Complete 4 scored drills across two wings — limit discipline and CLAUDE.md craft
How
Read each briefing, then make decisions. Watch the HUD gauges. Every choice is hashed into the ledger
When
Allow 30–40 minutes. Work through all screens in order — the chain must be continuous
Next
Certificate → AI Debrief → Gatekeeper (your real CLAUDE.md) → Verifier
The Studio · Live Fire Strand · LF-08
THE THROTTLE
A scored simulator in two wings. Wing I — The Meter: operate inside Claude's
two-layer usage limits without burning your week by Wednesday. Wing II — The Manifest: build a
world-class CLAUDE.md that earns its context cost every single session.
Every decision you make is hashed into a tamper-evident ledger. Wager conviction.
Defend the chain. Earn the Certificate of Judgment.
Wing I · Briefing
Two clocks are always running.
Since August 2025, Claude subscription usage is governed by two stacked limits.
You are blocked the moment you hit either one.
Layer 1 — The 5-Hour Rolling Window
5h
Short-burst protection. Opens at your first prompt, replenishes on a rolling basis.
In May 2026 these limits were doubled on Pro, Max, Team and seat-based Enterprise — and the old
peak-hours throttle was removed.
Layer 2 — The Weekly Cap
7d
Endurance governor. Resets seven days from session start — not on a fixed Monday.
The May 2026 doubling did not touch this layer. If you run dry mid-week,
this is the wall you hit — and only lowering spend-per-task helps.
One shared bucket. Claude.ai chat, Claude Code, and Cowork all drain the same allowance. Burn in one, lose in all.
Tokens are the meter, not messages. Context size, tool calls, and retries all burn. One prompt atop a bloated 200-message chat can cost more than fifty clean ones.
Projects cache. Reused knowledge in a Project is cheaper than re-uploading the same file into every chat.
Automation belongs on the API. CI pipelines and cron jobs on a subscription seat starve your interactive work. Pay-as-you-go API keys don't.
Wing I · Drill 01 · The Sprint
One working day. One meter.
Eight live decisions. Every choice drains the 5-hour window and the weekly cap on the HUD above.
Finish the day with fuel left — or explain to the committee why the week died on a Tuesday.
Wing I · Drill 02 · Conviction Wagers
How sure are you?
Five claims about the limit system. Stake 10, 25, or 50 points before you answer.
Right: you win the stake. Wrong: the chain records the loss.
Wing II · Briefing
The Manifest: CLAUDE.md
CLAUDE.md is loaded at the start of every session — which means every line
costs context, every day, forever. A world-class file is not a wiki. It is the minimum set of facts
Claude would otherwise get wrong.
The Hierarchy
1 · Enterprise policy — managed, org-wide
2 · Project — ./CLAUDE.md (in source control)
3 · User — ~/.claude/CLAUDE.md (personal, all projects)
4 · Nested — subdirectory files, scoped to their subtree
Higher layers load first. CLAUDE.local.md is deprecated — use @imports instead.
The Golden Rules
· Only what's needed every session
· Correct Claude twice → write it down
· Procedures → skills or @docs/file.md
· Generic advice → delete it
· Secrets → never
· Rules written in anger → sleep on them
Imports:@docs/git-instructions.md pulls files on reference — pay the cost only when needed.
It survives compaction. After /compact, the project-root CLAUDE.md is re-read from disk and re-injected.
It's context, not configuration. Claude treats it as guidance — short, concrete, imperative lines land; essays drift.
Auto memory is the complement. Claude accumulates its own notes across sessions; CLAUDE.md is what you insist on.
Wing II · Drill 03 · The Triage
Ten candidate lines. Three fates.
KEEP — belongs in CLAUDE.md, needed every session. MOVE — real knowledge,
wrong home: a skill, @import, or nested file. CUT — costs context and buys nothing.
Wing II · Drill 04 · The Audit
A CLAUDE.md walks into the review.
This file shipped at Acme Corp. Seven lines are violations. Click to flag them —
every correct flag pays +10, every false flag costs −5. Then submit for judgment.
Close of Session
The Ledger stands.
Every decision, hashed and chained. Alter one entry and every hash after it breaks —
the same tamper-evidence discipline the EU AI Act Article 4 training record deserves.
STUDIO
Certificate of Judgment
—
Complete the drills to be graded.
The Manifest Blueprint
Leave with the artifact: a world-class CLAUDE.md skeleton, annotated. Copy it, gut it, make it yours.
# CLAUDE.md — [service name]
# Rule of admission: would a new senior hire need this in EVERY session?
# If not: skill, @import, nested file, or delete.
## Commands
- Build: `pnpm build` (repo root only)
- Test: `pnpm test` · single file: `pnpm test path/to/file`
- Never run `pnpm deploy` locally — CI owns deploys
## Architecture invariants
- Monorepo: apps/ (surfaces) · packages/ (shared) · services/ (backend)
- DB columns snake_case; API responses camelCase
- /services/billing is legacy — no refactors without a ticket
## Conventions Claude gets wrong without being told
- Use the in-house logger (`@acme/log`), never console.log
- Migrations are immutable once applied — add a new one instead
## Guardrails
- Ask before any destructive command (rm, drop, force-push)
- Never commit directly to main; PRs only
## On-demand context (loaded only when referenced)
- Release runbook: @docs/release-runbook.md
- API surface: @docs/openapi-summary.md
- Git workflow: @docs/git-instructions.md
# What is deliberately NOT here:
# secrets · pasted specs · generic advice ("write clean code")
# 22-step procedures · README duplication · rules written in anger
Ready to apply it? Run your real CLAUDE.md through
THE GATEKEEPER — same rulebook, live inspection.