菜单
Use when building web apps or design tasks via /prewalk command. Frontier model explores, plans, completes exactly one task, then pauses for model switch. Executor continues from todo list.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/westfable/hermes-prewalk --skill prewalk命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
基于 SOC 职业分类
| name | prewalk |
| description | Use when building web apps or design tasks via /prewalk command. Frontier model explores, plans, completes exactly one task, then pauses for model switch. Executor continues from todo list. |
| version | 2.0.0 |
| author | Hermes Agent |
| license | MIT |
| platforms | ["linux","macos","windows"] |
| metadata | {"hermes":{"tags":["prewalk","planning","model-switching","web-app","design","cost-optimization"],"homepage":"https://stencil.so/blog/prewalk","related_skills":["plan","todo"]}} |
| attribution | Technique by Can Bölük / Stencil (2026-07-13). See NOTICE for details. |
Hand off the context window, not a plan document. The frontier model explores, creates a todo list, completes exactly one task (edit + verification), then pauses. The executor inherits all the exploration context, the todo list as a steering mechanism, and one completed, verified task as an in-context example.
Based on prewalk research by Can Bölük / Stencil: the expensive part of an agent's work is reading (91% of tokens), not editing (9%). Traditional /plan makes the frontier model read everything at frontier prices, write a 2K-token "postcard", then the cheap model re-reads everything to execute — you pay for reads twice, and benchmarks show /plan costs more than no handoff at all. Full research, benchmarks, and the cheating-reduction effect: references/concept.md.
The core prompt files (references/prompt-frontier.md and references/prompt-executor.md) are written to be harness-neutral — they can be used with any agent that supports a todo/task-list tool and mid-session model switches. Hermes-specific details are relegated to bindings notes at the end of each file. See references/prompt-frontier.md and references/prompt-executor.md for the prompts, and the "Using Prewalk Outside Hermes" section of this repo's README for agent-agnostic usage.
/prewalk slash command (or alias /pw)Don't use for:
delegate_task) per workstream. Prewalk is for contained scope; a game of telephone doesn't help with sprawl.Note on invocation shape: The skill triggers when the user requests a web app or design task complex enough to warrant planning, regardless of whether they typed /prewalk literally. The slash command is a convenience wrapper, not a requirement. If a user says "let's redesign X" and X fits the criteria above, follow this workflow naturally.
Phase 1 (frontier) Phase 2 (human) Phase 3 (executor)
explore → plan → review → /compress? → check off PAUSE →
complete task #1 → /model <executor> → execute tasks 2..N →
add ⏸️ PAUSE → stop "continue" verify each → report
The frontier model receives a structured prewalk instruction (full prompt: references/prompt-frontier.md):
browser_navigate to the page, browser_vision screenshot, browser_console DOM probes (getBoundingClientRect + getComputedStyle z-index/position of the section and its decorative layers) to identify the culprit element. Todos then name the exact class/element instead of "investigate X".todo tool — max 12 items (overridable with --max-todos N). Each item is a complete task: concrete file path + what to do + verification criterion. Item #1 is the foundational task everything else builds on.Turn budget: the frontier phase should land in ~7–10 turns (article median: ~7). This isn't just cost discipline — stopping the frontier model early, while it's still in its confident phase, is what suppresses the desperation behaviors (web-searching for known fixes, thrashing). If the phase balloons past ~12 turns with no todo list and no completed task, interrupt: the task is probably too big for prewalk, or the model is lost.
The ⏸️ PAUSE todo is the structured checkpoint. Review:
Manual mode (default): Run /model <whatever-you-want>. Typically a local LLM via LM Studio (Gemma 4 26B, Qwen 3.6 36B, etc.), Ollama, or any other backend. The model choice is entirely up to you — prewalk just pauses and lets you pick.
Auto mode (--no-pause or --into <model>): The system automatically swaps to the executor after the frontier turn completes — --into names it, --no-pause alone uses cline-pass/deepseek-v4-flash. No human checkpoint.
Context compression: If the exploration phase was long (many files read), consider running /compress before switching to a model with a smaller context window. Only when the conversation is visibly long.
Manual mode: After switching via /model, type continue (or your own continuation prompt). Auto mode: sent automatically. Full continuation prompt: references/prompt-executor.md.
The executor inherits:
The executor context is the opposite of desperation: the approach already survived contact with the code, so the imitation machine executes instead of flailing.
/prewalk <task description> [--no-pause] [--into MODEL] [--max-todos N]
<task description> — everything after the command, treated as the task--no-pause — auto-swap to the default executor (cline-pass/deepseek-v4-flash) after the frontier turn. Skips the human checkpoint--into MODEL — auto-swap to a named executor (mirrors omp's --prewalk-into). Implies no pause--max-todos N — override the default 12-item todo cap (use sparingly — too many items loses the executor)/pwHermes-internals wiring (registry, handler, auto-swap, context-pruning divergence from the article): references/implementation.md.
| Mode | Frontier | Executor | How to switch |
|---|---|---|---|
| Manual (default) | cline-pass/glm-5.2 | User's choice (local LLM via LM Studio, Ollama, etc.) | /model <name> |
Auto (--no-pause) | cline-pass/glm-5.2 | cline-pass/deepseek-v4-flash | Automatic |
Auto (--into M) | cline-pass/glm-5.2 | M | Automatic |
Local-executor considerations (context windows, quality floor, when to fall back to frontier) and the full cline-pass model list: references/model-selection.md.
