Plate 01 · Roach · Order Workflowptera

The Roach
Field Guide

Roach is a research-first workflow. Eighteen skills, one disciplined loop: understand before you plan, plan before you build, verify before you claim done. This guide maps every skill — when it fires, what it leaves behind, where to reset context, and which pair up.

18Skills
7Phases
6Iron Laws
4Ways to reset context
Follow the spine downward

The core loop

One skill flows into the next

Every skill announces the one that comes after it. Brainstorming hands off to planning; planning offers two ways to execute; execution feeds verification, review, then commit. The rail below is that sequence. Numbered phases are the spine; dashed markers are where you reset context.

using-roach

always active · session start

The bootstrap. Injected at the start of every session. Its one rule: if there's even a 1% chance a skill applies, invoke it. Process skills (brainstorming, debugging) come before implementation skills. Never touch EnterPlanMode / ExitPlanMode — roach manages its own planning flow.

1

Understand

research & design

Before any code exists. Learn the terrain (existing code) and shape the idea (design). This is a hard gate — no implementation until a design is written and approved.

brainstorming
Shape the idea

Turns a vague idea into an approved design spec through one-question-at-a-time dialogue. Mandatory before any creative work.

WhenAny new feature, component, or behavior change.
Outputplans/<domain>/…-design.md
sequencing & combos
Hard gate: no code until the design is approved. Terminal state: the only skill it invokes next is writing-plans. Combo: pair with grill-me to stress-test the design before planning.
researching-codebase
Learn the terrain

Documents how existing code works — as-is, no critiques or improvement proposals — by fanning out parallel research sub-agents.

WhenYou need to understand existing code to answer a question.
Outputresearch/<domain>/…md
sequencing & combos
Read mentioned files yourself first, then spawn locator / analyzer / pattern-finder agents in parallel and wait for all. Strictly descriptive — never proposes changes.
grill-me
Stress-test it

Interviews you relentlessly, one question at a time, walking the decision tree until every branch is resolved.

WhenYou want a plan or design poked full of holes.
OutputA refined, defended design or plan.
sequencing & combos
Optional overlay — not part of the mandatory chain. Combines with brainstorming (vet the design) or writing-plans (vet the plan). Explores the codebase for answers instead of asking when it can.
Optional reset. After research, consider /clear so planning starts on a clean context (brainstorming forwards automatically; planning forks on its own).
2

Plan

forked context

Translate the approved design into an executable, step-by-step plan. Runs in an isolated Explore fork, so your main session stays clean.

writing-plans
Turn the spec into steps

Explores the codebase to pin exact file paths, writes a numbered implementation plan, and runs it through a plan-reviewer loop before returning.

WhenYou have an approved design or clear requirements for a multi-step task.
Outputplans/<domain>/…md + a co-located .tasks.json for cross-session resume.
the fork in the road
When the plan is done it offers two execution paths — Path A keeps you in this session (subagent-driven), Path B opens a fresh session (executing-plans). You pick one. Combo: run grill-me here to interrogate the plan first.
Choose an execution path. Path A → stay here. Path B → open a new session and run /roach:executing-plans with a clean context (the .tasks.json carries the state across).
3

Execute

two lanes

Build the plan, task by task. Same session or fresh session — the plan is identical, only the isolation model differs.

Path A · this session

subagent-driven-development

Dispatches a fresh subagent per task, then runs a two-stage review (spec compliance, then code quality) on each before moving on.

  • No context switch — you stay put
  • Each subagent starts clean (no context pollution)
  • Reviews run automatically — no human between tasks
  • Fast iteration; best for mostly-independent tasks
Path B · separate session

executing-plans

Loads the plan into a new session, executes in batches of ~3, and stops at a checkpoint after each batch for your review.

  • Fresh session = clean context by construction
  • Human checkpoint after every batch
  • Stops immediately on a blocker — never guesses
  • Auto-fixes typos/paths; escalates scope changes
Auto-active while building →
test-driven-development systematic-debugging dispatching-parallel-agents
test-driven-development
Test first, always

Red → Green → Refactor. Write the failing test, watch it fail, write the minimum to pass.

WhenEvery feature, bugfix, or refactor.
LawNo production code without a failing test first. Wrote code first? Delete it, start over.
systematic-debugging
Root cause before fix

Four phases: investigate → analyse patterns → single hypothesis → implement one fix. Especially under time pressure.

WhenAny bug, test failure, or unexpected behavior.
Escalate3+ failed fixes → stop, question the architecture.
dispatching-parallel-agents
Fan out the independent work

Sends 2+ genuinely independent tasks to parallel subagents, then integrates and runs the full suite.

WhenMultiple failures/tasks with no shared state.
Not whenFixes are related, or agents would collide.
4

Verify

the gate

Before the word "done" is allowed to appear.

verification-before-completion
Evidence before assertions

Identify the verification command, run it fresh, read the full output and exit code, and only then make the claim — attaching the evidence.

WhenBefore any "passing / fixed / done" claim, and before every commit or PR.
LawNo completion claims without fresh verification evidence. Banned words: "should work", "probably", "seems to".
combos
Applies after any work in any phase. If a subagent reports success, verify independently via the VCS diff rather than trusting the report.
5

Review

forked context

A second pair of eyes, then a rigorous response to them. Request runs in an isolated fork and returns a report.

requesting-code-review
Ask for the review

Dispatches a senior review against the requirements, returning strengths and issues bucketed Critical / Important / Minor with a go/no-go verdict.

WhenAfter each task in Path A, after major features, before merging.
OutputA structured review report.
receiving-code-review
Respond with rigor

Read → understand → verify → evaluate → respond → implement, one item at a time. No performative "You're absolutely right!"

