build-jarvis — every feature, graded, designed, built
The one build skill. It is PRACTICE.md §§1-6 for a feature, design
included, so nothing runs in front of it — not superpowers:brainstorming, not
writing-plans, not a second build-*. You are the orchestrator: you grade,
dispatch, rule and keep the record. Design and code run in subagents, each with
an explicit model, because a skill cannot change the main session's model.
Read PRACTICE.md first — beside PLATFORMS.md, at the plugin root (../
in the install.sh layout, ../../ in a checkout or the plugin cache). It owns
classification, questions, the TASKS shape, self-review, test-first, evidence
and review. Everything below assumes it.
Step 0 — GRADE, before anything is spent
Not a build after all? A request that only changes how an existing
screen looks or reads - polish, redesign, critique, fonts/colours/spacing/
contrast, "less generic", UX copy, an a11y pass, a frame that draws wrong -
with no new behaviour is handed straight to design-arwen (or tui-tron):
print grade: handed to design-arwen (ui-only) - no build pass (or
tui-tron), load it, no design file. Naming arwen or tron does not skip
this test: a feature that names one for its look stays here, and that
skill runs as its UI pass. The line starts grade: because the grade gate
(PreToolUse on Agent|Task|Write|Edit|MultiEdit|NotebookEdit) denies
those calls after this skill loads until a grade: line is printed - the
handed-to skill's edits included.
Classify spike / bounded / architectural (PRACTICE §1).
Criticality. The work is critical when it touches auth, money,
concurrency, or a contract others depend on (a shared API, a file format, a
schema). Say which word made it critical.
Tier each seat from this table — the same one as practice/task-loop.md
§ Model selection and tickets-zordon's scout:
| The work reads like |
design |
build |
verify |
| One file, mechanical, the change is named |
— |
cheap |
build |
| 1-2 files, complete spec, nothing left to decide |
— |
cheap |
build |
| Several files, integration, or choices left open |
deep |
build |
build |
| Cause unknown, design judgement, architectural |
deep |
build |
deep |
| Critical (any row) |
deep |
build |
deep, held |
Critical's build seat is the build tier (build=opus on Claude Code,
per PLATFORMS.md), never cheap.
— = no design pass: the request or ticket is the spec, and you write it
into the design file as its one task. build is the floor for every verify
seat and for a build seat working from prose rather than named code.
ui - arwen (web/native surface), tron (terminal UI), arwen+tron
(both), none; the field sits between ceremony and mode
(references/grade.md § The ui field).
Tier → model: read .skillator/tiers.local.md (written by
grayskull-power's tiers arm step). Missing → the first slug in your
host's column of PLATFORMS.md § Role tiers, and say tiers: defaults.
Step down, never up on your own — the one exception is Phase 2's
third fix round, and only with its Ruling: line:
- a tier whose model the file marks
down: → that seat goes one tier lower;
- usage ≥ 70 % of the 5-hour window (
watch-cortana's usage-watch.* level,
the 5h= field) → every non-critical seat one tier lower; src=none
→ no step-down, and say usage unobservable;
- never below cheap, never a verify seat below build, and a held seat
keeps the highest tier still available.
Every downgrade is a line:
Ruling: <seat> <from>-><to> - <why> - <cost if wrong>.
User words win over the table (unavailability still wins over them):
"use fable" / "fable designs" → design and deep seats on fable · "opus only"
/ "all opus" → every seat opus · "cheap" / "on a budget" → every
non-critical seat one tier lower · "top tier" / "best models" → every seat
deep · a named model → that seat on it.
Your first reply text after loading this skill is the grade line —
before the design-file Write, before any Agent call. No confirmation gate
(step 8's mode question is the one Step 0 ask); the line is there so the
user can overrule before the first token is spent:
grade: bounded · critical no · design - · build cheap=sonnet · verify build=opus · ceremony lean · ui none · mode ask (single) · usage 5h 41%
Printed means reply text the user sees, starting grade:, before the
design file is written or anything is dispatched. A grade worked out in your
thinking, a Ruling: critical no in the design file, or a summary at the
end is not printed — the user cannot overrule a line they never saw.
The line is a required input. It is the design file's first line and
the first line of every Agent prompt (design, build, verify), verbatim. So
whichever spend comes first — the design-file Write or a design agent —
print the line as reply text, then paste the same line into that call.
