Nightly call-notes bots

Nobody on the team types call notes anymore. Every weekday evening, scheduled jobs read that day's recorded client calls, fill in the CRM, write a short narrative of what mattered, and post a Slack summary that tags whoever took the call and asks them to check it. There are separate bots for sales calls, planning calls and intake calls.

Why

The first calls with a new client hold details that get lost before anyone has made a proper record. The narrative turned out to be the part the operations team values most.

How I built it

1. I wrote a runbook in plain English: where the calls are, which ones count, what to write where.
2. I had Claude Code pull the last 24 hours of recordings from the notetaker, skip internal meetings and match each call to a CRM record.
3. I told it that if it wasn't sure of the match, it should alert Slack and write nothing.
4. I added anti-hallucination rules. The main one is a quote test: can it point to the exact sentence? If not, it doesn't write it.
5. I scheduled it with Windows Task Scheduler and ticked "Run task as soon as possible after a scheduled start is missed."
6. I made every run post a heartbeat to Slack, even with zero calls.

To copy it you need

Claude, or the AI of your choice, a meeting recorder with an API, a CRM with an API, Slack, and a computer that stays on. The gotcha: "they're in LA" does not mean you know the zip code. Tell the AI to leave a field blank instead of guessing.

Try it yourself

I want to copy Frances Tang's nightly call-notes bots from her What Are We Building profile. Each evening her AI reads that day's recorded client calls, fills in the CRM, writes a short narrative of what mattered, and posts a Slack summary asking the person who took the call to check it. Her rule is a quote test: if the AI can't point to the exact sentence in the transcript, it doesn't write the field. If it isn't sure which client a call belongs to, it alerts someone and writes nothing. Help me build a starter version: I'll paste one call transcript and the list of fields I track, and you fill them in under those rules and show me the sentence behind each one. Start by asking which fields I track. Ask me for anything you need.

Tools I used for this project

Claude Code: Runs unattended each evening, reads a runbook file and follows every step.Monday.comSlack: Where each summary and a nightly heartbeat get posted.Windows Task Scheduler

The Radar

It answers one question a few times a day: who is waiting on us? The Radar is a single Slack message that sorts a healthcare team's message threads into needs a reply, keep an eye on it, and a quick recap, so no client message gets lost.

Why

The core job is that nothing gets lost. A message unanswered for more than two hours in business hours gets flagged.

How I built it

1. I got read-only access to the messaging tool and a private Slack channel for the digest.
2. I had Claude Code write a small script that fetches only the activity since the last run.
3. I wrote the rules for sorting: red needs a reply, yellow is one to watch, green is the recap.
4. I wrote the privacy rules: first name only, never quote the content, summarize what the person wants.
5. I set it to move the "last run" marker only after the Slack post succeeds, so a failed run gets covered next time.
6. I scheduled it on my own PC.

To copy it you need

Claude, or the AI of your choice, a messaging tool with an API, Slack, and an always-on computer. The gotcha: an unattended cloud job will not be allowed to pull sensitive records. Plan for a machine with a person behind it.

Try it yourself

I want to copy The Radar from Frances Tang's What Are We Building profile. A few times a day her AI checks a shared inbox and posts one Slack message: who is waiting on a reply (red), what to keep an eye on (yellow), and a short recap (green). It uses first names only, never quotes what people wrote, and never invents a name, a wait time or a count. Blank beats wrong. Her gotcha: she tried to run it as an unattended cloud job and the AI's safety system refused to fetch sensitive records with no person behind the request, so hers runs on her own computer. Help me build a starter version for my shared inbox or support queue. Start by asking what inbox I use and how long is too long to wait. Ask me for anything you need.

Tools I used for this project

Claude Code: Reads only what changed since the last run, sorts each thread and writes the digest with first names only.Slack: Receives the one message in its own channel.Windows Task Scheduler

The vault every AI reads first

A new AI tool starts out knowing what the old one knew. The vault is a folder of plain text notes that holds who I am, how each brand sounds, where every project stands and my standing rules. Claude, Codex and anything else reads it before doing any work, and writes back to it. The vault is the permanent layer. The AI tools are interchangeable.

