Weekly update
CertifiedTurns a week of work into six lines your lead actually reads.
1.2 KB
How you do anything is how your AI does everything.
A Chrome side panel that builds and maintains a personal context file for use with AI assistants. Answer some questions. Get a file. Hand it to whatever AI you already use.
Try it in 60 seconds
The Mirror Test
What does your AI chat service say if you prompt it like this? Probably a thin, partially right sketch of things you have done in that tool.
When your service reads from your perspective and writes from your perspective, how much of you is in your perspective?
Polished outputs are no longer scarce; the polish is becoming the new normal. Outputs that are recognizably yours are growing increasingly scarce.
Paste into the AI you already use
ChatGPTClaudeGeminiCopilotPerplexity— any of them
Based on everything you know about me, describe me: my role, my goals, how I like to work, how I write, and what I care about. Then list what you would need to know to write something that sounds exactly like me, and be specific about what you’re guessing.
Nothing is sent by this page. The prompt goes on your clipboard and the answer stays between you and your assistant.
It is not wrong
For most people the answer is broadly true and completely thin — a partially correct sketch assembled from whatever you happened to type in that tool. Being not-wrong is a low bar for something writing on your behalf.
It knows the tool, not you
What it has is your history inside one product. Move to another assistant and you start over, because none of it was ever yours to take.
The second half is the useful half
The list of what it would need to know is a specification for your own context file — written by the thing that will read it.
Workbrainby Model Citizen
Then answer it properly, once.
That is the whole product. A guided interview turns the second half of that answer into a plain markdown file you own, and you hand it to whichever assistant you are using. When your service reads from your perspective and writes from your perspective, how much of you is actually in your perspective?
How it works
Answer questions. Get a file. Use it everywhere.
Workbrain is a Chrome side panel. It interviews you once about how you actually work, writes what you said into a plain-text file you own, and hands that file to whichever AI you already use. The whole of that is free and always will be. What you can buy is help keeping it current.
-
Answer questions
About an hour, in the panel, in your own words. No account, nothing sent anywhere.
-
Get a file
Plain markdown on your machine. Readable, editable, and yours to take anywhere.
-
Hand it to any AI
Paste it, or let the panel put it in the composer for you. It works the same in all of them.
-
Prove it, then prune
Run it against the baseline the app captured before you had a file, and cut what earns nothing.
That fourth step has its own page.
The Proving Grounds runs the head-to-head on a full screen — your ask answered with the file and without it, side by side, against the baseline the app captured on your first day. Nothing you paste there leaves the tab.
What you get now, and what you can add
The file is the product and it is free. The subscription does not unlock it — it keeps it from going stale, which is the thing that actually happens to a file somebody wrote once and never opened again.
Free
Right now$0forever
Everything it takes to have a context file and use it. No account, no trial, no card.
The full interview — every module, in your words
Context.md and Skills.md, yours to download and edit
Hand-off to any assistant, by paste or into the composer
The Proving Grounds — your file against your own baseline
Nothing transmitted. It runs on your machine and keeps running offline
Workbrain+
Optional$10/ month
For keeping a file true after the week you wrote it. Everything in Free, and then:
Question packs — deeper sets for a trade, a role or a situation the base interview does not reach
Staying fresh — it notices what has gone stale and asks you about that, rather than everything
The nudge — a few minutes a week, however suits you, instead of another hour in one sitting
Agents — name a job you repeat, give it your file and your bar, and keep it
There is a third rung, Workbrain Pro, for running a team of agents against your connected accounts. It is not built yet — the says what is live and what is not.
Not sure where to start?
Some people want to be taught it, in a room.
The hardest part is not the software. It is knowing what is worth writing down about your own work — and that is much easier with somebody asking you the questions and a few other people in the room doing it at the same time.
The Workbrain Accelerator is four weeks, ninety minutes a week. You leave with a file you built and can defend, not a certificate.
What is actually in the file
What is in the file
Open a section to see the kind of line it holds.
Why it goes stale
Different answers rot at different speeds. A job title lasts years; this quarter's goal lasts a quarter. Every answer carries the date you gave it and a half-life, so the panel can tell you which parts are still true instead of asking you to re-read the whole file.
Current — answered recently enough to trust
Drifting — past its half-life, worth a look
Unanswered — never filled in, and that is fine
Working Paper · September 2026
Durable Context
How LLM assistants consume what you tell them, and what reliable skill execution actually requires.
