A year of learning AI

Everything I learned, by month

November 2025

• A local model cannot search the internet (Local AI): What a model running on my own PC can do for a literature review, which is read the PDFs I give it, and what the step after that is.
• Running a language model on my own PC (Local AI): How I set up LM Studio on a home PC so a model runs with my data staying on the machine, and the GPU setting that took the longest to find.
• The first local model was too small (Local AI): An 8B model was not enough for technical questions on an RTX 4090, the 32B class is what the card should run, and training is not what I would be doing at home.
• A new chat does not remember the old one (Context and memory): In a Claude project the files carry over to a new chat, but the answers I typed into an earlier chat do not.
• A handoff file when the chat fills up (Context and memory): When a long chat hit its limit I had it write a markdown file for the next chat, with a note to look at the project files.
• Say which files it should read (How I talk to AI): Asked to review five uploaded documents, Claude reviewed the whole project folder instead; the fix was naming the files.
• Telling it to think like a professor (How I talk to AI): Asked for fundamental research questions, Claude drifted to lists of what other people did and to student sized projects; it took repeated, specific instructions to get a usable section.
• Ask me one thing at a time (How I talk to AI): When I asked Claude to interview me before drafting, it sent ten questions at once; one at a time, with the answers kept, is what I asked for instead.
• Retrieval over my own documents with AnythingLLM (Local AI): How I put AnythingLLM on top of LM Studio so a local model answers from a folder of my own PDFs.
• The context length error (Local AI): Why the first long PDF failed in AnythingLLM, with the error text, and the fix in LM Studio.
• A Drive link is not an upload (Context and memory): Claude cannot open a Google Drive folder from a link; the files have to be uploaded by hand.
• It writes the code and I run it (How I talk to AI): I told Claude plainly that I knew nothing about machine learning or Python, and we set the roles: it writes, I copy, paste, run and learn.
• What tokens are (Local AI): Tokens are the pieces a model reads text in, and context length is how many of them it can hold at once.
• Teach me from the ground up (Machine learning): At 4 AM I asked Claude to teach me machine learning from nothing; it asked three questions and offered three paths, and I took the parallel one.
• What prompt engineering is (How I talk to AI): I asked Claude what prompt engineering means and how to get the result I expect; six points came back, and a note that my prompts were already doing most of it.
• A first physics informed network on a reactor (Machine learning): On Colab I built a physics informed neural network for a first order reactor, in four parts, and watched it train.
• Colab and the first cell (Machine learning): I asked for a way to run Python without installing anything, got Google Colab, and saw my first output the same evening.
• Files I can answer and comment in (How I talk to AI): Instead of answering a long response in the chat, I asked for markdown files, a questionnaire, a roadmap and running notes, that I could fill in and paste back.
• A five phase plan to learn machine learning (Machine learning): I asked Claude to coach me through machine learning from Python up; the plan has five phases, and the first two evenings covered Python, NumPy, pandas, plotting and scikit-learn.
• You are making assumptions from what other people did (How I talk to AI): Claude had built a case for a predictive model on a small pilot dataset; I pushed back on where the idea came from and on what the data could support.

December 2025

• Explain the thing, then ask if it can be measured (How I talk to AI): An assistant wrote gibberish about a property it did not know, so I explained the physical picture and asked whether the property can be measured at all.
• Telling an assistant to think like a scientist (How I talk to AI): A document of research ideas came back generic twice, and what it took to say what real research starts from.
• A test matrix with phases nobody asked for (How I talk to AI): The assistant's first lab test matrix invented four phases and an optimization nobody requested, and one message with four points set it straight.
• Ask the assistant what the reasoning behind its matrix is (How I talk to AI): Asking Claude why its test matrix looked the way it did made it admit the matrix could not find the best ratio, and the optional component's range had to start at zero.
• Constrained mixture design for a three-component formulation (What I used AI for): A literature search for how a formulation test matrix is built came back with constrained mixture design, the frame the matrix had been missing.
• A sequential matrix before a formal design (How I talk to AI): When nobody has made the material before, a look-then-adjust matrix fits better than a formal simplex design, and the assistant agreed once I said why.
• No executive reads a wall of bullet points (How I talk to AI): The first draft of a document meant for a decision maker came back as bullet points, and one line turned it into prose.

