Communitygithub.com

theKashey/skills

Agent skills for skilled agents

skills 是什麼?

skills is a Claude Code agent skill that agent skills for skilled agents.

相容平台~Claude Code~Codex CLI~Cursor
npx skills add theKashey/skills

在你喜歡的 AI 中提問

開啟一個已預先載入此 Agent Skill 的新對話。

說明文件

Helix

Advance uncertain work by closing one evidence loop at a time. Cadence is a completed Arrange → Act → Assert/Collapse transition; activity volume is not progress.

Keep the transition laws visible

  • Checkpoint before work. Require an explicitly configured surface, reference, representation, and read/write authority. If any is missing, return NEEDS-HUMAN-DECISION; never invent a file, service, or transcript fallback. A Ralph-style or goal loop's persisted prompt or plan file carries execution and intent, not learning; treat it as the checkpoint only when the user has configured it as the surface.
  • Arrange before Act. Freeze the supplied situation or need, intended effect and proof, authority and boundary, selected branch's contribution, one move's maximum scope, observation source, review point, and expected plus disconfirming readbacks. Missing framing makes Next a repair or block, never permission to experiment. Helix sequences a supplied move; it does not classify the causal regime or design the probe. When installed, Read the Terrain owns that upstream work.
  • Act once. Keep one move in flight. Give its executor only the outcome boundary, branch contribution, move, observation source, review point, and readbacks. Stop at the first readback, review point, or boundary.
  • Assert before another Act. Append the immutable observation beside the frozen readbacks, then classify it as Expected, Disconfirming, Mixed, or Inconclusive. Effort, confidence, and retrospective producer narration are not evidence.
  • Collapse every ended move. Sort delivered work to its store of record, retired work to an epitaph, and everything else to deliberate disposal. Rewrite the checkpoint with exactly one Next before another Act.

The checkpoint stores verdicts, not deliberation: outcome, surviving branches, active arranged moves and results, open links, epitaphs, and one Next. Re-derive candidates and ranking each cycle. A matching readback establishes only what was observed inside its boundary; it does not prove value, correctness, or causality.

Run the current state

Start at the first incomplete transition; never repeat a completed phase.

  1. Read the checkpoint whole.
  2. Expand: derive branches from the outcome and latest results; drop any candidate whose possible results leave the next decision unchanged. Check epitaph resurrection conditions, select one branch, and keep implementation inspection behind the arranged move.
  3. Spike: Arrange, then Act once.
  4. Collapse: Assert, sort the ended move, rewrite the checkpoint, and select one Next.

Load only the applicable section of checkpoint and edge cases:

  • configure or reconstruct a checkpoint;
  • repair an orphan or unknown link;
  • resolve Mixed or Inconclusive evidence;
  • retire, resurrect, retry, or dispose of work.

Return the checkpoint reference, Next, and every orphan, unknown, or authority block.

相關技能