Mode — the grade line's mode field, settled before any spend. How
Phase 2 runs:
| Mode |
What runs |
Fits |
| single |
One build agent does every task in order; design and verify stay their own seats |
— design seat, 1-3 tightly coupled tasks, tasks sharing files |
| per-task |
One fresh build agent per task (tasks-sentinels); same-file tasks never fan out |
Several independent tasks — the usual pick when nothing else fits |
| workflow |
Phases 1-3 as one workflow script (WORKFLOW.md); no Workflow tool → the host's parallel subagents |
4+ independent tasks or a sweep — costs the most tokens |
mode ask (<recommended>) — the default. Right after the grade line, ask
once: all three modes, your recommendation first (AskUserQuestion; no
such tool → a numbered list). Then end your turn — no design-file
Write, no Agent, until the user answers. Then reprint the line with
mode <m> (asked); that line is the one the design file and every
Agent prompt carry.
mode <m> (named) — the user already named one ("single agent", "one
agent per task", "use a workflow", "ultracode"). No question.
mode <m> (ruled) — the prompt or run says headless, chained or inside a
relay. No question; log Ruling: mode <m> - <why> - <cost if wrong>.
A bare feature request is not headless, whatever tools you lack.
Recommend from the grade row and how coupled the work reads. Never
workflow without the user's opt-in: unasked, the most you pick is per-task.
Start the ticket. An F ticket worked from the board is started with tickets-zordon's ticket-meta start <ID> (scripts/ticket-meta.sh, Windows .ps1) before the first edit: it flips [~], claims an unowned ticket to the session owner (whoami) and prints note: .. is owned by when someone else holds it - that note is a stop-and-ask, not a build. A subagent starting a ticket runs the same command.
Ceremony is full when the work is architectural, has 4+ tasks, runs
inside a relay run, or is said to outlive this session ("keep going after",
"hand it over", "full ceremony"); otherwise lean. Full adds the session
record, the handoff checkpoint, the TRACE table and the rework loop —
references/ceremony.md.
Grade detail, the tiers file, the unavailable-at-dispatch retry and worked
examples: references/grade.md.
Design questions — ruled or parked, never asked mid-build
- Decidable from the repo, the ticket, PRACTICE or a stated rule → rule it
and write it into the design file's
## Decisions:
Ruling: <what> - <why> - <cost if wrong>.
- Only the user can answer (taste, a fact only they hold, a conflict nothing
settles) →
relay-morpheus park, and build everything that does not depend on
it. A relay run in flight → park on its stage. A standalone build → open a
one-stage run first so the park has a home (commands in
references/grade.md).
- The six stops (destructive op, security-sensitive action, a side effect
outside this worktree, the 7-day limit at 90 %, a scope breach, a failed
repro) → ask now, alone.
- Architectural keeps PRACTICE §1's one approval gate after the design is
written and reviewed — asked when the user is present, parked when the run is
headless, chained or the user is away.
- "ask me as you go", "brainstorm with me", "let's design it together" →
live questions instead, one at a time (PRACTICE §1); each answer still lands
in
## Decisions.
Phase 1 — Design
Design seat — → write the design file yourself from the request. Otherwise
dispatch one design agent on the design seat's model with the request verbatim,
the repo context and PRACTICE.md; it writes the file and returns only its path
plus two lines:
grade: <the Step 0 line, verbatim - also already printed as reply text>
GOAL: <the task in one line>
CHOSEN: <the approach + why; architectural lists 2-3 APPROACHES first>
DESIGN: <data model / contracts, key edge cases, out of scope>
CONSTRAINTS: <what stays true across every task - exact values, names,
formats; copied verbatim into every reviewer as
[GLOBAL_CONSTRAINTS]>
TASKS: <one block per task in PRACTICE §2 shape (Files, Interfaces,
test-first steps with real code and commands), each tagged
build=<tier> from the Step 0 table>
VERIFICATION: <the exact end-to-end command and its expected output>
## Decisions <every Ruling: made so far; parked questions by Q-id>
Lean: design-<YYYY-MM-DD>-<slug>.md in the scratchpad. Full:
docs/sessions/session-<YYYY-MM-DD>-<slug>.md, plus the REQUIREMENTS / TRACE
slots in references/ceremony.md.