January 2026

• Brainstorm in chat and ask before generating any documents (How I talk to AI): A planning chat with Claude went well while it asked me questions and badly when it handed me a report, so I now say at the start how I want to work.
• A truth check line by line on a draft Claude wrote (How I talk to AI): A one-page draft from Claude read well and carried a yield figure we never measured and equipment we do not own, so every line got checked against what is true.
• A tracker file when the to-do list outgrows the chat (Context and memory): When the steps Claude listed grew past what either of us could hold in the chat, it made a project tracker file and a next-steps file in the repository.
• Files that let me check code I cannot read (How I talk to AI): When I asked Claude how a non-developer could know its code was sound, it answered with an architecture decision record, standard patterns and a security checklist.
• GitHub Codespaces when the file download failed (Tools for AI work): The code Claude wrote would not download to my computer, so I opened a repository in a GitHub Codespace and ran it there with nothing installed.
• The first time Claude Code came up (Agents): Asking whether I could use Claude as an agent inside GitHub got me the name of Claude Code, a command line version that works in the repository itself.
• Elicit library and BibTeX instead of Zotero (What I used AI for): Because I write in LaTeX and had never used Zotero, the search results for my review go into the Elicit library and come out as BibTeX when it is time to write.
• Save my preferences and progress in markdown files (Context and memory): For a long piece of work I told Claude to keep my preferences and progress in its memory and in markdown files in the project folder, because chats run out and do not see each other.
• The first Elicit searches were too loose (What I used AI for): One broad Elicit query for my review found little, and what found the pilot-scale work was one query per theme, per named technology and per named feedstock.
• Replicate a published machine learning study before my own data (Machine learning): Before putting experimental data I had from earlier work through machine learning, I decided to take a published study with public data and redo its work end to end.

February 2026

• Do not just be a yes man (How I talk to AI): An assistant agreed with every correction and the drafts stayed poor until I told it to search and argue and defined my key term in one sentence.
• A saved profile file instead of restating my goals (Context and memory): The assistant kept forgetting my goals and undercounting my history, so I had it write a profile file for the project folder.
• Ask me multiple choice questions (How I talk to AI): When the assistant could not see my screen and my descriptions were not helping, I had it ask multiple choice questions instead.
• The citation link pointed at the wrong article (How I talk to AI): I downloaded an article from the link in an assistant's citation, worked on it for hours, and the link had not gone to the article it was citing.
• npm looks for package.json where you stand (Tools for AI work): Running npm install one folder above the project gave ENOENT on package.json; npm reads that file from the current directory.
• PowerShell would not run npm (Tools for AI work): A fresh Windows machine blocks the PowerShell script behind npm until the execution policy is set to RemoteSigned for the current user.
• VS Code and Git clone (Tools for AI work): My first code editor, signed into GitHub, with the clone command that did not show up and then did.
• A series needs something to compare against (Machine learning): One occupation's employment line meant nothing on its own, so I pulled the same series for neighboring occupations and indexed them to the first year.
• Ask where the numbers came from (How I talk to AI): The first chart in my notebook was drawn from numbers the assistant had typed in as placeholders, and nothing in the cell said so.
• Installing Python with the install manager (Tools for AI work): The Python install manager, a PATH entry, a fresh venv and a pip upgrade, two hours after the first error.
• Python was not found, with Python installed (Tools for AI work): Inside an activated venv pandas would not import, because the python command on Windows was the Microsoft Store stub and not a real interpreter.
• My first Python chart (Tools for AI work): The vapor pressure curve was the first chart I made in Python myself, and then I asked whether the fancy visualizations online can be made the same way.
• A playbook file carries the work between chats (Context and memory): For the modeling work I had the assistant write a playbook file, rolled to a new version at every stopping point, so a new chat starts where the last one ended.
• Out of the browser and into my own notebook (Tools for AI work): Watching the assistant run the models inside the chat was slow, so I moved the work to a venv on my machine, a script one step at a time, then Jupyter.
• The temperature column was the setpoint (Machine learning): The temperatures in a dataset I was working on were controller inputs, not reactor temperatures, and every earlier fit had to be redone.
• A script that failed on its own docstring (Tools for AI work): A backslash followed by U inside a Python docstring is read as a Unicode escape, and the whole file refuses to parse.
• An exponential through four points cannot be extended (Machine learning): An exponential fitted to four measurements and extrapolated back gave an impossible number on one feed and a plausible one on the other, and neither could be used.
• Every table and figure explained before I kept it (How I talk to AI): Running the analysis cell by cell, I pasted each output or figure back and had it explained before the next cell ran.
• pip was not recognized, so Jupyter from the venv (Tools for AI work): In plain PowerShell pip does not exist; it lives inside the venv, so I skipped the terminal detour and did the work in Jupyter started from the venv.
• The run that printed nothing and saved nothing (Machine learning): An Optuna hyperparameter search ran for 22 hours in a notebook cell with no progress output and wrote nothing to disk; everything sat in kernel memory.
• Bootstrap intervals, in plain words (Machine learning): Beside every score went an interval made by redrawing the rows with replacement many times and reading off the spread of scores.
• Leave one out, in plain words (Machine learning): On a small table every model was scored by fitting on all rows but one and predicting the one left out, row after row.
• 10_10 read from a CSV as 1010 (Machine learning): A neural network's layer sizes saved as 10_10 came back from the CSV as the integer 1010, with no error, and the model trained happily as the wrong network.
• Y-scrambling, in plain words (Machine learning): Shuffle the answers, refit, repeat a thousand times, and see whether the real score stands clear of what chance gives; scikit-learn has it as a permutation test.
• A sleeping PC and a closed browser both lose the run (Machine learning): Twice in one week a long notebook run was lost without the code failing, so every cell now writes its result to disk and the setup cells get re-run after a restart.
• A venv does not survive being copied (Tools for AI work): A virtual environment copied to another machine still pointed pip at the old path; the Python docs say recreate it, never move it.
• rmdir /s /q does not work in PowerShell (Tools for AI work): In PowerShell rmdir is an alias for Remove-Item, which takes -Recurse -Force and not the old slash switches.
• Ask whether it skipped a step (How I talk to AI): With the playbook open beside the notebook I noticed the assistant had jumped a step, and started asking at every phase what else it had skipped.
• KNN cannot be learning curved below its own k (Machine learning): A nearest neighbor model tuned to eleven neighbors crashed the learning curve at a subset of nine rows, twice, with an error that names both numbers.
• Learning curves on a small table (Machine learning): A learning curve refits the model on growing subsets and plots the fitted and held out scores against size; the first pass had no held out line at all.