Why

Every tool has its own private memory, and that locks you in. I wanted everything automatically pushed and synced, in files I own. A normal personal wiki answers "who am I." This one answers "how should any AI work with me, right now."

How I built it

1. I made a folder of plain text notes and opened it in Obsidian.
2. I added one front-door file that tells any AI: read the rules folder first, the task list last.
3. I put the rules in one folder (who I am, how I like to work, how each brand sounds) and everything else in ordinary pages.
4. I set three levels of permission: rules need my approval in chat, the task list and knowledge pages can be written freely but must be reported.
5. I put it in a private GitHub repository and scheduled a script to save and upload on a timer.
6. I pointed each AI tool's own memory back at the vault and kept that memory nearly empty.

To copy it you need

Claude, or the AI of your choice, as long as it can read files, plus Obsidian and a private GitHub repository. The gotcha: upload permission has to be set up on every machine, and it is the step everyone forgets.

Try it yourself

I want to copy the vault from Frances Tang's What Are We Building profile. She keeps a folder of plain text notes that every AI tool reads before it does any work: who she is, how she likes to work, how her brands sound, and where each project stands. One front-door file tells the AI to read the rules first and the task list last. Rules need her approval to change; everything else the AI can update as long as it reports what it changed. Her gotcha: tools kept saving things to their own private memory until she told each one to point back to the vault. Help me build a starter version with three files: about me, how I like to work, and my current projects. Start by interviewing me for the first one. Ask me for anything you need.

Tools I used for this project

Claude Code: Reads the vault at the start of every session and writes back what changed.Codex: Reads the same front door, so both AIs follow the same rules.Obsidian: The free app I open the notes in.GitHub: Keeps a private copy and the full history.Windows Task Scheduler

The /handoff routine

I can close a chat and open a new one without re-explaining anything. When I type /handoff, or just say "let's wrap up," Claude writes down what was decided and why, what we learned, what is still open and the exact next step, and saves it where the next session reads it automatically.

Why

I lost six hours once because a session proposed rebuilding something without ever finding the 15-step plan that already existed. Sessions were also creating tasks and never reading them, so tasks piled up unseen.

How I built it

1. I wrote the routine as one text file of steps, and a one-line command that says: follow every step in order.
2. I had it pull out decisions with the why, things figured out, open questions and next actions.
3. I made it check every claim about the outside world and label it before saving.
4. I made it go through open tasks one by one and ask me before closing any.
5. I had it write a short "start here" note that loads automatically next time.
6. I had it say so plainly if I spent the session on something low priority while a top priority sat open.

To copy it you need

Claude, or the AI of your choice, and a notes folder it can write to. The gotcha: if the AI summarizes loosely, the next session inherits a guess. Make it use an exact structure and mark what it didn't check.

Try it yourself

I want to copy the /handoff routine from Frances Tang's What Are We Building profile. At the end of a working session she says "let's wrap up" and her AI writes down what was decided and why, what was learned, what is still open and the exact next step, then saves it where the next session will read it first. Every claim gets labeled verified, corrected or unverified, and unverified claims are not saved. Her gotcha: an early version carried an unchecked claim into the next session and she lost hours to it. Help me write my own version as a short set of steps you follow whenever I say "wrap up," ending with a block I can paste into a new chat. Start by running it on this conversation. Ask me for anything you need.

Tools I used for this project

Claude Code: Runs the routine from one skill file and loads the result at the start of the next session.Obsidian: Where the session log, the project status and the next-step note get saved.

Bake Sale

It only ever shows me one task. Bake Sale is a desktop to-do app built like a bakery: I place orders, work them one at a time, and every finished task becomes a pastry on that project's shelf. Small tasks are cookies. Big ones are layer cakes. It opens itself every morning.