This is a working paper, and the name is the point. It is revised as the vendor documentation changes and as runs come back — the measurement section reserves its seat rather than claiming a result. Read it as the current state of the argument, dated, not as a finished thing.
Adam Young — Model Citizen. Vendor documentation surveyed and cited as of 9 September 2026.
Abstract
Every mainstream LLM assistant now offers some way to make context durable: memory that fills itself, projects that hold instructions, knowledge files, connectors, custom agents. Users reasonably assume these surfaces behave like configuration — set once, applied consistently. This paper examines the published documentation for five widely deployed services (Anthropic Claude, OpenAI ChatGPT, Google Gemini, Microsoft 365 Copilot, and Perplexity) and finds that assumption unsupported.
Two findings carry the argument. First, only one of the five documents what happens to a user's context when a conversation exceeds the model's window; the other four document it nowhere, leaving users unable to know whether their earlier instructions were summarized, truncated, or dropped. Second, across all five vendors combined, exactly three statements of instruction precedence exist. On four of five services, a user cannot determine from documentation which of their own durable surfaces wins when two of them disagree.
The conclusion is not that these systems are unusable. It is that reliability cannot be obtained from inside the chat window, because the chat window is not where the durable part lives. Context has to be treated as infrastructure — written down once, versioned, owned, and portable across the tools that consume it.
Contents
- Everyone is rebuilding the same foundation every morning
- Four surfaces people mistake for one
- What the vendors actually document
- The window got bigger. Your odds did not.
- The same question does not get the same answer
- What determinism you can actually buy
- Context as infrastructure, not input
- A reference model
- How you would know it worked
- What this evidence does not support
- A five-minute test
- References · Disclosure · Appendix A — vendor comparison, gaps shown as gaps
Library
Skills you can drop into your file.
Every skill is markdown, and every skill states its own objective — which is what the Proving Grounds later scores it against. Data, never code.
Ticket triage
Sorts by blast radius, not by who shouted loudest.
0.9 KB
Meeting recap
Decisions, owners, dates. Nothing else survives the cut.
0.7 KB
Exec brief
CertifiedOne page, one ask, one number worth arguing about.
1.1 KB
Hiring screen
The questions that separate the top two candidates.
1.4 KB
Incident writeup
What broke, what it cost, and what changes on Monday.
1.3 KB
The Accelerator
Want it taught in a room?
Four weeks, up to ten people, free to everyone who attends. The host provides the room; Model Citizen teaches.
Proving Grounds
Two things worth proving.
A context file is proved against the baseline the app captured before you had one. A skill is proved against the objective you wrote for it. Most people come back for both.
1
Your captured baseline
The app captured this on your first day — the ask and the answer, before the file existed. Bring it here as one paste.
In the app: Context Development → Take it to the Proving Grounds
2
Your file now
Arrives with your bundle — everything on this page stays in this tab and leaves when it closes.
1
Pick a skill from your file
Read from Skills.md · 3 of 9 skills shown
2
The objective you stated
Captured while you built the skill. This is the bar the run is scored against — not a generic idea of a good answer.
3
What the skill produced
1.4k tok · 3.1s
4
Scored against your objective
Edit the skill
FIX
Objective scoring is the model judging its own output against your words. Directional, and it moves when the skill genuinely improves.
A full pass: the file, then every skill on it
One queue. The context file goes first, against your captured baseline, because a skill that runs on a thin file fails for the file's reasons rather than its own. Then each skill is scored against the objective you wrote for it.
Context.md
vs baseline captured 12 Aug
+7 specifics
Weekly update
vs its stated objective
2 of 3 met
Ticket triage
vs its stated objective
1 of 3 met
Exec brief
no objective recorded — add one to score it
4 runs · about 40 seconds
File
+7specifics
Against the 12 Aug baseline. Two sections still contribute nothing.
Skills
3of 6 criteria met
Ticket triage is the weak one — its objective asks for something the file cannot supply yet.
Do this first
Name your two biggest clients in the file. It closes the file's largest gap and the triage skill's failing criterion at once.
What leaves your browser
On this page: nothing. Your bundle lives in this tab and dies with it; you carry the prompt to your own AI yourself and bring the answer back the same way. A hosted run — the prompt to a named model, discarded on response — is a possible later door, and it will be a door, never a default.
Workbrain+
The free thing stays free.
Building your file, proving it, and handing it to any assistant costs nothing, forever, with no account. Membership is for managing a Workbrain over years rather than an afternoon.
