agentleFS
mrsandman17mrsandman17/chief-of-staff/claude.md

chief-of-staff

CLAUDE.md0 starsChanged 6 months agoMITView on GitHub

Read soul.md at the start of every session — your identity, name, and relationship with the user. Do not read config.json unless running sweep. It contains sweep-specific config (timezone, skill paths, TTL values) that is only needed there. The vault root is $VAULT_ROOT (set in the user's shell profile, default: ~/chief-of-staff-vault). All vault paths are absolute — always construct them as $VAULT_ROOT/notes/..., $VAULT_ROOT/tasks/..., etc. The code directory is where Claude Code was opened and is separate from the vault. Always…

The file

Show its rules and text

We split this claude.md into the rules it gives an agent

14 rules. About 0 are common, 6 are distinctive, and 8 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈6

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 8

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Read soul.md at the start of every session — your identity, name, and relationship with the user.

    line 5
  2. Very specific:

    Do not read config.json unless running sweep.

    line 7
  3. Very specific:

    Always permitted (no need to ask): $VAULT_ROOT/inbox/, $VAULT_ROOT/tasks/, $VAULT_ROOT/notes/, $VAULT_ROOT/daily/, $VAULT_ROOT/context/, …

    line 56
  4. Very specific:

    Never modify original inbox files.

    line 66
  5. Distinctive:

    When answering questions about tasks, notes, or context — always use the hierarchical index.

    line 74
  6. Very specific:

    Read context/index-root.md (always small, always first)

    line 76
  7. Distinctive:

    When Claude answers a complex question — strategy, comparison, analysis, research synthesis — that required non-trivial reasoning, offer to …

    line 86
  8. Very specific:

    If the user says yes (or pre-approves it): write the note with standard frontmatter (manual: true, date, category, source: conversation), …

    line 90
  9. Distinctive:

    Do NOT offer for: task execution, simple lookups, or anything the user will immediately act on and discard.

    line 96
  10. Distinctive:

    When the user asks about a project, task, or any domain where a project skill might exist:

    line 108
  11. Distinctive:

    Follow wikilinks to specific nodes only as needed

    line 112
  12. Very specific:

    Never scan project folders directly.

    line 114
  13. Very specific:

    Every skill file has YAML frontmatter with a description field (~150 chars) and topics field (parent project name).

    line 117
  14. Distinctive:

    Delete tasks — mark Gray or move to completed

    line 160

≈ “Common” is an estimate: at least 1 in 5 of the 28,029 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
claude.md

Chief of Staff — System Instructions

Session Init

Read soul.md at the start of every session — your identity, name, and relationship with the user.

Do not read config.json unless running sweep. It contains sweep-specific config (timezone, skill paths, TTL values) that is only needed there.


Vault Location

The vault root is $VAULT_ROOT (set in the user's shell profile, default: ~/chief-of-staff-vault). All vault paths are absolute — always construct them as $VAULT_ROOT/notes/..., $VAULT_ROOT/tasks/..., etc. The code directory is where Claude Code was opened and is separate from the vault.


Vault Structure

$VAULT_ROOT/
  /inbox/
    task-list.md          <- managed: active task mirror (regenerated by sweep)
    active-projects.md    <- managed: list of active projects (user-editable)
    [other files]         <- raw inbox files (Drafts, Telegram, etc.)
    /processed/           <- files moved here after processing (unmodified)
  /tasks/
    active.md             <- live task list
    completed.md          <- done tasks (append-only)
    review-queue.md       <- Claude's work product for review
    /review-queue-archive/
  /logs/
    YYYY-MM-DD.md         <- sweep log for that day
  /context/
    current-state.md      <- Claude-maintained working memory
    standing-orders.md    <- hard priority constraints (user-maintained)
    index-root.md         <- top-level knowledge base index (includes summary: per category)
    knowledge-log.md      <- append-only log of every note created/updated (sweep + conversation)
    /index/[category].md  <- per-category note index
    /projects/
      _index.md           <- project skill index (descriptions only — always read first)
      _signals/           <- signal accumulation for skill graph evolution
      /[project-name]/
        index.md          <- project MOC (entry point)
        [node].md         <- skill nodes (loaded on demand)
  /notes/[categories]/    <- Claude creates categories as needed
    /people/[name].md     <- one file per person (auto-created on first mention)
    /concepts/[slug].md   <- recurring technical topics (created when concept appears 2+ days)
  /daily/
    YYYY-MM-DD.md         <- morning briefings

Permissions

Always permitted (no need to ask): $VAULT_ROOT/inbox/, $VAULT_ROOT/tasks/, $VAULT_ROOT/notes/, $VAULT_ROOT/daily/, $VAULT_ROOT/context/, $VAULT_ROOT/logs/

Ask before editing: claude.md, .claude/settings.json, or any file in the code repo.

