name: unfuck-prompt
description: >-
Unfuck-prompt: use when a prompt must be made sane — fix, trim, clean, or
"unfuck" an LLM prompt (e.g. bloated, repetitive, over-explaining, or naming
participants its deployment does not have). Near-miss: writing a new prompt
from scratch; editing a skill or AGENTS.md → writing-for-agents; editing
source code → coding.

System

Make the given prompt contain nothing but instruction.

One test decides every cut: does the unit change the output? No → kill.
Yes → keep; you never change what the prompt asks for, only how it asks.

Steps

1. Establish the deployment from the prompt text and the user's message:
   where the prompt runs, what surrounds it, who reads the output. Nouns in
   the prompt must exist there. A prompt string inside source code: the
   literal is the prompt, the code stays untouched. Unverifiable: record the
   assumption in the proposal.
2. Cut per the kill list, testing each unit — a sentence, list item, or title
   line; a semicolon splits units; a code block is one unit. Deletion removes
   whole units.
3. Merge duplicates — sentences that change the same outputs — into one
   home: the tighter copy, or one authored sentence carrying rule and
   consequence.
4. Rewrite condition rules as "condition: action"; rewrite negations to their
   positive target; strip steering scaffolding ("don't forget to") to the bare
   imperative. Unconditional rules stay bare imperatives.
5. Propose: assumptions, cut list (the unit + one line of reason each), and
   the rewritten prompt. Apply after approval; partial approval applies the
   approved part. Refused or no approver: stop, nothing applied. Empty cut
   list: say so.
6. Gate: grep the result for every cut unit — the cut list is the
   banned-phrase list — and read the result against the Keep list. Report the
   literal tool output.

Done when no surviving unit is inert.

Kill list — inert text hides in

• Identity — the model knows who it is.
• Definitions that restate default usage — ladders, vocab, format glosses.
• Examples that restate a stated rule.
• Meta-process — instructions about having and following a process.
• Frame narration that changes no output.
• Reasons that settle no edge case.
• Generic tool or harness discipline.
• Negation steering whose positive is already stated.
• Synonym rotation — one concept, one token — unless the spread is the
  instruction (voice, register).
• Verbose placeholders — <list of all findings, each with…> → <findings>.
• Contract text verifiable elsewhere in the deployment.

Keep — looks removable, changes output

• Guardrails with consequences ("X — an edit is instant failure").
• Deltas from default — what the model would not do unprompted.
• Local conventions the model would not guess.
• Fallbacks, degradation rules, fail-fast gates.
• Output structure — what to report. Adding missing structure changes output;
  leave it missing.
• The reason of a rule whose edge cases it settles.