Skip to content

refactor: make instruction quality rules domain-agnostic#773

Closed
FL4TLiN3 wants to merge 2 commits intomainfrom
refactor/domain-agnostic-instruction-rules
Closed

refactor: make instruction quality rules domain-agnostic#773
FL4TLiN3 wants to merge 2 commits intomainfrom
refactor/domain-agnostic-instruction-rules

Conversation

@FL4TLiN3
Copy link
Contributor

Summary

  • Remove coding-specific vocabulary from instruction quality rules (TypeScript interfaces, Jest, blessed/ink, src/engine paths)
  • Replace with domain-neutral language that applies equally to coding, writing, research, design, analysis, and operations
  • Remove coding-specific example from plan's Failure Conditions
  • Generalize verify-test's artifact execution and semantic review checks

Motivation

Generated experts were overly code-focused because the rules explaining "what not to put in instructions" used exclusively coding examples, causing write-definition to interpret everything through a coding lens.

Test plan

  • Run create-expert with a non-coding domain (e.g., research, writing) and verify instructions are domain-appropriate
  • Run create-expert with a coding domain and verify instruction quality is maintained
  • Verify typecheck passes

🤖 Generated with Claude Code

FL4TLiN3 and others added 2 commits March 13, 2026 15:55
Two changes to stop plan's over-detailed Expert Definitions from being
blindly copied into perstack.toml instructions:

1. Plan: explicitly forbid deliverables/constraints in Architecture
   section — "write ONLY: name, one-line purpose, and role"
2. write-definition: new "How to Use plan.md" section — Architecture
   for tree structure only, Domain Knowledge for instruction content,
   Failure Conditions for verifier instruction
3. Verifier placement rule: must be direct child of coordinator, not
   nested under an executor

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The instruction quality rules and examples in create-expert were heavily
biased toward coding (TypeScript, Jest, blessed/ink, file paths), causing
generated experts to be overly code-focused even for non-coding domains.

Replace coding-specific vocabulary with domain-neutral language across
plan, write-definition, and verify-test so the pipeline produces experts
appropriate for any task type.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@FL4TLiN3 FL4TLiN3 enabled auto-merge (squash) March 13, 2026 16:09
@FL4TLiN3
Copy link
Contributor Author

Superseded by #774 which includes these changes plus the binary instruction checklist.

@FL4TLiN3 FL4TLiN3 closed this Mar 13, 2026
auto-merge was automatically disabled March 13, 2026 16:20

Pull request was closed

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant