weave-prompt
Weave prompt: compare independent adversarial implementations, then adopt the single best one
- License 许可
- MIT
- 所属仓库
- tony/skills
- 最近更新
- 2026-11-08
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- weave-prompt
- metadata
- { "source": "plugins/weave/skills/prompt/SKILL.md", "argument-hint": "[--passes=N] [--timeout=N|none] [--mode=fast|balanced|deep] [--workers=subagents|model-clis] <implementation prompt>" }
- description
- Weave prompt: compare independent adversarial implementations, then adopt the single best one
- allowed-tools
- [ "Bash", "Read", "Grep", "Glob", "Edit", "Write", "Task", "AskUserQuestion", "EnterPlanMode", "ExitPlanMode" ]
- disable-model-invocation
- true
02 资源文件 9
- 参考文档agents/openai.yaml292 B
- 脚本references/closest_model.py5.9 KB
- 脚本references/engine.py142.7 KB
- 参考文档references/ensemble-core.md21.3 KB
- 参考文档references/mutating-run.md11.1 KB
- 参考文档references/present-results.md9.2 KB
- 参考文档references/prompt.md773 B
- 参考文档references/prose-rules.md4.5 KB
- 参考文档references/worker-backends.md10.8 KB
03 SKILL.md 正文
Weave Prompt
Run an implementation prompt across independent adversarial workers, using host-native sub-agents by default or separate model CLIs when selected. After all workers finish, compare the implementations and adopt the single best one, whole. Unlike the weave-execute skill, which cherry-picks the best parts of each worker's work, this command never merges code across implementations.
The implementation prompt comes from $ARGUMENTS. With no prompt, ask the user what they want implemented.
A request like this skill apply the latest session runs no workers: follow references/mutating-run.md (Apply a recorded session) with this skill's adopt and update commands from its final phase, and skip the other phases.
<engine> is references/engine.py resolved from this skill's directory to an absolute path, written literally in every command, and run with python3.
Worker selection
Before any other unresolved configuration choice or operational step, read
references/worker-backends.md (Caller contract). Resolve
worker_backend from --workers=subagents|model-clis using that reference;
if the flag is absent, ask its worker question first.
If interactive choice is unavailable, honor its documented headless default.
Orchestration Plan
Enter plan mode before the first edit. In Claude Code call EnterPlanMode; in Cursor, Codex, or Gemini use /plan or Shift+Tab. Where plan mode is unavailable, present the plan as text.
Write the plan per references/mutating-run.md (Shared phases, Orchestration plan).
Present it, wait for approval, and exit plan mode (ExitPlanMode).
Phase 1: Start the session
Follow references/ensemble-core.md (Start the session) with --command prompt --mutating; it asks each unresolved setting through ask-user-choice. Dispatch per references/mutating-run.md (Shared phases).
Phase 2: Implement in parallel
Use the worker prompt body of references/mutating-run.md (Shared phases), with the implementation prompt as <the task>.
Phase 3: Compare the implementations
Capture, run the gates, and record the pass per references/mutating-run.md (Shared phases). Compare per ensemble-core.md (Blind judging, Synthesis protocol), where the judge participant scores. The convergence mode is pick-winner: the highest-scoring implementation wins, and no code merges across implementations. Its brief asks it to verify what each implementation does against what its participant claims. A failing quality gate caps that implementation's Correctness at 3. Write the comparison (scores per implementation, quality-gate results, the winner and why) to <session_dir>/pass-NNNN/synthesis.md per references/prose-rules.md.
Phase 4: Multi-pass refinement
Follow references/mutating-run.md (Refinement passes), repeating Phases 2 and 3.
Phase 5: Adopt the chosen implementation
Follow references/mutating-run.md (Apply gate and presentation): show the comparison and the recommended winner, and let the user choose and wait before touching the main checkout.
-
Copy everything the winner captured into the main checkout, deletions included:
$ python3 <engine> adopt \ --session "<session_dir>" \ --pass <final N> \ --all <winner>If
adoptrefuses, followreferences/mutating-run.md(Apply gate and presentation). -
Run the project's quality gates, then record the state:
$ python3 <engine> update \ --session "<session_dir>" \ --event adoption_applied -
Clean up per
references/mutating-run.md(Shared phases) (Apply), checkinggit diff --stat HEADagainst the winner's patch. -
Read
references/present-results.mdandreferences/prompt.md, and apply them withRESULT_KIND=prompt,ARTIFACT_PATH=<session_dir>/pass-NNNN/synthesis.md,SESSION_DIR,PASS_COUNT=passes,IN_PLAN_MODE= false,LABEL_MAP_PATH=<session_dir>/pass-NNNN/label-map.json.
Finish
Follow ensemble-core.md (Finish). Report the winning participant, the quality-gate result, and unavailable participants with their failure class.
Portability notes
ask-user-choice— follow the source's choice contract. Hosts with a structured multiple-choice tool (Claude Code'sAskUserQuestion) should use it. Honor a documented headless default when the source defines one; otherwise print a numbered list and wait for a numbered reply. Never invent a choice.$ARGUMENTS— the text the user passed when invoking this skill. If your host does not substitute it, read it as the user's request in the current turn, and ask when there is none.- Bundled files — every relative path in this skill points at a file shipped inside this skill directory. Read them from here, not from the host's plugin tree.
04 同仓库更多 Skills
serene-bliss
Use when asked to make a CLI --help, docs page, error message, or tooling UX more serene or calmer for developer experience; returns text, no edits.
review
Weave code review: run independent adversarial reviews in parallel, then synthesize findings
refine
Use when asked to refine a draft, iterate on code until cleaner, or polish a README through repeated critique and improvement.
prompt
Weave prompt: compare independent adversarial implementations, then adopt the single best one
plan
Weave planning: compare independent adversarial plans, then synthesize the best one
fix-review
Fix weave review findings: validate, add test coverage, fix, and commit each as atomic changes