Run AI Like a Project, Not a Chat
Your best AI work has a half life of one chat session.
Someone on your team built a great prompt three weeks ago. It drafted a proposal outline that needed almost no editing. It lives in a chat window with a title nobody remembers.
Meanwhile a new hire is typing the same request into a blank chat, getting mediocre output, and concluding AI is overhyped.
Nothing failed. The work just had no place to live.
Part 1 covers the voice system: the profile, the preferences, the rules. Part 1: Teach AI Your Voice Once, Use It Forever This part builds the container it lives in.
The container is a project directory: one folder per initiative, a README as the front door, and skills, briefs, source material, drafts, and decisions in one place. That is the difference between an experiment and an operating system.
The 60-second version
- Give every initiative its own folder with a README explaining the project and how the AI should work on it.
- Keep a simple structure: briefs, source material, drafts, decisions, skills. Same layout every time.
- Write reusable skills for work that repeats: brief to draft, notes to summary, careful revisions, output comparisons.
- Roll out with three willing people on one shared project before you touch the rest of the team.
- Start with a 30 minute quickstart on real work, not a pilot exercise.
The chat trap
Chat is a fine place to think. It is a terrible place to keep things.
Work scattered across dozens of chats has four problems. Nobody can find anything. Searching chat history is a hope, not a method. Nothing is reusable. The good draft is buried under a hundred messages. Good prompts die quietly. And the work leaves when people leave.
Fragile approach: The knowledge lives in chat threads owned by individual people.
Portable approach: The knowledge lives in shared project folders that anyone on the team can open.
The second approach survives a chat deletion, a model switch, and a resignation letter.
Project directories as the operating system
One folder per initiative. Not per document. Per initiative.
A pricing review gets a folder. A proposal gets a folder. The board update gets a folder. Each one holds the same five compartments:
- README.md as the front door. What is this project? What is the current goal? Which voice profile and skill files apply?
- briefs/ for assignments. Each brief states the audience, the decision, the desired action, the format, and the facts the AI may use. An hour long brief is a brief that will never get written.
- sources/ for material the AI is allowed to use. Sanitized documents, approved numbers, prior drafts that worked. Approved location, proper access controls.
- drafts/ for work in progress. Name files so the sequence is obvious. Version numbers beat clever titles.
- decisions/ for what was chosen and why. It records which direction won and why, so future teams never reargue settled points.
Harborline Distribution is a fictional company; every number below is invented.
AI_Projects/
harborline_pricing_review/
README.md
briefs/
sources/
drafts/
decisions/
skills/
brief_to_draft.md
notes_to_executive_summary.md
# repeat this layout for each initiativeThe voice system from Part 1 stays shared across all projects. Project skills are workflows for the work each initiative repeats.
A skills library: four starter templates
A skill is a written workflow: Purpose, Inputs needed, Steps, Quality check. Write it once. Every assignment in that project follows it.
(a) brief_to_draft
Purpose
Turn an assignment brief into a complete first draft without inventing facts.
Inputs needed
1. A brief from briefs/ stating: audience, decision, desired action, format, maximum length.
2. The facts the AI may use, from the brief or sources/.
3. The project README and the shared voice profile.
Steps
1. Read the README, the brief, and the voice profile before drafting.
2. Draft in plain language, following the voice rules. If the audience or decision is unclear, ask before writing.
3. Attach a list of factual gaps: anything the draft needed but the brief did not provide. Do not invent them.
Quality check
1. Does the opening state the decision or problem in the first two sentences?
2. Is every number, name, and claim traceable to the brief or sources?
3. Does the ending name the next action instead of summarizing?(b) notes_to_executive_summary
Purpose
Turn rough meeting or call notes into a summary an executive will actually read.
Inputs needed
1. The raw notes.
2. The name and role of the receiver.
Steps
1. Identify the three or fewer points that require a decision or action.
2. Draft in this order: decision needed, key facts, background.
3. Keep it under 200 words unless asked for more.
4. Flag unclear or contradictory points instead of smoothing over them.
Quality check
1. Can the receiver act after reading only the first paragraph?
2. Is each claim present in the notes, not added?
3. Were ambiguities flagged instead of guessed at?(c) revise_without_changing_meaning
Purpose
Improve a draft's clarity and voice without altering what it says.