March 2026

• A meter swap is not a swap (What I used AI for): Putting a new flow meter on the old wires looked like a swap until the range, the output setting, the unlabeled buttons and the wiring variant each turned out to be different.
• Check a live 4 to 20 mA loop with a multimeter before the swap (What I used AI for): A controller showed a healthy flow number with nothing flowing, and a fraction of a milliamp across the signal pins said the number was not coming from the sensor.
• A spreadsheet is a record, not a task manager (What I used AI for): A project tracker built as an Excel file was fine for the record and useless for adding a task in a minute, so the tasks moved to a board.
• Find the manual and name the page (How I talk to AI): The assistant offered scaling arithmetic with no basis, and the fix was to make it find the manual, give me the link and the page, and answer from that every time.
• Advice that flips between sessions (Context and memory): The assistant pushed a course in one session and argued against it in the next, and the fix I settled on was written context files rather than better prompts.
• A brain file with the mistakes in it (Context and memory): Before moving to a new project I had the assistant write a Markdown file of everything it knew, including the mistakes it had made and how not to repeat them.
• Installing Claude Code on Windows from PowerShell (Agents): One line in PowerShell installed Claude Code, and the installer itself said the folder it went into was not on the PATH yet.
• The first figure had the air going the wrong way (How I talk to AI): I asked the assistant about air bubbles reaching a pump suction, and the sketch it drew had the air flowing upward when the sparger blows down.
• Two files to carry a project into Claude Code (Context and memory): A context file for CLAUDE.md took three versions before it held everything, and a work history written from the chats went with it.
• What did the simulation actually do (What I used AI for): I asked for a flow simulation to back a decision on a plate size, got a picture, and had to ask what the picture had computed.
• Typing plan mode at the chat (Agents): I typed a slash command into the chat app to see what it would do, and nothing came of it, because plan mode lives in Claude Code and is reached another way.
• A photo of a data sheet into a printable table (What I used AI for): Three passes turned a phone photo of a handwritten lab data sheet into an Excel table that prints on one A4 page.
• A menu path the assistant invented (How I talk to AI): The assistant told me where to find a raw value in the controller software, and the menu it named does not exist, because it guesses at interfaces it has not seen.
• Grounded thermocouples on a steel wall read zero (What I used AI for): New thermocouples read zero where they touched the reactor wall, and the reason was in one letter of the model number.
• Two meters pegging the same input (What I used AI for): An analog input read full scale with no flow under the old meter and again under the new one, which put the fault in the channel or the wiring, not in either meter.
• A multimeter on the amps range on a milliamp loop (What I used AI for): Checking a 4 to 20 mA output with a meter that only has a DC amps range gave a negative reading in whole amperes, which said the meter setup was wrong, not the loop.

