Long task state

Skill nubasu/Pseudo-Fable-Framework/frameworks/pseudo-fable-lift/.claude/skills/long-task-state

Written by Fable, for models that aren't — context frameworks that lift Opus/Sonnet-class agents to Fable-grade work discipline.

Install

npx -y skills add nubasu/Pseudo-Fable-Framework --skill long-task-state

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

External working memory for long tasks — a state file holding the contract, decisions, learned facts, failed approaches, and the next action, kept current so work survives context compaction and session breaks. Use at the start of multi-hour or multi-session work, and immediately after any compaction.

SKILL.md

1.9 KB, 367 tokens by cl100k_base, as published. Nobody here has run it

long-task-state — memory that survives

Context gets compacted and sessions end. The state file is the single source of truth for where the task stands. Trust it over your memory.

Create

At task start, create .claude/state/<task-slug>.md (add .claude/state/ to .gitignore unless the team wants it tracked):

# <task title>                     (started: <date>)

## Contract
<numbered requirements, from UNDERSTAND>

## Constraints & decisions
<each with WHY, dated>

## Plan
<milestones as checkboxes; mark the current one>

## Learned facts
<`path:line — fact` — things expensive to rediscover>

## Failed approaches
<what was tried + why it failed — never retry these blind>

## Next action
<exactly one concrete step>

Update — cadence

After each milestone; after any surprising discovery; after any decision; before any risky operation. Roughly every 30 minutes of work.
Keep it under ~150 lines. It is state, not a diary — prune completed noise; keep decisions and facts.

Recover — after compaction or in a new session

Read the state file FIRST, before acting on anything you "remember".
Re-verify the top 3 load-bearing facts against the actual code — files drift, summaries are lossy.
If your memory and the file conflict → the file wins; investigate the discrepancy before proceeding.
Continue from Next action.

Anti-patterns

Narrating history into the file — keep it state-shaped.
Updating the chat but not the file — the file is what survives.
Trusting recollection over the file after compaction.