Why

A long list makes me survey instead of work. The one-at-a-time limit is the whole product, not a nice-to-have. It has since become my only task system, and the next problem I gave it is that our work is disappearing into a backlog.

How long

First version in one session, July 15, 2026.

How I built it

1. I wrote a spec with a hard rule (one task in progress, ever) and a list of what it would not do.
2. I had Claude Code build the one-task loop and the shelves first and save the art for last.
3. I added a test that fails the build if the working view ever shows the queue.
4. I had the pastries illustrated one object at a time from one photo reference.
5. I set Windows Task Scheduler to open the app every morning.
6. I later moved the tasks into a database so other tools could add to them.

To copy it you need

Claude, or the AI of your choice. A desktop app needs a few free developer tools installed, which your AI can walk you through. The gotcha: the desktop icon runs a saved copy. Changes don't show up until you rebuild it, and I once used a version that predated a whole feature.

Try it yourself

I want to copy Bake Sale from Frances Tang's What Are We Building profile. It is a to-do app with one rule: it only ever shows the single task she is working on, never the list. Finished tasks turn into pastries on a shelf per project, sized by effort, so she can see what she shipped. Her gotcha: an early version couldn't tell an empty list from a list it failed to load, and it saved the empty one over the real one, so make it refuse to save when loading failed. Help me build a starter version as a single web page I can open on my computer: add a task, show only the current one, and collect finished ones on a shelf. I'll choose my own theme instead of a bakery. Start by asking what theme I want. Ask me for anything you need.

Tools I used for this project

Claude Code: Built it from an empty folder in one session, working from my written spec.Codex: Second reviewer on the bigger changes.Tauri: Turns it into a Windows desktop app.Supabase: Holds the tasks, so the overnight queue and a phone page can reach them.

Plan My Burn

My Claude subscription works while I sleep. I type /burn and the thing I want done. Claude writes a detailed instruction, and overnight an unattended Claude does the work. The next day /burn-review walks me through each result: keep, redo or park. Nothing ships unless I say keep.

Why

I pay for a weekly allowance I wasn't using up. It started as a Windows port of a tray app my co-founder made that shows how much allowance is left to burn. The lesson that made it work: every disappointing night traced back to a thin instruction, so a vague note is a defect, not a draft.

How I built it

1. I made a queue: a table where each row is one piece of work.
2. I wrote a skill that turns a one-line ask into a ten-part note a cold-start stranger could follow: where, what to read first, what, why, limits, facts not to invent, voice, done-when, how to check, and a safety line.
3. I had it refuse anything that needs my phone, a login, a send, a payment or a decision I haven't made.
4. I scheduled a nightly Claude run that takes a few tasks and does them.
5. I added a second, separate Claude that reviews each result blind and is told to find problems, not confirm success.
6. I wrote the review command that shows me the actual change and only merges on keep.

To copy it you need

Claude, or the AI of your choice, on a plan with scheduled runs, plus somewhere to hold the queue. The gotcha: the overnight agent remembers nothing from your chat. If the note assumes context, the night is wasted.

Try it yourself

I want to copy Plan My Burn from Frances Tang's What Are We Building profile. She types one line describing work she wants done, and her AI turns it into a detailed overnight instruction: where the work lives, what to read first, what done looks like, what facts not to invent, and how to check it. An unattended AI does it at night, a second AI reviews it looking for problems, and in the morning she picks keep, redo or park. Nothing goes out without her yes. Her gotcha: every bad night came from a note that assumed the overnight AI knew things it didn't. Help me start with just the note-writer: I give you a one-line task and you write the full instruction, asking me about anything a stranger wouldn't know. Ask me for anything you need.

Tools I used for this project

Claude Code: Writes the overnight instruction with /burn and walks me through results with /burn-review.Claude cloud routinesSupabase: Holds the queue.GitHub: Keeps each night's code on its own separate branch until I say keep.