April 2026

• A HAZOP taken as a checklist, node by node (What I used AI for): Before the first hot run on a small pilot unit I had an assistant walk the drawing with me as a HAZOP, and took the output as a checklist instead of a full worksheet.
• Three runs at the same conditions gave three yields (What I used AI for): Three flow-through runs at the same conditions gave three different yields, and I did not accept an explanation that blamed something equal in all three.
• A reactor from pipe and flanges, where the flange is the limit (What I used AI for): Checking whether a heavy-wall pipe with flanges welded on can be a high pressure pilot reactor, under the piping code and the flange standard, and finding the flange at temperature is what sets the rating.
• Asking for the references behind each design decision (How I talk to AI): After a reactor design built up in a chat, I asked for the standards behind every choice and cut the answer back from a report to one comparison table with links.
• Different steel grades weld fine, the filler does the work (What I used AI for): I assumed the pipe and the flanges had to be the same stainless grade to weld; closely related grades weld fine and the filler metal is what matters.
• Threaded ports are the weak point on a high class blind flange (What I used AI for): A threaded pipe port bored in a blind flange would have undone the rating of a high pressure reactor, and a weldolet with a tube connector was the answer.

May 2026

• The assistant invented reactor sizes nobody makes (How I talk to AI): An analysis from the assistant rested on autoclave sizes the manufacturer does not make, and I told it to stop making things up and give facts on real references.
• The manufacturer's rating of an autoclave governs, not the flange curve (What I used AI for): I asked whether an autoclave's spare pressure rating could be traded for extra temperature along a flange curve; the manufacturer's rating of the assembled unit is the one that governs.
• Ask for the list, not the lecture (How I talk to AI): A sourcing question came back twice as a long reply full of caveats, and a short list only once I said how long the answer should be.
• Asking for literature with and without stirring (How I talk to AI): For an unstirred batch extraction, asking whether it would work got me an essay; asking for a published comparison of residence time with and without stirring is the question that can give a number.
• Every answer needs references, saved to the assistant's memory (Context and memory): After asking for sources reply by reply, I told the assistant to save in its memory that every answer has to cite the sources it rests on.
• A guitar put to the assistant in its shoes (What I used AI for): I gave the assistant my fifteen years of on and off guitar as if they were its own, and asked for its game plan and what it would play in the next half hour.
• Furniture legs with the assistant at the hardware store (What I used AI for): Standing in a hardware store with a broken futon at home, I asked the assistant one thing at a time what I needed to make new legs.

June 2026

• Starting a robot project as something the two of us can figure out (What I used AI for): How I opened a robot build with an assistant: set the terms, answer its setup questions, ask for parts with pictures, then ask to see the thing before buying anything.
• A coffee maker designed as a bench scale process (What I used AI for): The first drawing for a bean to cup coffee maker was a drip rig; describing the unit operations in order got the right drawing the second time.
• Asking about tires as what the locals use (What I used AI for): Choosing tires for my truck by asking an assistant what people around me actually run, then giving it my driving conditions and the shop's inventory.

July 2026

• Offline home automation on a small board instead of an old tablet (What I used AI for): Picking a pet project, home automation that runs completely offline for privacy, and choosing a Raspberry Pi for the hub rather than wiping an old tablet.
• A profile file to paste into any new chat (Context and memory): Asking an assistant to write my profile and working preferences into one markdown file so a new project starts with the background already in it.
• Looking at a project car the way I would rebuild a pilot plant (What I used AI for): Reading car listings from screenshots with an assistant, and asking whether a car exists where every line could be replaced with stainless tubing the way we do on a pilot plant.
• A referenced search on an old interest, iron as a rechargeable fuel (What I used AI for): Going back to an old idea, burning iron powder and reducing the oxide with renewable power, by asking for a literature report with references and then the one unsolved problem.
• Asking how can we claim that before posting a line about research (How I talk to AI): An assistant offered a quotable line about the hardest part of research; before using it I asked for the evidence and got a referenced write-up with the claim's limits.
• A stool from a single board, where the drawing came on the third ask (What I used AI for): Asking an assistant for designs of folding furniture cut from one board, and sending the same request for a drawing three times before one arrived.
• Briefing an assistant on a chiller fault outside my field (What I used AI for): A chiller on a pilot unit kept losing control power; saying I am a chemical engineer, not an electrical one, and giving the history in order got a diagnosis beyond the first guess.
• Why a chiller cannot run off a drive meant for a pump motor (What I used AI for): Asking whether a spare chiller could be powered from the variable frequency drives feeding pump motors, and getting a no with the reason.
• Working out the liquid in a round pot from a level reading (What I used AI for): Delegating the spherical cap arithmetic for a round pot to an assistant, running it on two bases, and rerunning it when the level reading turned out wrong.