Inputs needed
1. The draft to revise.
2. The shared voice profile and preferences.
3. The improvement target: clarity, tone, length, or all three.
Steps
1. Read the draft and the voice profile.
2. Revise structure, rhythm, and language. Do not change claims, numbers, positions, or commitments.
3. Return the revision with a short list of what changed and why. Flag ambiguous sentences instead of resolving them.
Quality check
1. Does every factual claim match the original?
2. Do the changes follow the voice rules, not personal taste?
3. Is each change explainable with a rule an editor could verify?(d) compare_two_outputs
Purpose
Compare two AI generated drafts of one assignment and recommend which to develop.
Inputs needed
1. Both drafts, labeled clearly.
2. The original brief.
Steps
1. Score each draft from 1 to 5 on: accuracy, voice match, clarity, usefulness, editing required.
2. Recommend one draft as the base, with reasons tied to the scores.
3. Note whether the best parts of the other draft can be merged in.
Quality check
1. Are scores based on evidence, not impressions?
2. Were invented facts flagged in either draft?
3. Does the verdict say what to do next?Steal these templates. A mediocre skill people actually use beats a perfect skill in an empty folder.
Rolling it out to a small team
Do not start with the whole company. Three willing people, one shared project, one shared drive location.
Three people will find every usability problem. One project keeps it focused. One drive location ends the where do things live question.
Name one owner, not a committee. The owner keeps the structure clean, approves skill updates, and runs the weekly 20 minute review: what worked, what broke, what changes in the skills.
Expand only when the first project proves itself. Did the folder produce better work faster than blank chats? If yes, copy the structure. If no, fix the skills before adding people.
The 30-minute quickstart
A timer and a real assignment.
Minutes 0 to 10: Pick the project and write the README. Choose one initiative with a deadline. Write a README stating the project, the current goal, and the skills and voice files the AI should use. Under 200 words.
Minutes 10 to 20: Write the first skill. Adapt the closest template to this project: allowed sources, the approver, what must never appear.
Minutes 20 to 30: Run it on real work and fix what breaks. Use the skill on the actual assignment. Something will break. Fix the skill, not the draft.
You now have a project that runs. Not perfect. Working.
What this creates for a business
Three employees following one skill produce work that reads like one team.
A new hire opens the folder and sees how work gets done. Days, not weeks.
Someone leaves; their knowledge stays in the folder. A model changes; the skills still work. The business keeps the asset.
Frequently asked questions
We already have a wiki nobody reads. How is this different?
Wikis fail because they are archives with no owner and no rhythm. A project directory is small, has one named owner, gets used weekly on real work, and gets reviewed. If a skill is not used for a month, it gets retired. The maintenance loop is the difference between a library and a landfill.
Who owns the skills library?
One owner per skill, named in the skill file. That person runs a short weekly review: what worked, what broke, what changed. Ownership can rotate, but it is never everyone. Shared ownership of a living document means no ownership.
What if the team will not adopt it?
Do not launch it to the team. Start with two or three people and the single most painful repeatable workflow. Adoption follows relief, not mandates. When the pilot group finishes the hated task in half the time, the rest of the team will ask to join. Mandates create compliance theater; results create pull.
How do we keep proprietary data out of shared AI folders?
Three rules. First, the directory lives only in an approved location your IT already trusts. Second, an intake screen: nothing enters the examples folder without a check for client names, internal numbers, and deal terms. Third, when in doubt, fictionalize: keep the structural pattern, invent the facts. The AI learns the pattern either way.
Do we need IT approval or new software for this?
No. You are organizing files in drives and tools the company already licenses, not buying software. That said, loop IT in early as a courtesy, especially on data handling rules. It is much easier to get forgiveness for a folder than for a shadow software purchase, but you should not need either.
Final checklist
Confirm before going live:
- One folder per initiative, in one agreed shared location, with a README pointing to the voice files and skills.
- Briefs state audience, decision, desired action, format, and allowed facts.
- Sources contain no confidential information the location does not permit.
- Decisions record what was chosen and why, not just what happened.
- At least one skill documents a repeatable workflow with a quality check. One owner runs the weekly review.
- Three people used it on real work, and the skills absorbed what they learned.
- A human approves final output.
One question for you
Where does your team's best AI work go to die right now: chat history, email threads, or someone's head?
If you have an answer, you have your first project folder.