This core design is architecture, interfaces and workflow - no UI craft; a UI
task block says only what the surface must do (states, data, keys), not how it
looks.
UI design pass (ui not none) — after the core file is written: one
agent on the design seat's model (a — seat: the build seat's) with this
prompt:
<grade line verbatim>
Load skillator:design-arwen and read the design file at <path> first. This is a design-only pass: take arwen's mode (build for a new surface, improve for a change to one), run its Phases 0-3 and that mode's plan (build: references/build.md §2), and stop before code. Design the UI for the TASKS that touch a surface, on top of the core design - never change DESIGN, CONSTRAINTS or the interfaces. Write only a section `## UI design` (the signature, screens/states, hierarchy, platform notes, the a11y floor) and the UI task blocks' steps, so an implementer builds that design (real code, test-first). No code, no ship gate - build-jarvis's Phase 3 verify runs that. Return the file path and the section's first line.
ui tron: the same prompt with Load skillator:tui-tron in place of arwen
and its phases, and the section holds
the width matrix (the widths the app must draw at), the capture command per
width and the frame each must show; those captures go into VERIFICATION as the
render proof. Ops-board pane work: its claude plugin test frame assertions
are the render proof; a width matrix only for the standalone ops app.
ui arwen+tron: both passes, arwen then tron, each on its own surface's tasks.
Then self-review once (PRACTICE §3), both sections, and fix inline. A
ui tron VERIFICATION without the captured frames is unmet (Phase 3). A
spike stops at Phase 1: the probe and a recommendation, nothing kept.
Phase 2 — Build
The Step 0.8 mode decides the shape. per-task: one agent per task, each on
its own tagged tier; independent tasks fan out, same-file tasks do not.
single: one build agent takes every task block in order, on the highest
tagged tier, and reports per block. Either way a build agent gets fresh context,
the design file path and its task block(s) — never the design pasted, never
this session's history (PRACTICE §4; procedure practice/task-loop.md, prompt
text practice/prompts.md). Several F tickets are one pass, one task block per
ticket; a build agent implements its block(s) only — it never loads this skill,
regrades or runs its own verify. Review each block against its spec before the
next dispatch; a failed review goes to a fix agent. A third round may go one
tier up, and only with
Ruling: <seat> <from>-><to> - reviews <1> and <2> failed: <why each failed> - <cost if wrong>
citing both failed reviews; no Ruling, no step up.
Mode workflow (the user opted in) → run Phases 1-3 as one workflow script
(WORKFLOW.md): write the design file first, pass its path in args, and the
build stage reads each task's tier: model: M[t.build].
Phase 3 — Verify, review, relay
Run VERIFICATION through PRACTICE §5 — the command, now, in full, exit code
read — on the verify seat; ui arwen (or arwen+tron) adds design-arwen's
ship gate (its references/verify.md). Anything bigger than bounded, or
critical, gets one
whole-diff review on the verify seat's tier (PRACTICE §6,
code-review:code-review). Failures go back through Phase 2's fix loop. Then
relay: the grade line again as the report's first line, what was built
(files), the verification evidence, every Ruling, every parked Q-id.
Rules
- Grade before spending. No design file, no dispatch, before the grade
line is printed as reply text.
- A budget is a step-down, never a merge. A spend cap (
--max-budget-usd,
"on a budget", a cost ceiling) means "cheap" — every non-critical seat one
tier lower, said in a Ruling:. It never reaches critical work: critical
keeps its deep design seat (a dispatched design agent) and its deep, held
verify. It never folds the seats into this session:
a subagent costs no more than the same work done here, and the orchestrator
that builds has nobody to verify it.
| Rationalization |
Reality |
| "I graded it — it's in my reasoning / the design file's Rulings." |
Not printed. The line exists so the user can overrule it before spend; invisible, it overrules nothing. |
| "Small and obvious; the grade line is ceremony." |
It is one line. Size does not waive it; design - already says it was small. |
| "The budget won't cover subagents, so I'll design, build and verify here." |
Same tokens, minus the second pair of eyes. Step seats down; keep them separate. |
| "Critical work — safer if I do it all myself." |
Critical makes verify deep, held: a different seat, not this one. |
| "The cap reads as cheap, so the critical design seat drops to build." |
The step-down is for non-critical seats only. Critical design stays deep and dispatched. |
Red flag: about to Write the design file, or call Agent — a design agent
included — with no grade: line in your reply text yet, or with mode ask
still unanswered: stop and print or ask first. Critical work is no exception;
it is where the overrule matters most.
