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

Backshift — Sun, 04 Oct 2026 03:33:43 +0000

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:

Read the master CLAUDE.md.
Read the Tier 1 scoreboard, then every Tier 2 scoreboard. Target against actual.
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:

Which layer failed? Trigger, runtime, dependency, config, network, app logic, data.
Three hypotheses at that layer, including the boring ones.
The cheapest 30-second test that tells them apart.
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.

DEV Community: Backshift