Score candidates on six criteria — exploration surface (10+ files), contained scope (8–12 todos), existing conventions to copy, a natural foundational task #1, no blocking data dependencies, clear verification per todo. Full rubric, scouting procedure for existing codebases, and a worked example: references/task-selection.md. Task-description template with examples: templates/task-template.md.
Todo items too vague — "build the auth page" gives the executor nothing concrete. Each item needs a file path and a verification criterion: "Create path/to/file with [specifics]. verify: [checkable criterion]"
Task #1 too big — a multi-file feature as the first task defeats the purpose. Task #1 should be one file, one concept, verified. If it can't be, restructure the todo list until item #1 is genuinely foundational and self-contained.
Frontier model doesn't stop after task #1 — the prompt says one task, but sometimes the model keeps going into task #2. Interrupt with Esc or /stop. The whole point is to stop it in its confident phase.
Frontier model stops without verifying — an unverified task #1 is a plan fragment, not a completed task. The executor will copy the sloppiness. Task #1 counts only when its verification criterion ran and passed.
Too many todo items — 30 items means the executor gets lost (the article notes GPT-5.6 loves 60-item lists and batch-completes them). Cap at 12; group related work into single items with sub-steps.
Executor batch-completes or declares done prematurely — the documented small-model failure mode. The executor prompt forbids skipping ahead and requires re-reading the todo list before declaring completion. If it persists, the model is below the floor for this task — switch back to frontier: /model cline-pass/glm-5.2
Context window overflow — exploration reads too many files, executor model can't hold it all. Run /compress before switching models.
Model cheats (web search) — the frontier prompt forbids web search during planning. If the model searches anyway, interrupt and re-run.
Design system mismatch — task #1 must match existing conventions exactly, or the executor copies the wrong pattern. Read 2-3 existing components before making the edit.
Treating post-handoff refinements as new prewalks — when the user comes back with "improve X" or "fix Y on mobile" after the handoff, handle the refinement inline: add a todo, do the work, verify, done. Only restart prewalk if the refinement needs fresh exploration. Do NOT re-issue the PAUSE checkpoint for iterations. See "Iteration After Initial Handoff" below.
Over-verifying visual-only changes — for CSS-only or pure-styling refinements, npx tsc --noEmit is sufficient. Skip npm run build unless the change touches logic, routes, or imports; unrelated pre-existing route errors (e.g. an /account/mfa prerender failure) surface as false failures.
Skipping visual reproduction for visual bug reports — when the task describes a seen defect, do NOT plan from code alone. Reproduce it visually first (screenshot + DOM geometry probes), then write todos that name the exact element/class. In a ROGUE FM hero pass, four overlapping decorative layers were candidates and only geometry probing could attribute the artifact.
Narrow default viewport hides desktop regressions — the built-in browser screenshot renders at a narrow viewport by default; "desktop looks fine" shots miss ≥1024px layout bugs. Verify responsive/desktop CSS with Playwright pinned to explicit desktop dimensions (e.g. 1600×900) and probe ambiguous pixels with elementsFromPoint. Both recipes: references/frontend-verification.md.
Users iterate after the first pass ("make it responsive", "thicker progress bar", "add a pulse"). How to handle:
npx tsc --noEmit + lint, report. No re-pause, no re-explore.The ⏸️ PAUSE is for the initial handoff only, never for iterations.
Phase 1 exit criteria (frontier):
Phase 2/3 (handoff + execution):
/model (manual) or automatically (--no-pause/--into)Per-edit verification cadence:
npx tsc --noEmit (~5–10s)npm run lintnpm run build (expect unrelated pre-existing prerender errors on some routes)tsc is enough; defer full build to commit time| File | Contents |
|---|---|
references/concept.md | Research, benchmarks, turn-budget rationale, cheating-reduction data (condensed from the Stencil article) |
references/prompt-frontier.md | Full Phase 1 instruction (harness-neutral; Hermes bindings at end) |
references/prompt-executor.md | Full Phase 3 continuation prompt (harness-neutral; Hermes bindings at end) |
references/model-selection.md | Frontier/executor pairing guidance + full cline-pass model list |
references/task-selection.md | 6-point rubric + scouting procedure for picking prewalk tasks |
references/plan-design-review.md | Plan-mode design review: patterns to challenge before saving, Framer Motion gotchas, patch discipline |
references/frontend-verification.md | Playwright desktop-pinned screenshots + DOM geometry-probe recipes |
references/css-animation-performance.md | Diagnosing scroll jank: enumerate animated layers, classify property cost |
references/implementation.md | Hermes slash-command internals: registry, handler, auto-swap, --into, context-pruning divergence |
templates/task-template.md | Template + examples for structuring a prewalk task description |