Claude Code is a great engineer. Here is how to make it your Chief of Staff.

Claude Code is very good at a task. It is not, by default, good at everything you are running: several goals, a dozen projects, a budget, and a pile of decisions nobody is making.

Left alone, each session starts from zero. Task lists rot. Decisions pile up. "Progress" becomes a feeling instead of a number.

This is the structure that fixed it. It has run a real multi-project portfolio every day for six months. It is a folder layout and a handful of Markdown files, nothing else.

The shape: three tiers

Portfolio/
├── CLAUDE.md  SCOREBOARD.md              ← Tier 1: the whole plan, every target
├── 2.1-Goal-Name/   CLAUDE.md + SCOREBOARD.md        ← Tier 2: a thin monitor per goal
│   └── 3.1.1-Project-Name/  CLAUDE · ROADMAP · TASKS · PROGRESS · DECISIONS   ← Tier 3: where work runs

Measure at the top, work at the bottom.

• Tier 1 holds every target and the priority order.
• Tier 2 is a thin monitor per goal: the target, which projects feed it, how the number is measured.
• Tier 3 is where work actually happens.

The rule that keeps it from rotting: task lists live only at Tier 3. The moment a task list appears at Tier 1 or 2, two places claim the same work and both drift.

The session flow

Every session, Claude does the same thing:

1. Read the master CLAUDE.md.
2. Read the Tier 1 scoreboard, then every Tier 2 scoreboard. Target against actual.
3. For each active project in priority order: read DECISIONS.md and TASKS.md, do every unblocked task, update PROGRESS.md, list what needs you.

Top-down to load context. Bottom-up to do work. The order matters. A Claude that reads scoreboards before tasks optimises for the number, not for the to-do list.

Five files per project, one writer each

File	Who writes	Job
CLAUDE.md	Claude, with you	project context, runbook, rules
ROADMAP.md	you only	milestones, each with an exit test
TASKS.md	Claude	working list; done items move to Done, never deleted
PROGRESS.md	Claude	append-only session log
DECISIONS.md	you, Claude as scribe	your directives, verbatim, signed

Why five files and not one: each has one writer and one job. A file with two writers becomes a negotiation.

Three rules that carry the whole thing

1. The scribe rule. Claude writes a decision entry only when you state a directive in the session, in your words. Never on inferred intent. This prevents the most expensive failure: Claude quietly making a judgment call, then treating its own note as your authority next session.

2. Flag, don't block. Anything only you can do (an account, an API key, a payment, an approval) goes under ## CEO Actions Needed. Claude writes it down and keeps going on everything else. A session never halts waiting for you.

3. Decide what can be decided. Routine technical choices: Claude decides. Anything that commits money, time or public credibility, or touches your values: Claude surfaces it as a yes/no question with a default and one clause of why. Never an open question. Open questions are why decisions pile up.

Scoreboards

Every row has the same five columns:

target · current actual · delta · as-of · source

Start with manual entry. Keep the source column from day one. It is what lets a data feed replace a manual field later without restructuring the file.

Troubleshooting discipline

When something fails, before Claude recommends any fix:

1. Which layer failed? Trigger, runtime, dependency, config, network, app logic, data.
2. Three hypotheses at that layer, including the boring ones.
3. The cheapest 30-second test that tells them apart.
4. Stop. Wait for the result before proposing a fix that costs more than the test.

One observation is a hypothesis, not a verdict. Most root causes are mundane.

Try it

The skeleton is free and MIT licensed:

github.com/backshift-works/portfolio-os

Copy it into a folder, fill in the {{fields}}, open Claude Code and say: "Read CLAUDE.md and run the session flow."

If you want the operating layer on top (the full autonomy framework, the capital-flow gate, the weekly decision card that turns every pending decision into a yes/no you can clear from your phone, and the prompt pattern for scheduled runs that never stall on a permission prompt), that is the Portfolio OS Kit.

Clock off. The work keeps going.