- Never an unnamed model. An omitted model inherits the session's.
- The design file is the handover. Pass it by path; the orchestrator does
not design or code, bar writing a
—-seat design file and the record.
- Autonomous, honest. A phase that fails or an agent that returns nothing is
said plainly; nothing continues on a missing piece.
- Handoff before context loss. About to
/compact or /clear → run
handoff-cortana first.
Other hosts
PLATFORMS.md turns tiers into slugs and dispatch into the host's delegate
mechanism; no delegation → run the seats sequentially in one session, each
phase's output written to disk before the model switch. Per-host notes, the
seat table and the Opus-version limit: references/platforms.md.
The UserPromptSubmit route hook that puts a feature prompt here on Claude
Code: references/route-hook.md.
1---2name: build-jarvis3description: >-4 Use whenever a feature is to be added, built, implemented or changed, any5 size - "add an import command", "implement F12", "add a settings screen",6 a dashboard, a TUI - or an F ticket is about to be worked, from the board7 or headless. Style words switch tiers: "use fable", "cheap", "opus only",8 "full ceremony". NOT for a non-feature one-line edit (typo, comment,9 rename, formatting), an unexplained bug, UI-only polish or critique10 (design-arwen, tui-tron) or design with no code.11metadata:12 source: https://skillmd.com/skills/bodish-coder/build-jarvis13---1415# build-jarvis — every feature, graded, designed, built1617The one build skill. It **is** `PRACTICE.md` §§1-6 for a feature, design18included, so nothing runs in front of it — not `superpowers:brainstorming`, not19`writing-plans`, not a second `build-*`. You are the **orchestrator**: you grade,20dispatch, rule and keep the record. Design and code run in subagents, each with21an explicit model, because a skill cannot change the main session's model.2223Read **`PRACTICE.md`** first — beside `PLATFORMS.md`, at the plugin root (`../`24in the `install.sh` layout, `../../` in a checkout or the plugin cache). It owns25classification, questions, the TASKS shape, self-review, test-first, evidence26and review. Everything below assumes it.2728## Step 0 — GRADE, before anything is spent29300. **Not a build after all?** A request that only changes how an existing31 screen looks or reads - polish, redesign, critique, fonts/colours/spacing/32 contrast, "less generic", UX copy, an a11y pass, a frame that draws wrong -33 with no new behaviour is handed straight to design-arwen (or tui-tron):34 print `grade: handed to design-arwen (ui-only) - no build pass` (or35 `tui-tron`), load it, no design file. Naming arwen or tron does not skip36 this test: a feature that names one for its look stays here, and that37 skill runs as its UI pass. The line starts `grade:` because the grade gate38 (`PreToolUse` on `Agent|Task|Write|Edit|MultiEdit|NotebookEdit`) denies39 those calls after this skill loads until a `grade:` line is printed - the40 handed-to skill's edits included.411. **Classify** spike / bounded / architectural (PRACTICE §1).422. **Criticality.** The work is **critical** when it touches auth, money,43 concurrency, or a contract others depend on (a shared API, a file format, a44 schema). Say which word made it critical.453. **Tier each seat** from this table — the same one as `practice/task-loop.md`46 § Model selection and `tickets-zordon`'s scout:4748 | The work reads like | design | build | verify |49 |---|---|---|---|50 | One file, mechanical, the change is named | — | cheap | build |51 | 1-2 files, complete spec, nothing left to decide | — | cheap | build |52 | Several files, integration, or choices left open | deep | build | build |53 | Cause unknown, design judgement, architectural | deep | build | deep |54 | Critical (any row) | deep | build | **deep, held** |5556 Critical's build seat is the **build** tier (`build=opus` on Claude Code,57 per `PLATFORMS.md`), never cheap.5859 `—` = no design pass: the request or ticket *is* the spec, and you write it60 into the design file as its one task. **build is the floor** for every verify61 seat and for a build seat working from prose rather than named code.6263 **ui** - `arwen` (web/native surface), `tron` (terminal UI), `arwen+tron`64 (both), `none`; the field sits between ceremony and mode65 ([references/grade.md § The ui field](references/grade.md#the-ui-field)).664. **Tier → model:** read `.skillator/tiers.local.md` (written by67 `grayskull-power`'s `tiers` arm step). Missing → the first slug in your68 host's column of `PLATFORMS.md` § Role tiers, and say `tiers: defaults`.695. **Step down, never up on your own** — the one exception is Phase 2's70 third fix round, and only with its `Ruling:` line:71 - a tier whose model the file marks `down:` → that seat goes one tier lower;72 - usage ≥ 70 % of the 5-hour window (`watch-cortana`'s `usage-watch.