide
Search before you chat.
Inspect a local SQLite search baseline with three fictional notes, ranked excerpts, source IDs, and a missing-term test before adding AI answers.
Find the evidence
Resource · AI handoffs · Tutorial
Valid JSON can still be wrong.
Validate six fictional handoffs with JSON Schema, catch missing fields and a premature action, and see why a passing result still needs source review.
Read walkthrough
Resource · Context budgets · Tutorial
Your prompt has a budget.
Count two fictional context files locally with tiktoken, preserve the essential facts, and learn why a text count is only part of an API request.
Read walkthrough
Resource · Agent tools · Tutorial
Before you connect an agent, read the tool list.
Inspect a local MCP demo, review its inputs, and learn why declared hints are not permissions. Includes tested commands and a connection checklist.
Read walkthrough
Resource · Prompt evaluation · Tutorial
Make your prompt earn your trust. — build five checks with Promptfoo.
Grade fictional answers, catch two deliberate mistakes, and prepare an optional model trial. Includes a complete download and clear verification boundaries.
Read walkthrough
Resource · Open source · Tutorial
Before you prompt, fix the input. — inspect your document with MarkItDown.
Convert a fictional document to Markdown, check what survived, and try a source-grounded review prompt. Includes the actual HTML conversion output and clear limits.
Read walkthrough
Resource · Local AI · Tutorial
Less chat. More decisions. — build a local AI ticket classifier.
Turn a sample request into Billing, Technical, or Other with Ollama. Get the complete PowerShell example, test your categories, and keep a person in the review loop.
Read walkthrough Resource · Tutorial · FreePrompt Patterns That Hold Up — a worked walkthrough.
Watch one weak prompt get rebuilt, pattern by pattern, until it survives real inputs. The companion example to the Prompting Checklist.
Read tutorial Resource · ChecklistSecurity Checklist — ship React + TS + SQL Server safely.
A practical, no-hype application security checklist. Open it on new projects or audit what's already live.
Open checklist Resource · ChecklistPrompting Checklist — write prompts that hold up under real inputs.
Role, task, format, examples, constraints, evals. Use it before shipping any AI step.
Open checklist Resource · ChecklistOptimization Checklist — make it fast, cheap, and quiet.
Model choice, caching, batching, latency, token spend, observability. Use it when "it works" needs to become "it scales."
Open checklist
AI Anonymous / Field noteWho chose your side of the fork?
Big AI vs. sovereign AI. Building on a reel by @mattmurphyai: what is at risk, what sovereign looks like, and five questions for your own AI stack.
Read the field guide →A question. An idea.
A conversation.
Something from the field notes catch your eye? Have a question about a post or an idea to share? Get in touch.