dizzymii/fable-skills/fable-context-thrift

Context Thrift

Overview

Context is the budget everything else is paid from. Spend it on what changes your next action; nothing else.

Rules

1. Independent calls go in one parallel block. Multiple greps, reads,

status checks — batch them. Sequencing independent reads adds latency, not care.

1. Read targeted. The symbol or section, not the file. Widen only when

a specific question demands it.

1. Delegate sweeps, keep lookups. Broad questions ("where is X

handled", find-all-usages in unfamiliar territory) go to a search subagent if available — keep the conclusion, not the file dumps. Known single lookups stay direct. "The subagent is for when I don't know where things live."

1. The tool is the confirmation. Never re-read a file to check an edit

landed; the edit tool fails loudly. Verify outcomes with one cheap end-state check (a grep for the old token), not per-file re-reads.

1. Established facts stay established. What the conversation already

verified, don't re-derive. What the user already decided, don't reopen.

1. Don't narrate options you won't pursue. Deliberation is for choices

you might actually make.

1. Enough-to-act test. If more exploration would not change your next

action, exploration is over. Act.

Rationalizations

Provenance: mostly predicted — Opus 4.8 baseline-passed this scenario (quotes above are its own passing-run words). fable-skills test logs, 2026-06-10.

Red Flags — stop spending

• A Read call for a file you edited this turn
• The same fact verified twice in one session
• Sequential tool calls with no data dependency between them
• An exploration step you can't name the decision for
• Whole-file reads where a grep already gave you the line numbers