tishad.net
SearchCtrl K

A journey by AI

Lessons on using AI well, in the order a newcomer should meet them. Each one is the same shape: the ask people usually make, what comes back, and the ask that works. 72 so far, more as I learn. Drafted with Fable 5.1 from my own notes, then chosen and edited by me.

Part 1. Chat assistants and prompting

  1. 01Write a brief, not a line
  2. 02State the form and the length of the answer
  3. 03Have it ask you questions one at a time
  4. 04Say which files it should read
  5. 05Ask the question or ask for the change, not both
  6. 06Give it a role and a number
  7. 07Describe the process in order before asking for a drawing
  8. 08Name an object, not an adjective
  9. 09State what is fixed, or it invents structure
  10. 10When words do not converge, edit the file and hand it back

Part 2. Context and memory

  1. 11A new chat starts empty, so write a handoff note
  2. 12Cut a big job into steps that each leave a file
  3. 13Standing context lives in files, one per topic
  4. 14Give it the document and tell it to work from it

Part 3. Checking what an assistant tells you

  1. 15Ask what the claim rests on
  2. 16A citation is a lead, not a reference
  3. 17Check generated text line by line
  4. 18Ask where the numbers came from before charting them
  5. 19A cause written on a chart is a hypothesis
  6. 20A clean run is not a checked result
  7. 21Tell it not to agree with you
  8. 22Reading a document someone wrote with a model
  9. 23Using an assistant for a literature search

Part 4. Documents and data with an assistant

  1. 24A number read from a long document is a pointer to a page
  2. 25Tables from photos, one sheet per request
  3. 26Catalogue a large archive by script before a model reads it
  4. 27Write the timeline before asking for the cause

Part 5. Writing and publishing with AI

  1. 28Machine sounding drafts need a style guide first
  2. 29A scripted gate before anything goes out
  3. 30A bot speaks for a body of knowledge, not as a person

Part 6. Running a model on your own computer

  1. 31A local model reads what you give it, at the context you set
  2. 32Write the jobs down before buying hardware

Part 7. Python, notebooks and the tools around them

  1. 33Start in a hosted notebook and install nothing
  2. 34Run the analysis where you can run it again
  3. 37Getting instrument files into a shape Python can load
  4. 38A column heading does not say what kind of number it holds
  5. 39A long run writes to disk as it goes
  6. 40Git for one person

Part 8. Machine learning on small data in plain terms

  1. 41Machine learning words an engineer already knows
  2. 42A first model on a problem you can solve by hand
  3. 43Reproduce a published study before your own data
  4. 45Flag outliers, do not remove them
  5. 46What a flexible model adds on a small table
  6. 47Tree ensembles in plain terms
  7. 48Kernel models in plain terms
  8. 49A small neural network in plain terms
  9. 50A hyperparameter search in plain terms
  10. 51Leave one out validation, and what its score leaves out
  11. 52Predicted against actual is the first chart
  12. 53A bootstrap interval in a few lines
  13. 54Shuffle the target and refit
  14. 55A fit cannot be extended beyond its points
  15. 56Before a second dataset can test a model
  16. 57What a SHAP value says and what it does not
  17. 58Prediction and interpretation are two uses of a model
  18. 59An ablation study in plain terms
  19. 60Five checks on a small data model

Part 9. Figures, drawings and deliverables made from code

  1. 61Make figures from code, not from generated pictures
  2. 62Draw from a record, and put every correction in a checker
  3. 63A deliverable carries specifications, not its history

Part 10. Coding agents, commands and skills

  1. 64A tracker and written decisions before any build
  2. 65Review a plan one role at a time
  3. 66Approving a plan is not telling it to start
  4. 67The first session on an empty repository
  5. 68The installed version is not the one the plan named
  6. 69Tag the version that works before a redesign
  7. 70Ask for the count of what was checked
  8. 71A process an agent starts can outlive the session
  9. 72Do the task once, write the corrections as a guide, then save a command
  10. 73Test a skill against a run without it
  11. 74Separate the agent that reads a source from the one that writes
  12. 75Two copies kept in step need a rule for which one wins