August 2026

• Two pH meters against the standard buffers (What I used AI for): Both bench pH meters read high on the standard buffers, and without the right buffers the error could be measured but not calibrated out.
• The basic questions asked again in one message (Machine learning): The simulation and machine learning terms I kept meeting, asked about in one message and then compiled as a tutorial with pictures.
• A digester motor that lifted the drum and flipped back (What I used AI for): A new rotary digester would not rotate, and the fault was torque and wiring, not a control setting.
• A table with not stated where the value is missing (How I talk to AI): Turning a messy bench log into a clean table with one instruction that keeps the assistant from filling the gaps.
• Code that runs clean and reports nothing (How I talk to AI): An exercise where the assistant's code is run as the next cell and the output is read against what you know of the material, not of the code.
• The same notes with and without a brief (How I talk to AI): Two prompts from the same bare notes, one of them with a full brief in front, came back as two different reports.
• Half an hour of keys before I gave it the manual (How I talk to AI): Setting a cook temperature on a pilot unit's controller took a photo, a lot of wrong key presses and finally the manual.
• Find the manual online, then give me the steps (What I used AI for): A photo of a bench pH meter and one prompt got the manual found and a step by step calibration written for that model.
• Do not just agree with me (How I talk to AI): Asking the assistant to critique a knowledge base idea instead of agreeing with it.
• Planning the home PC as a fully local AI workstation (Local AI): What the home PC with an RTX 4090 should do as an AI workstation, and the rule that nothing leaves the machine.
• npm install in a Downloads folder is not a dev environment (Tools for AI work): A first npm install, run inside a Downloads folder with (1) in the path, failed with EBUSY, and the note I kept was to use a proper folder.
• Asking whether to use a much stronger acid (What I used AI for): I asked whether a much stronger acid would do for bringing a batch of process liquor down to a low pH, and took the argument against it.

September 2026

• First session with a command line agent on an empty repository (Agents): What the command line agent did, in order, when I asked it to write the rules file for a repository that had nothing in it yet.
• Set the git author globally so agent commits carry my name (Tools for AI work): An agent's fix to the commit author did not survive its container, so the setting belongs in the global git config on my own machine.
• Asking an assistant for guitar patches with the gear named (What I used AI for): I asked for two patches on my guitar processor and the assistant asked how the boxes were wired before it wrote anything.
• Reading codes on the car with an OBD reader (What I used AI for): I sent the assistant photos of a fault code from my car, cleared it, and had to tell it that one red light was only the seat belt.
• A context file is for the author, not the outline (Context and memory): I carried a context file into a new project for a long writing project and the assistant built the chapter spine from it before I had given the thesis.
• A render that finishes is not a finished video (Agents): A 4K manim render came out in two minutes at 3.39 MB, and only a look at the bitrate showed what was wrong; what the agent found about the encoder, the PNG export and the cache.
• A venv on an older Python because manim had no build for the newest (Tools for AI work): The machine had Python 3.14 and manim did not ship wheels for it yet, so the agent set the virtual environment up on Anaconda's 3.13 instead.
• The command line agent cannot see my chat projects (Agents): The command line agent has no access to the chat side's project folders; it wrote the script from what I had pasted and repaired the mangled characters.
• What Hugging Face is (Local AI): I asked the assistant what Hugging Face is, then read its own documentation: a hub of models, datasets and demo apps kept in Git repositories.
• Ask the question and hold the file (How I talk to AI): The assistant rewrote a procedure I had not asked it to change and skipped the question I had asked; the ask that works is the question on its own.
• Is it worth learning all this (How I talk to AI): The question I asked the assistant after midnight about whether to keep going with AI, what I want from it as a chemical engineer, and the next question two days later.
• A cost column in my own calculator, not a report (What I used AI for): For an order of magnitude estimate the right ask was a cost column added to my own Excel calculator and a one page summary tab, with the other tabs left alone.
• Obsidian for all my notes, with a me file (Context and memory): I moved my notes out of chat projects into Obsidian, plain Markdown files on my own disk, with one me.md file holding my preferences, memory and instructions.
• Upgrading a 2024 desktop for local models against a unified memory machine (Local AI): The questions I asked about my Core i9 and RTX 4090 desktop for running local models, and why they all came back to memory, about 128 GB.
• You are overdoing this (How I talk to AI): I pasted a posting that wanted partners and asked how we would do the work; the assistant wrote a plan when a raised hand in three sentences was the reply.

