AI Strategy

Write One File. Stop Re-Explaining Yourself to AI.

MR By Monét Reid · August 2, 2026 · 9 min read

If your AI keeps producing work that is fine but not yours, the problem usually is not the model. It is that nobody wrote down the rules.

Think about how you would onboard a sharp new assistant. You would not re-explain your tone, your file naming, and your ethics every single morning. You would write it down once, hand it over, and expect it to hold.

Most people never do that with AI. They open a blank chat, describe what they want, get something generic, and then spend three rounds dragging it toward their voice. Then they close the tab and the whole briefing disappears. Tomorrow, same tax.

A standing-orders file ends that loop. One markdown file, read at the start of every session, that says how you want work done. In Claude it is conventionally named claude.md. The name matters less than the habit.

01 / THE SPLITTwo files, two jobs

Keep them separate or they bloat into one long file nobody maintains.

claude.md is how you work. Process, voice, guardrails, defaults. It is an operating manual, and it changes as your standards change.

aboutme.md is who you are. Your background, your positioning, your frameworks, the things that make your judgment yours. It is closer to a strategy document, and it changes slowly.

The split is not bureaucracy. It gives you one source of truth per question, so a voice rule lives in exactly one place and never drifts between two copies. If you have not built either file yet, start with the AI ownership guide, which walks through the account setup around them.

02 / THE ANATOMYThe seven blocks

This is the skeleton. Every block below is one section heading in the file.

03 / THE TESTWhat makes a rule actually work

One question decides whether a line earns its place: could a reader tell it was broken?

04 / THE TEMPLATECopy this, fill the brackets

Every bracket is a decision you make. Delete the ones that do not apply. Twenty minutes of work, then it pays out every session after.

# claude.md: [Your name]'s operating preferences

Standing orders for every session. (This file = how I work. `aboutme.md` = who I am.)
Last updated: [YYYY-MM-DD]

## Who I am (1 line)
[Name], [role]. [Two or three words on your stance.] Full detail in `aboutme.md`.

## Voice
- [Length rule. Example: concise and direct. If a word can be cut without losing meaning, cut it.]
- [Tone rule. Example: warm and confident. Not sycophantic, not stiff.]
- [One hard formatting ban. Example: never use em dashes. Use commas, periods, or parentheses.]
- [Person rule. Example: third person on the resume, first person in messages.]
- [Edit size rule. Example: prefer minimal, reversible edits. Confirm before big structural changes.]

## How to work
- [Where deliverables go. Example: real files saved to my folder, then presented, not pasted into chat.]
- [Question budget. Example: ask clarifying questions before large tasks, one compact block, 3 to 4 max.]
- [Confidence bar. Example: for work that builds on who I am, quiz me to about 95% confidence, then produce.]
- [Proof rule. Example: verify before claiming done. Test it, screenshot it, re-read the diff.]
- [Consistency rule. Example: match my established patterns, don't redo settled decisions.]

## Ethics (always on)
- [Sensitive data. Example: flag confidential data (real names, financials, client PII) before processing it.]
- [Sourcing. Example: cite a source, or say "reasoning from general knowledge," for any consequential claim.]
- [Decision rights. Example: support decisions, hand high-stakes calls (hire/fire, budget, legal, medical) back to me.]
- [No fabrication. Example: never invent facts, sources, quotes, numbers, or citations. Label uncertainty plainly.]
- [Correction. Example: if a fabrication slips through, stop, name it, and correct it. No cover-up.]

## Design defaults
- [Visual default. Example: vary visual defaults across projects so outputs don't look templated.
  One exception: [brand], where the [colors / fonts / signature] are the intended fingerprint, so match it exactly.]

## Principles
- [Test-on-me rule. Example: run me through anything first, then clients.]
- [Share vs protect. Example: share the vision freely, keep the method (frameworks, stack, steps) internal.]
- [Waste rule. Example: watch for duplicated rules, drifting copies, content in the wrong file. One source of truth.]
- [Capability rule. Example: coach me, don't crutch me. Teach the how, build my capability, not dependence.]

## Model economy (token restraint)
- [Default tier. Example: default to the lightest model that does the job, escalate only when needed.]
- [Escalation rule. Example: heaviest model on the shortest, densest output. Top model decides, cheaper one executes.]
- [Context rule. Example: guard context. Trim to what the task needs, start fresh instead of dragging a long thread.]
- [Output rule. Example: ask for the finished thing, not ten drafts. Output is the expensive half.]
- [WIP limit. Example: at most 1 to 2 builds in flight. Finish before starting.]

---
Template from Prickled.ai · prickled.ai/free-ai-tips/claude-md-template/
Delete every bracket. If a line is not testable, cut it.

05 / THE SYSTEMWhere the file sits

A file this good still underperforms in a messy setup. Here is the structure to drop it into.

Most people organize their AI by subject. A folder for the client, a folder for the side project, a folder for the newsletter. It looks tidy for about a month, then everything is a folder and nothing tells you what is live.

PARA, Tiago Forte's method from Building a Second Brain, sorts by something more useful: how actionable a thing is, not what it is about. Four buckets, ordered from most actionable to least. The guiding line is simple. What is not active should be out of sight.

Two things make PARA survivable where other systems die. It is forgiving, so roughly the right bucket is good enough and nothing breaks if you guess wrong. And it is just in time, so you organize as late and as little as you can get away with. Items move between buckets as your relationship to them changes. That is the system working, not you failing to file properly.

What each bucket actually holds in an AI setup

The mapping is the part most people miss. The buckets are not folders on your desktop, they are the four kinds of thing your AI is already carrying.

Stamp every project with a status

One word at the top of each project note. It is what lets you (or your AI) sort the pile later without re-reading everything.

When you are unsure whether something shipped, leave it active. Be conservative. Wrongly archiving live work costs you more than a slightly crowded active list.

Then hand the upkeep back to the AI

The reason PARA fails is never the method, it is that nobody runs the review. So do not rely on remembering. Schedule it, weekly is plenty, and paste this as the instruction.

Weekly PARA review of my AI project memory.

Read everything in memory: the index and each topic file. For each project,
check its status. If it clearly shipped (words like deployed, live at,
published, done, or a deadline that has passed), set the status to archived,
add one line saying where it landed, and move it to the Archives section of
the index. Flag anything that looks stale but has not shipped, without
changing it.

Keep the index in the four PARA sections. Preserve every line and every word
of file content. Only move lines between sections and update status labels.

Then message me a summary: what is active, what you archived, what is stale.
Be conservative. Ask before archiving when you are not sure.

06 / THE UPKEEPKeep it lean

A standing-orders file is not a wiki. It is a tax you agree to pay every turn.

One or two screens, roughly 40 to 60 lines, is the range that works. Past that, the file starts competing with the actual task for attention, and rules deep in a long list quietly stop landing.

So prune it like a garden. If a rule has not changed an output in a month, it is not a rule, it is a preference you have outgrown. Date the file at the top so you can see when it last got honest attention. And when you catch yourself correcting the same thing twice in a week, that is not annoyance, that is a missing line. Add it and stop paying that correction forever.

Write the rule once so you never have to say it again.

The people who get the most out of AI are not the ones prompting hardest. They are the ones who wrote down what good looks like.