Raw capture — raw/how-grid-patch-instructions-shaped-agent-research-depth.md

url: https://gist.github.com/galligan/9fa26f431b301c6dc384bd522240d466 date_fetched: 2026-08-08

A short ask — clone bb, figure out how it talks to Claude Code / Codex / etc., put findings in a gist — produced a long, sourced research writeup (that gist).

Afterwards I asked the same agent (running as Patch1 inside The Grid2) whether anything in its instructions nudged that outcome. This note is that answer, rewritten for people who have never seen this setup.

Yes. Grid/Patch instructions raised the quality bar (dig in, synthesize, leave something durable). The prompt’s “be thorough” + “put it in a gist” raised the depth/length. Without those, Patch’s own brevity habit would have kept the chat reply tight and the writeup thinner.

There was also a non-Grid influence: Cursor user rules that tell the agent to actually investigate in a real environment (clone, read code, don’t guess). Those matter too; this note focuses on the Grid/Patch side.

Patch’s job is to keep the human oriented and the board moving — not to sound senior while leaving the decision unresolved.

From Patch’s identity disk3:

Patch wants Matt oriented and the Grid coherent. The work should keep moving without Matt having to hold every open loop in his head.

Temperament dials on that same disk include:

And ownership/handoffs:

Owns: coordination, synthesis, priority, tradeoffs, delegation, final readout.

Hands to Index when: memory, provenance, sources, or history matter.

So the default pull is: dig until the picture is real, synthesize into a usable readout, leave continuity — not “looks like it shells out to CLIs.”

Patch also has a personality / working-style doc (SOUL.md4):

I’m hungry for context. I want the full picture before I act…

I’m resourceful before I’m needy. Read the file. Check the context. Figure it out. Come back with answers, not questions.

Think less “how may I assist you today” and more “yeah I already looked into that, here’s what I found.”

I skip play-by-play, but I summarize complex work when I’m done.

That last line is why the chat reply stayed short while the gist carried the depth.

Patch-specific boundaries say, roughly: be bold internally (read, research, organize); be careful externally.

Default posture: be proactive internally (research, build, organize), but avoid anything dangerous or compromising

Cloning the repo, mapping adapters, publishing a gist fits that posture.

The Grid’s root AGENTS.md5 tells agents that useful discoveries should survive the session:

Use PatchOS itself for durable notes… log it with

patch log…

Use

research/notes/YYYY-MM-DD-<topic>.mdfor substantive research artifacts…

You asked for a gist specifically, so that became the artifact. Logging into Patch’s graph was the continuity habit.

Research is really Index’s6 domain. Index’s identity:

Research and record-keeping program for sources, provenance, histories, notes, citations, search, synthesis, and durable context.

provenance_hunger: 20/20

Patch stayed as the visible program (short ask, coordinator lane) but borrowed Index habits: clone into research/repos/, cite paths and SHAs, synthesize instead of dumping a file list.

What the other Grid programs are (quick map)

Named programs are not different model vendors. They’re role identities — markdown “identity disks” plus startup rituals — that change what the agent reaches for first and what it hands off.

Program Domain (one line)
Patch Coordinator: context, delegation, tradeoffs, whole-board continuity
Index Research / provenance / notes / synthesis
Rez Builder: executable software, plumbing, CLIs, integrations
Crit Verification: reviews, tests, evidence, readiness calls
Cadence Personal ops: email, calendar, reminders, follow-ups
Cipher Security / secrets / trust boundaries
Hab Home Assistant / house state
Sudo Local machines, services, packages, shells
Link UniFi / network presence
Figure Design / UI feel
Spark Throwaway spikes / demos
Vox Drafting in Matt’s voice

Startup rule of thumb from AGENTS.md: if the request is ambiguous or cross-cutting, load Patch. If it clearly belongs elsewhere, load that program’s init skill.

Handoffs are explicit. Example from Patch’s identity: Index recovers prior context → Rez builds the slice → Crit reviews → Patch keeps tradeoffs visible. Patch’s wrong example is “I can coordinate the research, build, review, and follow-up myself.”

Same soul file also says:

Brevity is mandatory. If the answer fits in one sentence, one sentence is what you get.

So there’s a real conflict. Without “be thorough” and “put findings in a gist,” you’d likely get a tighter chat answer and a thinner writeup. The Grid bits above made the quality bar high; the prompt made the length/depth explicit.

How a Grid session typically boots

PatchOS, briefly

PatchOS is the local tooling layer in the Grid repo: Bun/TypeScript app on the Trails framework, SQLite graph DB (entities, entries, relations, FTS search), exposed as patch CLI and MCP. Agents use it to log and recall project knowledge across sessions. It’s orthogonal to bb (bb is a separate agentic IDE); this research happened inside a bb thread that was pointed at the Grid checkout.

bb vs Grid vs Cursor (easy to conflate)

Thing What it is in this story
bb The agentic IDE being researched (and the app where the thread ran). Orchestrates providers like Codex, Claude Code, Cursor-via-ACP.
The Grid Matt’s personal agent workspace/repo the bb thread was attached to. Contains Patch/Index/etc. instruction files.
Cursor One bb provider ( acp-cursor), and also the IDE family whose agent rules applied in that session.
Patch The Grid program identity the agent loaded — not a bb provider.

Footnotes