October 2026

• A big ask has to be cut into steps (Context and memory): Asking for notes on every past conversation at once ran into the context and the usage limit, so I asked how to do it in steps.
• A task done once by hand became a command (Agents): The first time Claude Code turned an article PDF into a vault note it did it by hand; the second time it was a skill I ran with one command.
• Creating a project needs its own word (Agents): A setup command that drafted a scope from old chats when the intake folder was empty taught me that init should only set up and stop.
• Loading less per command beats adding commands (Agents): I asked for a review of my skills and got back the view that the larger gain is in loading less per command, not in adding skills.
• Moving the folders broke the commands (Agents): Moving my Workspace and vault off OneDrive broke the skill junctions and stranded the memory notes under the old path, and here is what the repair touched.
• Testing a skill like a lab method (Agents): A unit checking skill was tested with a reference problem of planted errors, pass criteria and blind runs on two models, and the transcripts were audited for a stray file.
• The assistant counted my re-scopes back to me (How I talk to AI): Twice in one night, before changing a plan the way I asked, Claude read me the plan's own count of how many times I had changed direction.
• Work and personal never mix in the tools (Agents): A skill built for work refused a personal project, so I had the separation of work and personal written into it as the rule above all others.
• A skill reads whatever it can reach (Agents): Watching a skill run showed it reading files it did not need, so I named its three inputs and nothing else.
• Questions one at a time, with choices (How I talk to AI): A block of open questions is not how I want to be asked; one question at a time with multiple choices is, and the next question should come from my last answer.
• A drawing carries specifications, not its own history (How I talk to AI): Notes on a drawing that said where each change came from told the reader a machine drew it; the sheet now carries ratings, set pressures and line sizes only.
• A guide written from one session of corrections (Agents): After an evening of corrections on one drawing I had them written down as a style guide, and the guide became a skill the same night.
• A model trained on pictures of drawings would check nothing (Agents): I asked how to train a custom model on drawings; the answer I acted on was a structured record of the plant with a rule checker, and a drawing drawn from the record.
• It answered my Bangla in Hindi (How I talk to AI): A question I asked in Bangla came back answered in Hindi, and I had to tell the assistant which language my non-English messages are.
• Remote Control steers the session, it is not a remote desktop (Agents): Turning on Remote Control let me follow a Claude Code session from my phone, but the session and the files stay on the desktop PC.
• Too much text describing every detail feels AI generated (How I talk to AI): I told the assistant I do not want text explaining every detail of a deliverable, and that it should notice how I prefer things and save that.
• Two kinds of memory, one I write and one it writes (Context and memory): Claude Code carries two things across sessions, a rules file I write and the notes it writes itself from my corrections, and both got written to that week.
• A plan grilled one role at a time (How I talk to AI): A plan reviewed as a dev lead, then a data scientist, then a project manager, then a domain expert, each pass finding what the one before did not.
• Drive letters out before the vault goes into git (Tools for AI work): Putting my Obsidian vault on GitHub meant cleaning it first and taking the drive letters out of every path so it works on any device.
• Good to go was taken as the word to merge (Agents): Saying good to go got a branch merged, tagged and pushed, and later the same day the rule went in that approving a plan is not the start signal.
• Months of work run in parallel agents (Agents): Turning an export of old chats into daily notes went month by month until I asked for all the remaining months at once, and Claude ran them as ten subagents in parallel.
• The transcripts are saved, but not for ever (Agents): Claude Code keeps a transcript of every session on the machine and deletes old ones after a retention period, so I had them fed into my daily notes.
• A keyword search on one index is a scoping pilot (What I used AI for): A keyword search on an open index is a pilot; a search you can report needs one master string, every hit retrieved, a written protocol and the databases named.
• A screen a person can do from title, abstract and conclusion (What I used AI for): Three rich templates for screening articles by hand went in the bin; what stayed is one row per article, three things to read and two ratings.
• If it takes time, I do not want it (How I talk to AI): I asked for a quick desktop app to read fetched abstracts, said I did not want it if it took time, and the honest advice was not to build it.