Free
$0forever
The install is the sign-up. No email, no identity.
The full interview and both files
Freshness tracking and drift verdicts
Your captured baseline, and the copy-paste proof loop
The whole public library
Workbrain+
Membership$10/ month
Identity lives here on the site. The extension takes a pasted member key — no accounts in the app itself.
Private skills library — add your own and other people's skills to your profile instantly
Unlimited hosted proving — full passes across the file and every skill, with run history
Evergreen monitoring — drift nudges arrive here by email, never as extension notifications
Sync across devices — by your own hand or inside the browser's own quota
Every paid feature is an enhancement over a manual path that keeps working on a plane.
Workbrain Pro
In development$100/ month
Context and skills plus guardrails make a role. Pro runs a team of them — connected to your accounts, with one lead you talk to that hands out the work and reports back.
Everything in Workbrain+
Roles — context, skills and guardrails combined into a dependable assistant for one part of your work
A team of roles, coordinated by one lead
Named integrations, with human approval on anything risky
Not built yet. The roadmap says what is live and what is not.
Taught in a room
Two rungs somebody else pays for
Everything above is bought by a person, monthly. These are bought once, by whoever convenes the room — and they are free to everyone in it.
Workbrain Accelerator
$5,000 · per cohortFour weeks, up to ten people, ninety minutes a week. The host provides the room and the people; Model Citizen teaches. Free to attend.
Workbrain Accelerator Pro
from $25,000The same three strands run privately in one week for a team of up to twenty. Paid by the business, or through a training grant it applies for.
Money never enters the extension. Purchases, scheduling and service requests all live here on the site; the panel deep-links out and takes a pasted key.
Workbrain Accelerator
Four weeks. One goal. A file that is theirs.
A short class in working with AI, where the thing everybody leaves with is their own context file — built in the room, defended out loud, and theirs to keep. Two ways to run it, depending on who is paying and who is in the room.
Workbrain Accelerator
Hosted$5,000per cohort
An organisation convenes the room and pays once. Everyone in it attends free. Schools, libraries, workforce boards, community programmes.
Four weeks, ninety minutes a week — the same twenty-eight day arc
Up to ten people per cohort
Free to attend — the host pays, the room does not
Often funded by a grant the host already holds
12 months of Workbrain+ for anyone who earns Certified
Everything below — the gates, the settings, the paperwork — describes this track.
Accelerator Pro
For a businessfrom $25,000
The same three strands and the same twenty-eight day arc, with the teaching delivered in consecutive days rather than weekly evenings. A business pays for its own team, on its own schedule.
A full day to several, depending on the size of the team
Up to twenty people, one team
Private — your work, your examples, under NDA if you need it
Payable by the business, or through a training grant it applies for
12 months of Workbrain+ for anyone who earns Certified
Talk about Pro
Scoped per engagement. Tell us the team and the week and we will come back with a number.
How a cohort actually runs
Both tracks teach the same three strands. The difference is the room, the length and who paid — not the material.
The shape
Twenty-eight days · either format
Everyone arrives on day one with a goal of their own. They name it, build their file around it, put it to work, and plan what comes next. The arc is twenty-eight days in both tracks — what changes is how the contact time is spent.
Community — four weekly sessions of ninety minutes, up to ten people
Business — a full day to several consecutive days, sized to the team, up to twenty
Either way the measurement lands on day twenty-eight, because that is when there is something to measure — a file that has been used for a few weeks, not one that was finished in the room.
Three strands, every session
Each week goes deeper
Business fundamentals, so the goal is worth something. AI foundations, so people understand what the tools can and cannot do. And Workbrain, where each lesson becomes answers in each person’s own file.
Day twenty-eight
Measured, not assumed
That day-one goal is measured against where they started. Everyone leaves with a completed context file, a thirty-day plan to keep it improving, and proof — their own work with and without the file, side by side.
Hosting a community cohort
Everything from here is the detail of the hosted community track — the gates, the settings, the credentials and the paperwork. A business cohort runs the same twenty-eight day arc and the same three strands; what differs is that the room is one team, the contact time is consecutive days, and the company is both host and payer — so most of what follows is handled inside the engagement rather than negotiated with a venue.