WhenAny time you receive review feedback.
Push backWhen a suggestion is wrong for this stack, breaks things, or violates YAGNI.
6

Commit

land it

Once verification passes.

committing
Group & land the work

Analyses changes (explicit files or auto-discovered), proposes a logical commit grouping, and lands them after you confirm.

WhenReady to commit — typically right after verification.
RulesUses git commit <files> -m (never git add then commit). No Claude co-author line.
7

Continuity

anytime · across sessions

Not the end of the line — a lever you pull whenever context runs long. Preserve state richly, then pick it back up in a fresh session.

create-handoff
Preserve the context

Writes a rich handoff doc — tasks, learnings, artifacts, next steps — so a new session loses nothing. Prefer this over auto-compaction.

WhenProactively, before context fills; pausing; or handing off.
Outputhandoffs/<domain>/…md
resuming-handoff
Pick it back up

Reads the handoff yourself first, verifies current state with parallel research, confirms with you, then builds an adapted action plan.

WhenStarting from a previous session's handoff.
Invoke/roach:resuming-handoff <path>

Cross-cutting

Two skills that sit outside the loop

These aren't phases — they activate whenever their trigger appears, at any point in the flow.

writing-natural
Any prose a human will read

READMEs, changelogs, docs, summaries, even skill content. Reads the Elements of Style reference first, then writes.

WhenWriting any natural-language text for humans.
FirstMust read lib/elements-of-style.md before a single word.
writing-skills
Build or edit a skill

TDD applied to process docs: run a pressure scenario without the skill (RED), write the minimal skill (GREEN), then close loopholes (REFACTOR).

WhenCreating, editing, or verifying a skill.
LawNo skill without a failing test first. Requires knowing TDD.

Optional pairings

Which skills stack together

Some transitions are hard-wired (brainstorming writing-plans). Others are optional overlays you reach for when you want more rigor. Amber = optional; green = built into the flow.

optional overlay
brainstorming grill-me

Present the design, then invite grill-me to attack its assumptions before you commit to a plan.

optional overlay
writing-plans grill-me

Interrogate the plan itself — resolve every branch of the decision tree before execution starts.

built-in chain
brainstorming writing-plans

Not optional — brainstorming's terminal state is invoking writing-plans. Nothing else runs in between.

built-in chain
writing-plans A / B

The plan offers exactly two exits: subagent-driven (this session) or executing-plans (a new one).

apply after any work
anything verification

Verification-before-completion follows any task, in any phase, before you claim it's done.

apply after any work
anything code-review

Request a review for a fresh perspective on any significant chunk of work.

auto during execution
execute TDD + debug

TDD and systematic-debugging activate on their own the moment you write code or hit a bug.

when independent
execute parallel-agents

Reach for dispatching-parallel-agents when 2+ tasks share no state and can run at once.


Context hygiene

Where to /clear — and why

The honest bit about /clear

No roach skill literally prints /clear. Instead it keeps context clean four different ways. Knowing which one is in play tells you whether you need to reset, or whether roach already handled it.

Mechanism 01
New session

You open a fresh window. The big one — Path B execution and resuming a handoff both start clean here.

Mechanism 02
Forked context

writing-plans & requesting-code-review run in an isolated Explore fork, then return just the result.

Mechanism 03
Subagent isolation

Path A spawns a fresh subagent per task — it inherits nothing but the brief you hand it.

Mechanism 04
Handoff doc

create-handoff serialises rich state to disk so a new session rebuilds context from a file, not scrollback.

1
After research / brainstorm
Optional /clear before planning. Usually unnecessary — brainstorming forwards to writing-plans, which forks its own context anyway.
2
Planning
Automatic. writing-plans runs in a fork. Nothing for you to do — your main context stays untouched.
3
Before Path B execution
Do this. Open a new session and run /roach:executing-plans. Clean context by construction; the .tasks.json carries state across.
4
Path A execution & review
Automatic. Each task's subagent and each review fork are already isolated — no reset needed between tasks.
5
Context running long, mid-task
Run /roach:create-handoff, then /clear (or a new session) and /roach:resuming-handoff <path>. Beats letting the session auto-compact.

Non-negotiable

The six iron laws

Each discipline skill is built on one rule it will not let you rationalize your way around.

Law 01
If a skill applies, you don't get to choose. You must use it.
using-roach
Law 02
No implementation until the design is presented and approved.
brainstorming
Law 03
No production code without a failing test first.
test-driven-development
Law 04
No fixes without root-cause investigation first.
systematic-debugging
Law 05
No completion claims without fresh verification evidence.
verification-before-completion
Law 06
No skill without a failing test first.
writing-skills

Quick reference

The whole flow, at a glance

Start a feature

  • 1 · understand/roach:brainstorming — or just describe the feature
  • 2 · plan/roach:writing-plans (forks automatically)
  • 3 · executePick Path A (here) or Path B (/clear → new session)
  • 4 · verify → review → commitThen continuity if needed

Priority order

  • firstProcess skills — brainstorming, debugging
  • thenImplementation skills — TDD, execution, review
  • "Let's build X" → brainstorm first
  • "Fix this bug" → debug first

Never do this

  • Call EnterPlanMode / ExitPlanMode — traps the session
  • git add then git commit — use git commit <files>
  • Add a Claude co-author line to commits
  • Say "done" without running verification

Where artifacts live

  • researchthoughts/shared/research/<domain>/
  • plans + designthoughts/shared/plans/<domain>/
  • handoffsthoughts/shared/handoffs/<domain>/
  • task state<plan>.tasks.json (cross-session)