The _MANIFEST.md: Five Minutes That Save Hours

The full series

1. The Prompt Is the Least Important Part
2. Three Files That Replace a Chief of Staff
3. You Don't Have a Productivity Problem. You Have a Setup Problem.
4. Stop Telling AI What to Do. Tell It What Done Looks Like.
5. The _MANIFEST.md: Five Minutes That Save Hours (you are here)
6. Plugins Are Composable. Stack Them.
7. The Monday Briefing That Runs Without You
8. Custom Skills: Build Once, Run Forever
9. AI as a Team, Not a Chatbot
10. Cowork Is Not Chat. Stop Using It Like Chat.

You gave Claude access to your project folder. Claude read everything, including the version you scrapped three weeks ago, the meeting notes that reversed direction, and the first draft that contradicts the third. The output now blends old thinking with new. The fix takes five minutes.

A project folder with 6 files produces clean output.

A project folder with 30 files causes confusion, not because Claude can't handle volume, but because it treats all 30 equally, and 12 of them contain thinking you abandoned weeks ago.

We've seen what these folders look like. Old drafts sit next to current versions, superseded strategy documents live alongside the brief that replaced them, and meeting notes from January contradict the decision made in March. The AI treats every file as equally valid input, blending outdated thinking into current work without any indication that something has gone stale.

The result looks almost right.

Recommendations pulled from an abandoned pricing model. Tone shifts mid-document because two different brand-voice references compete for attention in the same folder. Summaries include data points you retracted weeks ago, and the AI carries every contradiction forward into the final deliverable without flagging any of them.

The output isn't hallucinated. It is a faithful reflection of a messy, contradictory folder.

Accidental context. The most common source of confused output that nobody thinks to diagnose.

The problem nobody sees

When output quality drops, most people blame the prompt or the model.

We've watched this pattern play out repeatedly in consulting, product management, and operations. A practitioner runs a task, gets muddled results, rewrites the prompt with more constraints and more detail, gets marginally better results, and calls it solved. Then the same confusion surfaces two days later in a different deliverable.

The prompt was never the problem. The folder was.

Every file in a project folder becomes part of the AI's working context when a session begins. Cowork doesn't skip files that look old, doesn't deprioritize documents with earlier timestamps, and doesn't distinguish between a current strategy and the draft it replaced. It reads everything available and treats every document as equally authoritative, regardless of when it was created or whether a newer version exists.

This works when every file is up to date, accurate, and relevant to the task at hand.

This fails when even one outdated file contradicts a current one.

The AI won't flag the contradiction. It will average the two sources and produce output that sits somewhere between them, wrong in a way that looks polished enough to miss on first read. We've seen entire deliverables ship with retracted data baked in, because a superseded spreadsheet was still sitting in the working folder alongside the corrected version, and nobody remembered to remove it.

Deliberate context. Not accidental context.

Three tiers

Data Lifecycle Architecture

The fix is a single Markdown file called _MANIFEST.md, placed at the top level of any project folder, that tells Claude what to read first, what to read conditionally, and what to ignore entirely.

Three tiers.

The whole system.

Tier 1, Canonical. The source-of-truth documents that Claude must read first, every session, with no exceptions. The current project brief, the active brand guidelines, and the latest strategy document. Two or three files at most. If you can't limit this tier to a handful of documents, some of them aren't truly canonical and need to move down.

Tier 2, Domain. Supporting files grouped by topic are loaded only when a task touches that specific area. Research under a /research heading, pricing analysis under /pricing, client feedback under /feedback. A content strategy task doesn't need the pricing files, and a competitor analysis doesn't need the internal process documentation. Domain files provide depth on demand without adding noise to unrelated work.

Tier 3, Archival. Old drafts, superseded versions, outdated notes, and reference material you need someday but not today. Claude ignores all of it unless you specifically ask. The files stay in your folder and don't get deleted, but they stop contaminating every session with context that no longer reflects your current thinking.

A clean folder gives you a current strategy without the ghost of the version you abandoned in January.

A clean folder provides supporting research that loads only when the task requires it, keeping the working context focused on everything else.

A clean folder gives you the same clarity from thirty files as from six, because only the right files actually get read.

A clean folder gives you output that reflects where the project stands today, not a blended average of everywhere it has been.

How to write one

The underscore prefix in _MANIFEST.md is deliberate. It sorts the file to the top of any directory listing, which means Claude encounters it first, reads your tier assignments before touching anything else, and applies them to every task in that folder automatically.

Five minutes of writing.

# _MANIFEST.md

## Tier 1: Canonical (always read first)
- project-brief-v3.md: Current scope, goals, and constraints
- brand-guidelines.md: Voice, tone, and formatting standards
- stakeholder-map.md: Key contacts, roles, and decision authority

## Tier 2: Domain (read when task is relevant)
### /research
- market-analysis-q1.md: Competitive landscape and positioning
- customer-interviews.md: Synthesized findings from 12 interviews

### /pricing
- pricing-model-current.md: Active pricing tiers and logic

### /content
- content-calendar.md: Planned topics and deadlines through Q2

## Tier 3: Archival (ignore unless explicitly asked)
- project-brief-v1.md: Superseded by v3
- project-brief-v2.md: Superseded by v3
- notes-jan-kickoff.md: Direction changed since this meeting
- draft-strategy-old.md: Abandoned early planning approach

The short descriptions after each file name tell Claude why the file matters and, for archival files, why it no longer does.

These descriptions aren't decoration. They're routing instructions that help the AI determine relevance in seconds, without having to open and read every file to do so.

Making it stick

A manifest file works only if Claude knows to look for it.

One line in Global Instructions handles this permanently.

Add this to your existing instructions:

When starting any task, look for _MANIFEST.md first. Load Tier 1 files.
Only load Tier 2 files when the task explicitly touches that domain.
Never load Tier 3 files unless I specifically ask.

This loads at the start of every Cowork session, silently and automatically.

Every project folder with a manifest now includes context control. Every folder without one behaves the way it always has, with all files read at equal weight, which works fine for small, clean folders where everything is current and nothing contradicts anything else.

The manifest complements Folder Instructions rather than competing with them. Folder Instructions define how Claude should behave for a specific project, covering the client name, the deliverable format, the terminology, and the quality bar. The manifest defines what Claude should read and what it should skip.

One governs behavior. The other governs attention.

We've seen teams adopt manifests across every active project within a week, and the most common feedback is that output quality improved noticeably without changing a single prompt. The confused, blended output stopped appearing because the contradictory source material stopped loading. We've also seen practitioners catch files they forgot existed, sitting in working folders for months, quietly degrading every session.

Five minutes. Once per folder.

Do This Today

1. Open your most cluttered project folder and count the files. If the number is above ten, the AI is probably reading files you forgot were there.
2. Create a file called _MANIFEST.md at the top level of that folder.
3. List your two or three source-of-truth documents under Tier 1, Canonical. These are the files that represent your current thinking, with no caveats attached.
4. Group your supporting files into Tier 2, Domain, organized by topic. Add a one-line description to each file explaining what it contains and when Claude should read it.
5. Move everything else to Tier 3, Archival. Then add the manifest instruction to your Global Instructions: "When starting any task, look for _MANIFEST.md first. Load Tier 1 files. Only load Tier 2 when the task touches that domain. Never load Tier 3 unless explicitly asked."

---

Next in the series: one plugin makes Cowork useful. Two plugins stacked in a single task make it a compound task. Most people install one and stop. The real value is in the combination, and it takes four steps with no code to set up.

Read Part 6: Plugins Are Composable. Stack Them. →