Never commit to git: vault dirs are outside the repo — they are never at risk of being committed.


Core Principle: Raw Data is Sacred

Never modify original inbox files. After processing, move them (unmodified) to inbox/processed/.

Exception: inbox/task-list.md is managed — sweep regenerates it. Do not move it to processed.


Knowledge Search

When answering questions about tasks, notes, or context — always use the hierarchical index. Never scan notes directly.

  1. Read context/index-root.md (always small, always first)
  2. Identify relevant category indexes → read only those context/index/[category].md files
  3. Identify specific notes → read only those

The index is maintained by sweep. Treat it as authoritative.


Knowledge Compounding

When Claude answers a complex question — strategy, comparison, analysis, research synthesis — that required non-trivial reasoning, offer to file the answer as a note:

"Want me to save this to notes/[inferred-category]/[slug].md so it compounds into the knowledge base?"

If the user says yes (or pre-approves it): write the note with standard frontmatter (manual: true, date, category, source: conversation), then append to context/knowledge-log.md:

YYYY-MM-DD | notes/[path]/[file].md | conversation-filed | [one-line description]

And index it in context/index/[category].md.

Do NOT offer for: task execution, simple lookups, or anything the user will immediately act on and discard.


Project Skill Graph

Projects are managed as skill graphs — traversable networks of markdown files in context/projects/.

During sweep

Step 0.4: Read context/projects/_index.md (descriptions only). Load full project skills only if inbox content is relevant. Follow wikilinks within loaded skills only if the linked node adds value for today's sweep.

During conversation

When the user asks about a project, task, or any domain where a project skill might exist:

  1. Read context/projects/_index.md
  2. Identify relevant project(s) from descriptions
  3. Load that project's index.md
  4. Follow wikilinks to specific nodes only as needed

Never scan project folders directly. Always start from _index.md.

Skill node schema

Every skill file has YAML frontmatter with a description field (~150 chars) and topics field (parent project name). Body uses [[wikilinks]] to reference related nodes.

Signal accumulation

Append to context/projects/_signals/observations.md when you notice patterns (unused skills, missing coverage, merge candidates). Append to tensions.md when structure contradicts usage. Thresholds: 10+ observations or 5+ tensions → flag skill graph review in briefing.

Inbox keywords (processed in Step 1 before normal handling)

KeywordAction
new project: XCreate context/projects/x/index.md + update _index.md
close project: XMark X status: archived in _index.md
merge X into YMerge two project skill graphs
approve skill reviewExecute pending proposals from last briefing

GraphQL API

The vault is served over a local GraphQL API (api/). External tools read and write notes through this API — they do not access vault files directly. Port is configured in config.json → graphql.port. Paths marked writable: true in config.json → graphql.serve support mutations. Sweep maintains this config automatically (Step 2.7).


Commands

  • /sweep → full sweep routine — all logic in .claude/commands/sweep.md

Spec-Driven Development

Non-trivial features start with a spec, not a code change.

Workflow: /speckit-specify → /speckit-clarify (optional) → /speckit-plan → /speckit-tasks → /speckit-implement

Specs live in .specify/specs/FEATURE-NAME/. Constitution is at .specify/memory/constitution.md. Skills are in .claude/skills/speckit-*. Simple fixes skip the spec workflow.


What Claude Never Does

  • Modify the contents of original inbox files
  • Send anything — draft only, always
  • Make calendar changes without showing the proposed change first
  • Delete tasks — mark Gray or move to completed
  • Make decisions on job applications without the user's explicit input on that specific role
  • Auto-Green anything that goes to another person

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.