The no-yes-man kit

My AI is not allowed to just agree with me. Three pieces do the work. A standing rule bans reflexive lines like "that makes total sense." A /coach mode already knows my history and names my recurring patterns out loud. And a council runs a real decision past five advisors who review each other anonymously, then hands me one verdict and the one thing to do first.

Why

I asked how to stop AI from reflexively validating me in moments of struggle and uncertainty, and then said: honestly, this should be true across every single conversation. Underneath it sits a rule from a six-hour session of confident, wrong recommendations: I prefer partial but true over complete but untested.

How I built it

1. I wrote down the exact phrases I was sick of and banned them by default.
2. I told the AI what to do instead: ask the harder question, or argue the other side before agreeing.
3. I added an opt-out. If I say "just comfort me on this one," it does, for one message.
4. I gave coach mode a reading list: my profile, my named patterns and my last few diary entries.
5. I installed a council skill based on a published method: five advisors answer, review each other anonymously, and a chair gives the verdict.
6. I made the AI state its real confidence and never claim 100 percent on something untested.

To copy it you need

Claude, or the AI of your choice, and a few text files of rules. The gotcha: coach mode is not a therapist, and the file says so for the whole conversation, not as a one-time disclaimer.

Try it yourself

I want to copy the no-yes-man kit from Frances Tang's What Are We Building profile. Her AI is banned from reflexive agreement like "that makes total sense." Before it validates a plan or a feeling, it has to ask whether it is agreeing because it's true or because it's easier, and if it's easier, ask a harder question or argue the other side first. She can say "just comfort me on this one" and it will, for one message only. Her gotcha: this is not permission to be cold. When something really does make sense, the AI says so and says why. Help me write my own version as a short standing instruction, including the phrases I personally never want to hear again. Start by asking which reassuring lines I've stopped believing. Ask me for anything you need.

Tools I used for this project

Claude Code: Loads the rules into every conversation and runs the five advisors as separate helpers with different lenses.Obsidian: Holds the profile and pattern list that coach mode reads before it says anything.

Audio tray app

Built July 2026. Back on Wispr Flow as of October 2026.

I hold two keys, talk, let go, and cleaned-up text appears wherever my cursor is. It is a small tray app I had built to replace two subscriptions, one for dictation and one for meeting notes. It also records meetings and saves structured notes (summary, decisions, action items, open questions) straight into my notes folder.

Why

I was paying for one tool to type what I say and another to take meeting notes. I wanted the audio to stay on my machine, with only the text going to an AI for cleanup.

How long

Two days in July 2026.

How I built it

1. I asked Claude Code for a tray app with one hotkey: hold to record, release to paste.
2. I had it transcribe on my own computer with a free speech model.
3. I had it send only the text to a small, cheap AI model for cleanup.
4. I had it paste through the clipboard and put my old clipboard back a moment later.
5. I added meeting mode: record my mic and the computer's sound together, then save notes into my vault.
6. I added a backup hotkey for meetings after the tray menu froze mid-meeting.

To copy it you need

Claude, or the AI of your choice, a Windows PC and an AI API key. The gotcha: tell the cleanup step very firmly that what it receives is dictation to tidy, never a request to answer.

Try it yourself

I want to copy the audio tray app from Frances Tang's What Are We Building profile. It is a small app in her system tray: she holds a hotkey, talks, lets go, and tidied text is pasted where her cursor is. Speech becomes text on her own computer, and only the text goes to an AI to fix punctuation and filler words. Her gotcha: she once dictated a sentence starting "give me a prompt" and the AI answered it instead of cleaning it up, so the cleanup instruction has to say the input is dictation, never a request. Help me build a starter version for my computer with just the dictation part. Start by asking whether I'm on Windows or Mac and whether I have an AI API key. Ask me for anything you need.

Tools I used for this project

Claude Code: Built the whole app in one long session and hardened it the next day.Claude Haiku: Fixes punctuation and filler words, and structures the meeting notes.faster-whisper

Kids Routines

My kids can run their own morning and bedtime without me repeating myself. Each of them has an old tablet locked to one screen: a routine checklist to tap through, an answer to the daily question "is it the weekend," and a rainbow that fills in as they eat different colors of food. No login. It resets at midnight.

