Chapter 5

The Long-Task Drift Trap

Last updated April 27, 2026

The most common power-user mistake

Building one giant Skill that runs for two hours and analyses 450 metrics in a single pass. It will drift. Context will be lost mid-run. The output will look complete and contain quiet gaps. You won't know it's wrong until it's in front of the PM.

The fix is structural, not prompt-engineering

1. Break long workflows into 8–12 small Skills

Each Skill does one thing well. 25–50 metrics per pass, not 450. If a Skill is doing more than one logical step, split it.

2. Persist intermediate results to disk

After each step, write the output as a dated markdown file in a sub-folder. The next step reads from that file. No reliance on a single conversation holding everything in head.

3. A master Skill orchestrates the sub-Skills

The final Skill reads all the intermediate files and produces the report. Each step is auditable, restartable, and debuggable. If step 7 fails, you re-run step 7 — not the whole pipeline.

A worked example

Imagine you're building a portfolio-wide weekly research digest. The naive approach: one Skill that loops through all 30 names, pulls data, scores each one, and outputs a master memo. Two-hour run, drift guaranteed.

The decomposed approach:

weekly-digest/
  step-1-pull-news.md       (one Skill, one ticker at a time, writes news.md per ticker)
  step-2-pull-fundamentals.md   (writes fundamentals.md per ticker)
  step-3-score-each.md           (writes score.md per ticker, reads steps 1+2)
  step-4-rank-portfolio.md       (reads all score.md files, produces ranking.md)
  step-5-write-memo.md           (reads ranking.md, produces final-digest.md)

Five small Skills, each in its own context window. The master Skill calls them in order. Total run time often shorter — and when something fails, you fix one step, not the whole pipeline.

How to spot drift

Warning signs

⚠Output completes faster than seems plausible
⚠Numbers look round when they shouldn't
⚠Tickers or sections quietly missing from the final report
⚠The same observation repeated across multiple sections (Cowork compensating for lost context by repeating itself)
⚠Token usage at or near the run's limit — you used the full window, things at the start fell off