Overview

What problem this solves, the three levers it actually pulls, and the whole idea in one picture.

A big change does not fit in one Claude session. So you split it — and immediately pay for the split twice.

Long sessions rot. As the context window fills, a model's recall and precision degrade. The last third of a very long session is measurably worse work than the first third. You do not get a warning; you just get sloppier code.

Short sessions are expensive in a different currency. Every fresh session has to find its footing again: read the plan, read what the last session did, re-explore the code. That bootstrap is a fixed tax, and chopping work into many tiny sessions pays it over and over. It also throws away the prompt cache, which is what makes a warm session cheap.

    one long session   ███████████████████████████████████
                       └────────── quality decays in the tail ──────────┘

  many tiny sessions   ███▏███▏███▏███▏███▏███▏███▏███▏███
                       ▏ = a full bootstrap + closeout, paid every single time

         right-sized   ██████████▏██████████▏██████████
                       enough work to amortise ▏, few enough ▏ to stay sharp

And you lose the thread. Once work spans a dozen sessions, "which piece is next?" stops being obvious. Piece 7 might be ready while pieces 2 and 3 are still open. A note that says "currently on phase 5" goes stale the moment you finish something out of order — and then you build on a base that was never finished.

The three real levers

Contrary to the common belief that long sessions cost quadratically, Claude Code caches the conversation prefix, so a warm session is roughly linear in turns. The things that actually hurt:

Lever	What it is	What this skill does about it
Context rot	quality degrades as the window fills — usually bites long before cost does	sizes every session to ~60% of the window, so no session runs into the bad zone
Cache-busting	switching model, changing effort, `/compact`, a >5-min idle gap — each forces a full-price re-read	makes a model switch an explicit session boundary; discourages mid-phase `/compact`
Bootstrap tax	each fresh session re-reads plan + handoff + code before it can do anything	batches adjacent work into one session so the tax is paid once, not five times

The idea in one picture

Work is a dependency graph, not a checklist. Each phase declares which phases must finish before it can start. Then one rule decides everything:

A phase is ready when it has not been started and every dependency is done.

Readiness is computed from the set of finished phases — never from a counter. That single choice is what makes out-of-order and fan-out progress safe.

flowchart LR
    P1["1<br/>schema"] --> P2["2<br/>pricing"]
    P1 --> P3["3<br/>payments"]
    P2 --> P4["4<br/>cart API"]
    P3 --> P5["5<br/>refunds"]
    P4 --> P6["6<br/>ship"]
    P5 --> P6

    class P1,P2 done
    class P3,P4 ready
    class P5,P6 waiting

    classDef done fill:#3FB68B,stroke:#248063,color:#06251A,stroke-width:2px
    classDef ready fill:#FFB627,stroke:#B8790C,color:#2A1C00,stroke-width:3px
    classDef waiting fill:#8A9BA3,stroke:#5A6B73,color:#0E1B22,stroke-width:1px

Phases 1 and 2 are done. That makes 3 ready (its only dependency, 1, is done) and 4 ready (its only dependency, 2, is done) — one finished phase unblocked two. 5 and 6 wait, because a dependency of each is still open.

Two consequences worth internalising:

Finishing a phase can unblock several. Pick whichever you like. Whether two of them may run at the same time depends on their scope — the repos each touches, from the plan's Repos column. Disjoint scopes go in parallel; anything sharing a repo runs one at a time, because two Claude sessions in one working tree overwrite each other mid-edit.
"Finished" means every phase is done, not "we reached the highest number". You can complete 1 → 4 → 6 and the board will still, correctly, show 3 and 5 as unfinished.