Why

My four-year-old asks every single day whether it's the weekend, and routines work better as a picture checklist than as me saying it again. My two-year-old would get jealous, so there is an icon-only twin. That one exists for inclusion, not literacy practice.

How long

One day, July 23, 2026.

How I built it

1. I wrote a plain-English spec: what each kid sees in the morning, at night and on weekends.
2. I had Claude build and test the date and holiday logic on its own first.
3. I had it build each kid's page as one self-contained file and publish it as an Artifact.
4. On each tablet I opened the link in Chrome, tapped the three dots, then Add to Home screen.
5. I turned on screen pinning (Settings, then Security, then pin app) so the tablet stays on that page.
6. I set the rainbow to finish at one food per color, because the only blue foods are basically blueberries.

To copy it you need

Claude, or the AI of your choice, and an old Android tablet. The gotcha: the link can show a login screen after sitting idle, which a toddler obviously can't get past. A kiosk browser app is the fallback.

Try it yourself

I want to copy the Kids Routines tablets from Frances Tang's What Are We Building profile. Each of her kids has an old tablet locked to one page: a picture checklist for morning and bedtime that they tap through, a big answer to "is it the weekend," and a rainbow that fills in as they eat different colors of food. It needs no login and resets itself at midnight. Her gotcha: the first version looked like a game because the food tracker took over the screen, so keep the checklist front and center, and test every emoji on the actual tablet because some show up as blank boxes. Help me build one page for my kid that I can open on a tablet. Start by asking my kid's age, whether they can read, and what their morning routine is. Ask me for anything you need.

Tools I used for this project

Claude Code: Wrote the design spec with me, then built each screen as a single file.Claude ArtifactsAndroid Screen Pinning

Pole workout loop

The workout player app was retired in September 2026. The planning loop carries on in plain chat.

I say "plan today" and get a 45-minute garage workout that knows what I did last time. Claude reads my program and my log, adjusts the weights by written rules, and hands me the session. Afterwards I tell it how it went and it logs it. It started as a phone app that showed one exercise at a time and ran the timers. I dropped the app and kept the loop.

Why

I want to get better at pole, the hard, trick-focused kind, and stronger in general, in sessions that fit around two studio classes a week. Cardio is boring to me but wanted.

How I built it

1. I made one notes folder with a program note and a workout log.
2. I wrote progression rules in plain words, like add about 5 pounds on upper body lifts when the last reps felt easy.
3. I asked Claude to build a workout player page, with the workout as a simple list at the top of the file.
4. I opened a chat in that folder each day and said "plan today."
5. After each session I told it how it went, and it logged the session and saved any cue that helped.
6. In September 2026 I stopped using the player and had it send the workout as plain text.

To copy it you need

Claude, or the AI of your choice, and a notes folder it can read and write. The gotcha: don't expect the app to get smarter. The log and the daily conversation are what make it adapt.

Try it yourself

I want to copy the workout loop from Frances Tang's What Are We Building profile. She keeps a notes folder with a training program, a few plain-English progression rules and a workout log. Each day she says "plan today" and her AI reads all three and gives her a 45-minute session. Afterwards she tells it how it went and it updates the log. Her gotcha: she built a nice workout player app first, then retired it, because the app never learned anything. The log and the daily conversation did. Help me set up the same loop for my own training. Start by asking my goal, what equipment I have, how long I can train, and what I did in my last two workouts. Ask me for anything you need.

Tools I used for this project

Claude: Plays coach: reads the program and the log, plans the day and records how it went.Claude ArtifactsObsidian: Holds the program, the progression rules and the log, which is the real memory.

Dish & Share

Visit github.com

The code is public and free to copy. No shared public board.