The agreement, plainly
What you're agreeing to host.
| Who is in the room | Adults, or students with guardian consent. Cohort of up to 10. |
|---|---|
| How long | Four sessions of 90 minutes, weekly. Then a weekly board night, ongoing and optional to attend. |
| The board night | Run for you, not for one group. Every cohort you have hosted is welcome in the same room, week after week. |
| What participants do | Answer questions about their own work, write plain text files, use an AI assistant they already have. |
| What they install | Nothing required. An optional free Chrome extension. No account, no signup. |
| What data the tool collects | None. It runs in the browser and transmits nothing. |
| What we collect | Name and contact at enrollment, for scheduling and the completion record. Nothing else. |
| What leaves the building | Nothing generated in the sessions. Participants keep their own files. |
| What we need from you | A room with power and wifi, the same night each time, and a screen. |
| What we provide | Facilitation, curriculum, printed materials, loaner devices, certificates. |
| Insurance | General liability carried by Model Citizen. Certificate on request. |
| Background checks | Completed and current for all facilitators. Documentation on request. |
| What we ask of participants | Attendance and a short voluntary survey. |
| What we never do | Sell to participants during sessions, collect payment in the room, or require an email address for anyone to receive their own files. |
| If the cohort can't fill | Under 6 confirmed a week out, the cohort is rescheduled. Never billed as a partial cohort. |
| If you need to cancel | Full refund of any deposit before device and printing commitments are made, roughly two weeks out. Case-by-case after that. |
| A second cohort | No discount yet — priced the same as the first, revisited as real per-cohort costs are established. |
We do not sell to participants. Not during sessions, not after. The program is free to attend and there is nothing to buy at the end of it.
Two columns, no fine print
What each side provides.
You provide
- A room with power and wifi
- The same night each week
- A screen to teach from
- Promotion through your channels
- The audience you convene
We provide
- The facilitator, background-checked and insured
- The curriculum and printed materials
- Loaner devices for anyone without one
- Certificates and the credential record
- Reporting at 90 and 180 days
The program at a glance
Five gates. Six contact hours. Four weeks.
Four teaching sessions plus a weekly board night, for cohorts of up to 10. Each session produces one artifact the participant keeps.
-
1
Session 1 · 90 minutes
Baseline.md — one real task, run cold, and the answer it got, saved
-
2
Session 2 · 90 minutes
Context.md — who they are, what they do, what they want next
-
3
Session 3 · 90 minutes
Skills.md — one repeated task, written as a method
-
4
Session 4 · 90 minutes
Skills.md, the side-by-side proof, and the Certificate of Completion
-
5
Board night · weekly, optional
Work Brain Certified — earned by returning and showing what they built since
Three ways it runs
After school, over a summer, or on library evenings.
After-school
Students build files about school projects, first jobs, and applications. Guardian consent comes first, the directory stays off by default, and every file leaves with the student.
Summer
The same gates on a faster rhythm, inside a youth employment or summer program — baseline to proof lands within one season.
Community adults
Evenings at the library or community center. Job seekers, small-business owners, career changers — the program meets each person at their own work.
What participants leave with
Three files they own, and a credential.
Baseline.md
Context.md
Skills.md
Certificate of Completion
Issued at Session 4, for finishing the four sessions and producing the files. Attendance plus evidence.
Work Brain Certified
Earned later, at a board night, by showing what you built since. Demonstrated use — that is what keeps the first tier from being a participation ribbon.
Included
12 months of Workbrain+
Added automatically to the file they built in the training — a year of question packs, the nudge and the staleness checks, at no cost. The credential says the file was real; the year is what keeps it that way.
Outcomes
What we measure, and what we refuse to claim.
We measure what we can actually attribute. Participants complete a task before any teaching and a comparable task at the end, and we record how much context they had to type, how many exchanges it took, and whether they used the result. We also track how many complete each stage and how many return.
We do not claim employment outcomes. Participants report what happens next and we pass that on as reported, not as program effect. Assistant outputs also vary between runs for reasons unrelated to any program, so we report a group of measures rather than one headline number.
Privacy and data handling
Nothing to hand over, by design.
The tool runs entirely in the browser. There is no account, no server, and nothing is transmitted. Files are plain markdown, written to the participant's own device. Participants who never come back keep everything they made.
Sessions do teach participants to inspect what their AI assistant has already stored about them — several services retain inferred personal details by default. That is done privately, on each person's own screen. No one is asked to read anything aloud.
