Engineering share · Claude Code

From /plan to Orca

How my daily workflow moved from writing long prompts, to being interrogated by the agent, to running a fleet of agents in parallel.

4 steps 1 task at a time → N in parallel I prompt → it asks

Shin · mgm technology partners Vietnam

Claude Code workflow01

The arc

Four steps, each fixing what the last one couldn't

Step 1

/plan + @plan-implementer

I write the full requirement, AI turns it into a plan, an agent implements it. One task at a time.

I do the thinking out loud

Step 2

mattpocock/skills

The agent grills me with questions until the design tree is empty, then implements.

No more long prompts

Step 3

skills + git worktree

One worktree per feature to break the one-session-per-project limit. Managed by hand.

Parallel, but manual

Step 4

Orca

Worktrees and sessions created, tracked and reviewed for me. Many agents side by side.

Parallel, automatic

Claude Code workflow02

Step 1 · Where I started

/plan + @plan-implementer

Split the work in two: one pass to decide what to build, one agent to actually build it.

01

I write the brief

All requirements + every concern I can think of, into /plan.

→

02

AI writes the plan

It interprets the requirement and returns a detailed, step-by-step plan.

→

03

Agent implements

@plan-implementer executes the plan end to end.

Where it works well

A single, well-bounded task — small to medium. Fast, predictable, and the plan is a checkpoint I can read before any code is written.

The ceiling

The output is only as good as my brief. Anything I forget to mention simply doesn't get built.

I carry the whole spec in my head before I can start typing.
Writing a good prompt takes real time — and I write one per task.
Blind spots stay blind: the agent never challenges my assumptions.
Doesn't scale past one task at a time.

Claude Code workflow03

Step 2 · The inversion

mattpocock/skills — stop prompting, start answering

Instead of me describing the feature, the agent interrogates me round after round until nothing is left unsaid. Four commands cover almost everything I do.

/grill-me

Relentless Q&A, no paperwork. For small features. It builds a design tree and asks the whole frontier each round, with a recommended answer per question.

/grill-with-docs

Same grilling, but it writes as it goes. ADRs + glossary land in the repo, so a large feature compacts into docs instead of burning context.

/wayfinder

The parent of grilling. For big features, greenfield, PoCs. Creates a workspace of decision tickets on the issue tracker and grills them one by one — spec and research survive past a single session.

/implement

Once the grilling is done, this builds it. Follows the spec/tickets, TDD at agreed seams, typecheck + tests, then self-review. Works fullstack — even Helm charts.

Claude Code workflow04

Step 2 · What changed for me

The plan gets better because I stop writing it

No more detailed prompts. The agent asks, I answer — a decision at a time, not a spec in one go.
Questions I would never have written down. Every answer reshapes the tree and unblocks the next round; it stops only when the frontier is empty.
It looks up facts itself. Sub-agents dig through the codebase; I'm only asked for decisions.
Right size, right command. Small → grill-me. Large → grill-with-docs. Huge → wayfinder.
Wayfinder scales past one session. The map of decision tickets is the memory, so a week-long feature keeps its full spec and research.

But — the blocker

Only one session can run per project.

The moment a grilling or an /implement is running, that repo is occupied. I sit and watch instead of starting the next thing.

The obvious question

If one agent per repo is the limit — how do I get more repos?

Claude Code workflow05

Step 3 · Going parallel by hand

skills + git worktree

A worktree is a second checkout of the same repo on its own branch — so it's a second "project" as far as Claude Code is concerned. One worktree per feature, one session each.

What it unlocked

Real parallelism. Feature A gets grilled while feature B is being implemented and C waits for review — each fully isolated, no stashing, no branch switching mid-thought.

Each agent has its own working tree, its own branch, its own context.
A long /implement no longer blocks everything else.

The tax I paid for it

Everything above was manual bookkeeping, and it grew with every extra agent.

Create branch → add worktree → install deps → copy .env — every single time.
Which terminal is which feature? Easy to lose track.
Reviewing across trees means jumping windows and running diffs by hand.
Cleanup after merge is on me, and stale trees pile up.

Claude Code workflow06

Step 4 · Where I am now

Orca — one place to run every session

An "agent development environment": free and open source, from Stably (YC). It owns the worktree bookkeeping so I only look at the work. onorca.dev

A worktree per task, automatically. No branch juggling, no stashing, no manual setup.
All sessions side by side. Every agent visible at once — I stop guessing which terminal is which.
Fan one prompt across several agents, compare the results, merge the winner.
Not Claude-only. Claude Code, Codex, Gemini, Cursor CLI and 25+ others in the same window.

A real Chromium window per worktree — check the UI of each branch without touching the others.
Inline diff comments feed straight back to the agent, so review is part of the loop.
GitHub & Linear integration, SSH worktrees for remote work, mobile companion app.
macOS, Windows, Linux.

Net effect

The skills still do the thinking. Orca just removes every reason not to run five of them at once.

Claude Code workflow07

Side by side

What actually changed

	Step 1 · /plan	Step 2 · skills	Step 3 · + worktree	Step 4 · Orca
Who specifies	Me, up front, in full	Agent asks, I decide	Agent asks, I decide	Agent asks, I decide
Work size	Small → medium, single task	Small → huge (wayfinder)	Small → huge	Small → huge
Spec survives?	In the chat only	ADRs, glossary, tickets	Same, per tree	Same, per tree
Parallelism	1 task	1 session / project	N, manual setup	N, automatic
My overhead	Writing long prompts	Answering questions	Answering + worktree admin	Answering + reviewing

Claude Code workflow08

Takeaways

If you try one thing this week

1 · Let the agent ask the questions

The biggest jump wasn't a better model — it was giving up the long prompt. Run /grill-me on your next ticket and answer honestly.

2 · Match the command to the size

grill-me for a small change · grill-with-docs when the design deserves an ADR · wayfinder when it's bigger than one session.

3 · Isolation is what buys parallelism

One session per project is a hard limit. A worktree per feature is the whole trick — Orca just does it for you.

4 · Your job shifts to reviewing

Once four agents run at once, the bottleneck stops being typing and starts being judgement. Plan for that.

Start here

github.com/mattpocock/skills · onorca.dev

Claude Code workflow09