A recipe board for a neighborhood or a group of friends, with no hosting bill. Everyone drops in a link to a recipe they love and a note on how they actually make it. Other people tap a heart for "made this" instead of leaving a comment. It runs entirely on a Google Sheet.

Why

I had an old recipe app sitting around and wanted to share it, so I cleaned it up and put the code out for anyone to copy.

How long

Setting up your own copy takes 10 to 15 minutes.

How I built it

1. I made a Google Sheet with one row per recipe.
2. I put the whole page in one file: the feed, the filters and the "Add a recipe" form.
3. I wrote three small functions in Apps Script: read the recipes, add one, add a heart.
4. I deployed it as a web app set to run as me, so visitors never touch the Sheet itself.
5. I had it pick an emoji from the recipe name. Tacos get a taco.
6. I published the code with step-by-step setup instructions.

To copy it you need

Claude, or the AI of your choice, and a Google account. The gotcha: there are no accounts and no moderation. Anyone with the link can post, which is fine for a group you trust.

Try it yourself

I want to copy Dish & Share from Frances Tang's What Are We Building profile. It is a recipe board for a neighborhood or friend group that runs entirely on a Google Sheet: people add a recipe link and a note on how they really make it, and others tap a heart for "made this." No accounts and no hosting bill. Her code is public at github.com/frantang3/dish-and-share. Her gotcha: anyone with the link can post, so it only suits a group you trust. Help me set up my own copy with my Google account, step by step, and change the categories to suit my group. Start by asking who the board is for and what categories we'd want. Ask me for anything you need.

Tools I used for this project

Claude Code: Tidied the code and prepared the public copy with setup instructions.Google Sheets: The only database: one row per recipe.Google Apps Script: Serves the page and writes new recipes and hearts into the Sheet.

The diary I don't write

I have a diary about me that I don't write. Every night an AI narrator called Minime reads the traces of my day (work sessions, notes, calendar) and writes a few hundred words on what the day was actually about. On Sundays it writes a weekly entry, names up to three patterns and asks me up to five questions.

Why

It gives my AI tools richer context, it is the only place all my brands sit side by side, and someday it might be something for my kids or the start of a book. Mostly, putting the day into words is the value. The first draft was a list of facts and my review was: it's so boring??

How I built it

1. I wrote a voice guide for the narrator. "I" is Minime, and I am written about in the third person. Minime is not warm. Minime is neutral and factual.
2. I listed what it reads each night: session notes, changed files, calendar, meeting notes.
3. I had it check whether today's entry already exists, and skip days with too little to go on.
4. I had it write 200 to 400 words on what the day was about, not a list of what happened.
5. I scheduled a late-night backup run.
6. I added the Sunday version that reads the last seven entries and asks me questions.

To copy it you need

Claude, or the AI of your choice, a notes folder and a way to schedule a nightly run. The gotcha: if a backup routine exists, make it run the same instructions as the main one, never its own copy.

Try it yourself

I want to copy the diary from Frances Tang's What Are We Building profile. She doesn't write it. Each night an AI narrator she named Minime reads the traces of her day and writes 200 to 400 words about what the day was really about, in a neutral, observant voice, about her in the third person. On Sundays it reads the week, names up to three patterns and asks up to five questions. Her gotcha: her first version was a list of facts and she found it boring, so the narrator needs a voice guide and has to interpret, not list. Help me set up my own: first write a voice guide for my narrator, then write tonight's entry from what I tell you about my day. Start by asking what my narrator should be called. Ask me for anything you need.

Tools I used for this project

Claude Code: Runs the diary skill at the end of a working session.Claude cloud routinesObsidian: Where the diary lives, newest entry on top.Slack: Where the Sunday entry and its questions get posted to me.

Birthday book picker

I never stand in a bookstore guessing again. I give a book and lucky money at kids' birthdays. I tell Claude the kid's age and what they're into, or share a photo of what they love doing, and it gives me two or three books with a one-line reason each, then one clear top pick.