Compliance
The paperwork, on request.
| Insurance | General liability carried by Model Citizen. Certificate on request. |
|---|---|
| Background checks | Completed and current for all facilitators. Documentation on request. |
| Under-18 participants | Guardian consent covers enrollment, follow-up contact, and reporting choices. The public directory defaults to off. |
| Facilitator credentials | A named, background-checked facilitator per cohort, introduced before the first session. |
| Participant records | We keep your name, contact details and completion record for as long as the credential is valid, so we can verify it if someone asks. You can ask us to delete it at any time. |
| Participant files | Never stored by us. We never have them — the tool transmits nothing. |
What it costs
Free to attend is not free to run.
A free program with no numbers reads as unserious, so here are the numbers.
$5,000One funded cohort, up to 10 people
$0Attendance is free, always.
| Facilitation and delivery | Four 90-minute sessions, the weekly board night, and follow-up reporting at 90 and 180 days. | Our rate card |
|---|---|---|
| Materials and the credential | Printed materials, and a certificate in both physical and shareable digital form. | Actual cost, per participant |
| Program overhead | Liability insurance, upkeep and replacement of the loaner laptops, and the participant records. | Actual, allocated per cohort |
| Not billed | Curriculum development and program administration, donated by Model Citizen. |
A funded cohort is paid by the hosting organization or its sponsors. A full line-item breakdown, with the figure behind each row, is available in the program packet on request. Attendance is free, always.
Evidence
Read before you decide.
Curriculum summary
Session by session, artifact by artifact. It ships in the host packet — request a cohort and it comes with our reply.Facilitator background
Background-checked, insured, and named before the cohort starts. Documentation on request.Start a cohort
Request a cohort.
Tell us who you would convene
After-school, a summer program, or community adults — the cohort is shaped around the room you can fill.
We reply within two business days with the program packet: the session-by-session summary, the host agreement, and available dates. Nothing about your community is required to ask.
The writing
Eight arguments, in three series.
Each one is a deck you scroll rather than an article you skim — built around something you can actually operate, because a claim you can test beats a claim you have to take. Every piece also reads as plain text if you would rather have it that way.
Series one
The Slop Trilogy
What slop actually became, what the trade cost, and why the machine reverts to the mean.
The evolution of AI slopIt arrived as a slur about laziness. It is now a condition, and the tell is not quality — it is absence.
14 actsThe convergence · 48 outputs
The Blind CompromiseEvery compromise you consciously made came with a moment. This one did not, so it just accrues.
11 actsThe ledger · both columns
Super Markov BrothersA real order-2 chain over two corpora. It is not biased toward the average — it IS the average.
11 actsThe chain · same seed, two corpora
Series two
The Amplifier
Three metaphors for the same mechanism: the tool multiplies what you bring to it, and brings nothing of its own.
New dog, better tricksEvery service is a remarkable animal nobody was taught the commands for.
11 actsThe trick deck · five dogs
Slop shotsEverybody got the magic cue. Then a few players started calling their shots in three rails.
14 actsThe bumper count · a real table
Cesar Millan was training peopleThe dog changes in ten minutes. The owner takes longer, and the owner is the one getting the instruction.
11 actsTwo owners · and the too-late slider
Series three
Fundamentals
The unglamorous part. Doing the small things the same way every time, before anything clever.
A-Iron ChefThe same basket, every time. What a fixed set of ingredients says about working from what you actually have.
Not written yet
Wooden’s shoelacesHe started every season by teaching how to put on socks. The small thing done right, before anything else.
Not written yet
The long one
Durable Context
The argument made properly, with citations and a measurement section that reserves a seat for results rather than claiming them. It is what the Accelerator and the Civic work both cite.
Start with the one-minute version
Every piece ends in the same place, so you may as well start there. Ask the assistant you already use to describe you, then to list what it would need to know to sound exactly like you.
Download
Two minutes to a file.
Install the side panel, answer the first module, and download Context.md. There is nothing to sign up for.
Add to Chrome — free
Chrome 114+ · also Edge and Brave
01
Install and open the panel
It lives in Chrome's side panel, beside whatever you are already working in.
02
Answer the first module
Four questions about who you are. The app captures your baseline on the way in.
03
Download and hand it over
Paste it into your assistant, or keep it in a project so every chat starts there.
The Chrome listing discloses no data collection, because there is none to disclose. Your file stays in your browser's own storage.
Privacy & guardrails
The promise, and its one amendment.
Never transmit the person's content anywhere. Not to us, not to their AI without them seeing the exact text first, not to a pack publisher.
What stays
Every answer you give, in your browser's own storage
Both files, generated locally, downloaded by you
Your captured baseline, and every proof run's history
What can leave, and only if you press it
A proof run: your file or skill plus the task, to a named model, discarded on response
Feedback you choose to write and send
Nothing else. Ever, and not quietly.
