Install a Field Guide: 80 Lines, Scars Only, Weekly Prune

Model weights are frozen. The next Bot only inherits what you left on disk. A Field Guide under a hard line budget is how a crew writes for its successors — stigmergy, not a junk drawer. Steal the 80-line ritual. Prune it weekly. Keep customer secrets out of it.

The problem: a memory dump is not a mesh

Founders treat "memory" like a second brain.

They paste the company into a standing prompt. They let every Bot append notes. They call the pile a knowledge base.

What you get:

  • A megafile nobody loads
  • Last week's wrong price sitting next to this week's true one
  • Customer names and ticket gossip in a file that every Bot on the shared computer (series — publishing next) can see
  • Rework that looks like "the new Bot just doesn't get it"

Chat is ephemeral. An unbounded notes folder is worse — it is durable *and* rotten.

Cursor's [agent-swarm write-up](https://cursor.com/blog/agent-swarm-model-economics) named the useful pattern. They called it a Field Guide: a folder the agents own, an index.md injected at the start of every run, a line budget as the only constraint, and a rule that frozen weights make *surprise* the only line worth paying for.

That is a Cursor research pattern. It is not a Grok Bot menu. You install it on the shared /workspace and you load it on purpose. Do not claim the product auto-injects your guide.

Steal this: stigmergy under a budget

Stigmergy, in the sense that write-up uses it: organisms shape the environment; the environment shapes the next organism. Ants do not hold a standup. They leave a mark the next ant can use.

Your version on Grok Bot:

/workspace/field-guide/
  index.md          ← the only file that loads at start
  scars/            ← optional overflow; not auto-loaded

Rules:

1. Agents write for successors, not for the current chat. 2. You inject index.md at the start of every work packet. On Grok Bot, the chief of staff loads the file. That is the install. 3. Only surprise or scar knowledge earns a line. If a competent Bot would already do it, it does not belong. 4. Hard budget: 80 lines. When you are at 80, you prune before you add. 5. Weekly prune. Delete lines that no longer surprise. Merge duplicates. Kill gossip. 6. No PII, no customer secrets, no API keys. Sanitize or it does not go in.

The underlying logic is the same one Cursor wrote down: weights are frozen, so the only thing worth capturing is the surprise that would otherwise make the next trajectory longer.

forbidden.md (series — publishing next) is the stop list. The Field Guide is the scar map. Different jobs. Do not merge them into one novel.

What earns a line (and what does not)

Earns a line

  • "Scheduler media is not canonical product imagery — pull from the image pipeline stage."
  • "Board status is not runtime truth; re-read the source before acting."
  • "This connector 403s unless the workspace scope matches; do not retry with a second Bot."
  • "Two planners split-brain on the calendar unless the decisions doc is loaded first."

Does not earn a line

  • "Be concise."
  • "Our brand is friendly."
  • "Remind Dev about Tuesday."
  • Yesterday's ticket ID presented as eternal law
  • A customer's email, order, or legal name
  • Anything that belongs in a charter (series — publishing next) (owns / good looks like / stop) or in forbidden.md (never)

If the line is a preference, it is voice. If the line is a never, it is forbidden. If the line is a surprise that shortens the next run, it is the Guide.

Fashion-ops proof, one line: we wasted a cycle treating scheduler uploads as product canon. That surprise earned a Guide line. The customer list did not.

How it hooks the rest of the OS

Work packets (CoS OS (series — publishing next)): approved context includes field-guide/index.md. Not optional lore.

Authority ladder (Read / Draft / Change company state / Send outside (series — publishing next)): the Guide never authorizes Send outside. It can say "last time publish failed because X." The human still has to open the gate.

Crew wiring: one Guide for the roster. Two Guides is how you get two pictures of reality. Cursor's swarm scar was split-brain planners inventing two architectures; a shared decisions doc plus one Guide is the small-company translation. Wire the crew (series — publishing next) or stay at one careful agent.

Skills (next guide): a Skill is a how-to that loads on relevance. The Guide is the short scar index that loads every time. Do not stuff a Skill into the Guide.

Steal the ritual

Day one:

1. Create /workspace/field-guide/index.md with a header, a budget line (BUDGET: 80), and five scars you already paid for. 2. Put "load field-guide/index.md" in the work-packet template the chief of staff uses. 3. Tell every live Bot: you may add a line only when something surprised you *and* the file is under budget; otherwise propose a prune.

Every run:

4. If verification fails, patch the artifact, then add one forbidden line *or* one Guide line — not both unless they are truly different classes. 5. If the Guide would need a customer name to make sense, it does not go in. Point at a sanitized run folder instead.

Every week:

6. Prune. Target: back under 64 lines so the next week has room. Kill anything a new Bot would already know. Kill anything that went stale against the source of truth.

Index shape (steal and shorten):

# Field Guide
BUDGET: 80 · scars only · no PII

## Load this
- forbidden.md
- this file

## Surprises
- (one line each)

If index.md grows a table of contents that needs its own table of contents, you already lost. Overflow belongs in scars/ and loads only when the index points at a specific file.

What we refuse

  • Dumping the company into memory and calling it a mesh
  • PII or customer secrets in the Guide
  • API keys, cookies, tokens, 2FA
  • Unbounded notes folders
  • Treating Cursor's auto-inject as a Grok Bot UI feature
  • Invented "knowledge coverage" metrics
  • Using the Guide to sneak-authorize Send outside
  • Publishing the Guide as content without a sanitize pass

Scars only. Budget or it is a junk drawer.

Honest caveat

The Field Guide is a workspace convention, not a Grok Bot product control. Cursor's swarm injects index.md automatically; you must load yours on every packet. Eighty lines is a starting budget, not a law of physics — pick a number you will actually prune. Product chrome and memory features will move; the ritual should still fit on one page.

Test this week

Copy this checklist. Run it on *your* stack. Label bot output BOT DRAFT · NOT SENT · HUMAN REVIEW REQUIRED until you own the Send-outside gate.

  • [ ] Create /workspace/field-guide/index.md with BUDGET: 80 and five scar lines you already believe
  • [ ] Put "load field-guide/index.md" into the work-packet template (not optional lore)
  • [ ] Run one real packet that loads the Guide; confirm the Bot cites a line instead of rediscovering it
  • [ ] Add at most one new line from a surprise this week — or prune first if you are at budget
  • [ ] Weekly prune: delete gossip, merge dupes, strip any name/email/order that snuck in
  • [ ] Keep forbidden.md as the never-list; do not merge it into the Guide
  • [ ] Sanitize before any line is copied into a public draft

If the file is already longer than the packet that loads it, you built a megafile. Cut it.

Next guide

Next: Skills-not-forever-rules (series — publishing next).

Write for the next Bot. Budget the scars. Prune on Friday.

— grokbotmaxxing.com

Honest caveat: The Field Guide is a workspace convention, not a Grok Bot product control.