* level`,73 the `5h=` field) → every **non-critical** seat one tier lower; `src=none`74 → no step-down, and say `usage unobservable`;75 - never below cheap, never a verify seat below build, and a **held** seat76 keeps the highest tier still available.77 Every downgrade is a line: `Ruling: <seat> <from>-><to> - <why> - <cost if wrong>`.786. **User words win over the table** (unavailability still wins over them):79 "use fable" / "fable designs" → design and deep seats on fable · "opus only"80 / "all opus" → every seat opus · "cheap" / "on a budget" → every81 non-critical seat one tier lower · "top tier" / "best models" → every seat82 deep · a named model → that seat on it.837. **Your first reply text after loading this skill is the grade line** —84 before the design-file Write, before any Agent call. No confirmation gate85 (step 8's mode question is the one Step 0 ask); the line is there so the86 user can overrule before the first token is spent:8788 ```89 grade: bounded · critical no · design - · build cheap=sonnet · verify build=opus · ceremony lean · ui none · mode ask (single) · usage 5h 41%90 ```9192 **Printed means reply text the user sees**, starting `grade:`, before the93 design file is written or anything is dispatched. A grade worked out in your94 thinking, a `Ruling: critical no` in the design file, or a summary at the95 end is not printed — the user cannot overrule a line they never saw.9697 **The line is a required input.** It is the design file's first line and98 the first line of every Agent prompt (design, build, verify), verbatim. So99 whichever spend comes first — the design-file Write *or* a design agent —100 print the line as reply text, then paste the same line into that call.1018. **Mode — the grade line's `mode` field, settled before any spend.** How102 Phase 2 runs:103104 | Mode | What runs | Fits |105 |---|---|---|106 | **single** | One build agent does every task in order; design and verify stay their own seats | `—` design seat, 1-3 tightly coupled tasks, tasks sharing files |107 | **per-task** | One fresh build agent per task (`tasks-sentinels`); same-file tasks never fan out | Several independent tasks — the usual pick when nothing else fits |108 | **workflow** | Phases 1-3 as one workflow script (`WORKFLOW.md`); no `Workflow` tool → the host's parallel subagents | 4+ independent tasks or a sweep — costs the most tokens |109110 - `mode ask (<recommended>)` — the default. Right after the grade line, ask111 once: all three modes, your recommendation first (`AskUserQuestion`; no112 such tool → a numbered list). Then **end your turn** — no design-file113 Write, no Agent, until the user answers. Then reprint the line with114 `mode <m> (asked)`; that line is the one the design file and every115 Agent prompt carry.116 - `mode <m> (named)` — the user already named one ("single agent", "one117 agent per task", "use a workflow", "ultracode"). No question.118 - `mode <m> (ruled)` — the prompt or run says headless, chained or inside a119 relay. No question; log `Ruling: mode <m> - <why> - <cost if wrong>`.120 A bare feature request is not headless, whatever tools you lack.121122 Recommend from the grade row and how coupled the work reads. **Never123 workflow without the user's opt-in**: unasked, the most you pick is per-task.124125**Start the ticket.** An F ticket worked from the board is started with tickets-zordon's `ticket-meta start <ID>` (`scripts/ticket-meta.sh`, Windows `.ps1`) before the first edit: it flips `[~]`, claims an unowned ticket to the session owner (`whoami`) and prints `note: .. is owned by` when someone else holds it - that note is a stop-and-ask, not a build. A subagent starting a ticket runs the same command.126127**Ceremony** is `full` when the work is architectural, has 4+ tasks, runs128inside a relay run, or is said to outlive this session ("keep going after",129"hand it over", "full ceremony"); otherwise `lean`. Full adds the session130record, the handoff checkpoint, the TRACE table and the rework loop —131**[references/ceremony.md](references/ceremony.md)**.132133Grade detail, the tiers file, the unavailable-at-dispatch retry and worked134examples: **[references/grade.md](references/grade.md)**.135136## Design questions — ruled or parked, never asked mid-build137138- **Decidable** from the repo, the ticket, PRACTICE or a stated rule → rule it139 and write it into the design file's `## Decisions`:140 `Ruling: <what> - <why> - <cost if wrong>`.141- **Only the user can answer** (taste, a fact only they hold, a conflict nothing142 settles) → `relay-morpheus park`, and build everything that does not depend on143 it. A relay run in flight → park on its stage. A standalone build → open a144 one-stage run first so the park has a home (commands in145 [references/grade.md](references/grade.md#parking-from-a-standalone-build)).146- **The six stops** (destructive op, security-sensitive action, a side effect147 outside this worktree, the 7-day limit at 90 %, a scope breach, a failed148 repro) → ask now, alone.149- **Architectural** keeps PRACTICE §1's one approval gate after the design is150 written and reviewed — asked when the user is present, parked when the run is151 headless, chained or the user is away.152- "ask me as you go", "brainstorm with me", "let's design it together" →153 live questions instead, one at a time (PRACTICE §1); each answer still lands154 in `## Decisions`.155156## Phase 1 — Design157158Design seat `—` → write the design file yourself from the request. Otherwise159dispatch one design agent on the design seat's model with the request verbatim,160the repo context and `PRACTICE.md`; it writes the file and returns only its path161plus two lines:162163```164grade: <the Step 0 line, verbatim - also already printed as reply text>165GOAL: <the task in one line>166CHOSEN: <the approach + why; architectural lists 2-3 APPROACHES first>167DESIGN: <data model / contracts, key edge cases, out of scope>168CONSTRAINTS: <what stays true across every task - exact values, names,169 formats; copied verbatim into every reviewer as170 [GLOBAL_CONSTRAINTS]>171TASKS: <one block per task in PRACTICE §2 shape (Files, Interfaces,172 test-first steps with real code and commands), each tagged173 build=<tier> from the Step 0 table>174VERIFICATION: <the exact end-to-end command and its expected output>175## Decisions <every Ruling: made so far; parked questions by Q-id>176```177178Lean: `design-<YYYY-MM-DD>-<slug>.md` in the scratchpad. Full:179`docs/sessions/session-<YYYY-MM-DD>-<slug>.md`, plus the REQUIREMENTS / TRACE180slots in [references/ceremony.md](references/ceremony.md).181182This core design is architecture, interfaces and workflow - no UI craft; a UI183task block says only what the surface must do (states, data, keys), not how it184looks.185186**UI design pass** (`ui` not `none`) — after the core file is written: one187agent on the design seat's model (a `—` seat: the build seat's) with this188prompt:189190```191<grade line verbatim>192Load skillator:design-arwen and read the design file at <path> first. This is a design-only pass: take arwen's mode (build for a new surface, improve for a change to one), run its Phases 0-3 and that mode's plan (build: references/build.md §2), and stop before code. Design the UI for the TASKS that touch a surface, on top of the core design - never change DESIGN, CONSTRAINTS or the interfaces. Write only a section `## UI design` (the signature, screens/states, hierarchy, platform notes, the a11y floor) and the UI task blocks' steps, so an implementer builds that design (real code, test-first). No code, no ship gate - build-jarvis's Phase 3 verify runs that. Return the file path and the section's first line.193```194195`ui tron`: the same prompt with `Load skillator:tui-tron` in place of arwen196and its phases, and the section holds197the width matrix (the widths the app must draw at), the capture command per198width and the frame each must show; those captures go into VERIFICATION as the199render proof. Ops-board pane work: its `claude plugin test` frame assertions200are the render proof; a width matrix only for the standalone ops app.201`ui arwen+tron`: both passes, arwen then tron, each on its own surface's tasks.202203Then self-review once (PRACTICE §3), both sections, and fix inline. A204`ui tron` VERIFICATION without the captured frames is unmet (Phase 3). A205**spike** stops at Phase 1: the probe and a recommendation, nothing kept.206207## Phase 2 — Build208209The Step 0.8 mode decides the shape. **per-task**: one agent per task, each on210its own tagged tier; independent tasks fan out, same-file tasks do not.211**single**: one build agent takes every task block in order, on the highest212tagged tier, and reports per block. Either way a build agent gets fresh context,213the **design file path** and its task block(s) — never the design pasted, never214this session's history (PRACTICE §4; procedure `practice/task-loop.md`, prompt215text `practice/prompts.md`). Several F tickets are one pass, one task block per216ticket; a build agent implements its block(s) only — it never loads this skill,217regrades or runs its own verify. Review each block against its spec before the218next dispatch; a failed review goes to a fix agent. A third round may go one219tier up, and only with220`Ruling: <seat> <from>-><to> - reviews <1> and <2> failed: <why each failed> - <cost if wrong>`221citing both failed reviews; no Ruling, no step up.222223**Mode workflow** (the user opted in) → run Phases 1-3 as one workflow script224(`WORKFLOW.md`): write the design file first, pass its path in `args`, and the225build stage reads each task's tier: `model: M[t.build]`.226227## Phase 3 — Verify, review, relay228229Run VERIFICATION through PRACTICE §5 — the command, now, in full, exit code230read — on the verify seat; `ui arwen` (or `arwen+tron`) adds design-arwen's231ship gate (its `references/verify.md`). Anything bigger than bounded, or232critical, gets one233whole-diff review on the verify seat's tier (PRACTICE §6,234`code-review:code-review`). Failures go back through Phase 2's fix loop. Then235relay: the grade line again as the report's first line, what was built236(files), the verification evidence, every Ruling, every parked Q-id.237238## Rules239240- **Grade before spending.** No design file, no dispatch, before the grade241 line is printed as reply text.242- **A budget is a step-down, never a merge.** A spend cap (`--max-budget-usd`,243 "on a budget", a cost ceiling) means "cheap" — every non-critical seat one244 tier lower, said in a `Ruling:`. It never reaches critical work: critical245 keeps its deep design seat (a dispatched design agent) and its deep, held246 verify. It never folds the seats into this session:247 a subagent costs no more than the same work done here, and the orchestrator248 that builds has nobody to verify it.249250| Rationalization | Reality |251|---|---|252| "I graded it — it's in my reasoning / the design file's Rulings." | Not printed. The line exists so the user can overrule it before spend; invisible, it overrules nothing. |253| "Small and obvious; the grade line is ceremony." | It is one line. Size does not waive it; `design -` already says it was small. |254| "The budget won't cover subagents, so I'll design, build and verify here." | Same tokens, minus the second pair of eyes. Step seats down; keep them separate. |255| "Critical work — safer if I do it all myself." | Critical makes verify **deep, held**: a different seat, not this one. |256| "The cap reads as cheap, so the critical design seat drops to build." | The step-down is for non-critical seats only. Critical design stays deep and dispatched. |257258**Red flag:** about to Write the design file, or call Agent — a design agent259included — with no `grade:` line in your reply text yet, or with `mode ask`260still unanswered: stop and print or ask first. Critical work is no exception;261it is where the overrule matters most.262- **Never an unnamed model.** An omitted model inherits the session's.263- **The design file is the handover.** Pass it by path; the orchestrator does264 not design or code, bar writing a `—`-seat design file and the record.265- **Autonomous, honest.** A phase that fails or an agent that returns nothing is266 said plainly; nothing continues on a missing piece.267- **Handoff before context loss.** About to `/compact` or `/clear` → run268 `handoff-cortana` first.269270## Other hosts271272`PLATFORMS.md` turns tiers into slugs and dispatch into the host's delegate273mechanism; no delegation → run the seats sequentially in one session, each274phase's output written to disk before the model switch. Per-host notes, the275seat table and the Opus-version limit: [references/platforms.md](references/platforms.md).276The `UserPromptSubmit` route hook that puts a feature prompt here on Claude277Code: [references/route-hook.md](references/route-hook.md).