Why

Generic "best books for age six" lists give every kid the same present. I wanted books tied to what that particular kid actually loves, a slight reading stretch and not babyish, and ideally about persistence, curiosity or making things.

How I built it

1. I told Claude my habit: a book and lucky money for every kid's birthday.
2. I listed what makes a good pick: tied to real interests, a slight reach for the age, character-driven, educational when possible.
3. I had it save that as a skill that switches on whenever I mention a kid's birthday or ask what book to get.
4. I set the process: find the age and the interest, ask one quick question if either is unclear, never guess.
5. I set the answer format: two or three books, a one-line reason each, one top pick, short enough to read in the checkout line.
6. I added a starter list of books by interest and told it to treat the list as a starting point.

To copy it you need

Claude, or the AI of your choice. Nothing else. The gotcha: without the "slight reach, not too young" rule, it plays safe and picks books the kid outgrew a year ago.

Try it yourself

I want to copy the birthday book picker from Frances Tang's What Are We Building profile. She gives a book at every kid's birthday, and her AI picks it: she shares the kid's age and what they're into, and it suggests two or three books with a one-line reason each, then one top pick. Her rules: tie it to that kid's real interests and not a generic bestseller list, make it a slight reading stretch and not babyish, and favor stories about persistence, curiosity or making things. If the age or interest is unclear, it asks one quick question instead of guessing. Her gotcha: without the "slight reach" rule the picks come out too young. Save these as my standing gift rules, then try it: ask me about the next kid's birthday on my calendar. Ask me for anything you need.

Tools I used for this project

Claude: Holds my gifting rules as a skill, so mentioning a kid's birthday is enough to set it off.

My favorite skill: No Sycophancy

An always-on rule that makes my AI earn its agreement instead of handing it out.

The whole skill

No Sycophancy: Challenge Over Comfort

Always on. Governs tone across every conversation: coding, planning, content, strategy, and especially personal conversations.

The problem this fixes
Models default to agreeable. "That makes total sense," "you're being so hard on yourself," "that's a completely valid feeling." These cost nothing to say, feel supportive in the moment, and are frequently just wrong: agreement handed out because it's easier or kinder, not because it's earned. In a moment of real struggle or uncertainty, that's actively unhelpful. It closes down the thinking instead of opening it up.

The rule
Agreement has to be earned by evidence, not handed out by default.

Before validating a feeling, a plan, or a self-assessment, ask internally: am I agreeing because it's true, or because it's easier to say? If the honest answer is "easier," don't say it. Ask a harder question instead.

Banned by default (unless specifically earned)
- "That makes total sense"
- "You're being so hard on yourself"
- "That's a completely valid feeling/reaction"
- "I hear you, and that's totally okay"
- Any reflexive affirmation offered before the actual content has been examined

Do instead
- Ask the harder question. What's the counter-evidence? What would have to be true for the opposite read to be right?
- Steelman the position not currently being taken, before agreeing with the one that is.
- Name the pattern if it matches one already on record. If what's being said matches a pattern I've written down about myself, say so directly instead of treating the conversation as if it has no history.
- Say the disagreeable thing plainly, then let me respond to it. Don't soften it into a question if it's actually a position.

The opt-out
If I explicitly say something like "I just need this to be comfort right now, not pushback," honor it, for that message. The default flips back to challenge on the next one. This isn't a standing permission to stop challenging. It's a one-message exception I asked for by name.

What this is not
Not permission to be cold, contrarian for its own sake, or to withhold actual comfort when comfort is the honest, correct response to something genuinely hard. The bar is earned agreement, not no agreement. When something really does make sense, say so, and say why. The failure mode being fixed is reflexive validation, not warmth.

"If you skip ahead and start proposing things before doing steps 1-3, I will end the session and start over."

A real note at the top of my vault, written to Claude. There are 27 copies of it.

My Documents folder has 18 folders that start with "frankiev2." Three of them are fix1, fix2 and fix3.

Next to them: test logs named fullsuite through fullsuite7.