The amendment: hosted proving
A hosted test environment can transmit the file, which the guardrail above forbids. Rather than make that a quiet exception somebody finds in a network tab, it is written down here.
- What is sent
- The prompt — your file or a single skill, plus a task. No identifiers, no session, no answer store.
- Where it goes
- Through us to a named model, and discarded on response.
- Is it a default
- No. Copy-and-paste stays, and stays first. This is a door.
- Hidden prompts
- None. If extra instructions exist, you read them before they send.
One thing we will not claim: that a model can report your environment, location or device. It sees the prompt. Anything else it says about your machine is invention, and tuning a file on invention is worse than not tuning it.
Roadmap
What's live, what's being built.
A short honest list beats a page full of shells. Rough timing, updated as things land. Last updated September 2026.
Live
The app
The side panel builds Context.md and Skills.md today. The Chrome Web Store listing is being prepared; the download page tracks it.
Live
Proving Grounds — Context mode
Bring the bundle from the app and run the head-to-head in your own AI, on this site, with nothing transmitted.
Live
The civic route’s program page
The program, the agreement, the numbers — the page a host or funder reads first.
In build
Civic host kit
The cohort request flow, the downloadable program packet, and a page per cohort location.
Next
Library starter set
Six to ten skills you can read before you keep. The Library returns to the nav when it is real.
Next
Proving Grounds — Skills mode
Run a skill against the objective you wrote for it, on the site.
Planned
Workbrain+
The paid membership. Priced when it exists, not before.
Planned
Credential verification
A page per credential ID, plus the alumni directory for graduates who opt in.
The standard
Good, defined.
Every AI program promises "better." Almost none will tell you what good is, in terms you could test. Here is our definition — measurable, public, and open to attack. If it's wrong somewhere, the fix is an edit to a file you own.
The ladder
Three rungs. Each one testable.
1
Started
The file exists, and a fresh assistant reading it knows who you are. Everyone reaches this on day one.
2
Good
Passes the five-minute test below — one run, four checks, no judgment call anywhere in it.
3
Great
Holds up across repetition, rewording, and time — the three forces that actually break these tools.
The bar
Good is the five-minute test.
Open a new chat. Paste your file. Ask for the thing you make or ask most often — and type nothing else. Your file is Good when all four hold:
The answer knows who is asking, and who it's for.
It uses your names for things — your systems, your teams, your words.
It arrives in the shape you would have corrected it into.
You typed zero words of setup beyond the ask itself.
That's the whole test. Most files in the wild — including most "custom instructions" — fail the second or third check. Yours can pass it this week.
The summit
Great is what survives.
Good is one run on one day. Great is what stays true after that:
Reword it ten ways, same quality. Written methods hold; improvised prompts drift with every retyping.
Proven against your before. The same ask, captured without the file — and the with-file run beats it, on record.
Knows its own age. The file carries its date and tells your assistant to flag what looks stale.
Short on purpose. The smallest set that works beats the pile — the research is unambiguous about this.
Works without you in the room. A stranger's assistant could act on it — names defined, formats shown, methods followable.
Every repeated job has a recipe. And every recipe states its objective first — what a good result lets you do — so the output can be scored, not just admired.
And underneath all six: the file is yours — plain text, on your device, readable by whatever assistant is in front of you. That line is on the list because most programs fail it first.
The contents
What a great file holds.
Seven things, each earning its place — plus the file's own rule for how an assistant should read it, and its date.
Who you are
Who your work is for
The systems you work in
Your names for things
What's already settled
Where you're headed
One real example, marked "like this"
Notice what's missing: everything else. A great file is defined as much by what it leaves out — irrelevant context doesn't just waste space, it measurably degrades answers.
The claims
Disagree? Show your measurement.
Six positions this standard stands on. Each one is falsifiable — that's what makes it a standard instead of a slogan.
- A file that has never been tested against a before is a belief, not a tool.
- A definition of good with no measurement attached is not a definition.
- More context is not better context. Curated beats complete.
- A method that lives only in your head changes every time you type it.
- Anything you can't export, edit, and carry is a vendor's feature, not your file.
- A credential for showing up certifies attendance. Certification should require evidence of use.
And if someone finds a real flaw in this standard? Your file is plain text. The fix is an edit you can make this afternoon. Critique makes these files better — that is the point of owning them.
Where to take this