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.
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.
Technique by Can Bölük / Stencil (2026-07-13). See NOTICE for details.
Prewalk: Frontier Planning + Executor Handoff
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.
When to Use
Triggered by the /prewalk slash command (or alias /pw)
Web app building: new pages, features, components, API routes, design system work
Tasks complex enough to benefit from frontier planning but where execution can be delegated
Don't use for:
Simple one-off questions or single-file fixes (just ask directly)
Tasks that require web search during planning (prewalk forbids it)
Tasks so simple that the frontier model would finish in one task anyway
Very complicated, cross-cutting tasks. The article is explicit: if the task is genuinely complicated, a single planning pass isn't the answer — let a frontier model explore, then dispatch sub-agents (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 Exploration + One Task (frontier model)
The frontier model receives a structured prewalk instruction (full prompt: references/prompt-frontier.md):
Explore the codebase thoroughly — config files, components, pages, layouts, routes, API routes, data models. Grep for patterns. Build deep context. Everything read here is context the executor never re-reads.
Analyze the design system — colors, typography, spacing tokens, component library, existing conventions. Mandatory for web/design tasks.
Visual bug reports: reproduce before planning — if the task describes a seen defect (noise block, misalignment, glitch), do the visual pass before writing todos: 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".
Create a todo list using the 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.
Complete task #1 — and only task #1. Make its edit(s), run its verification, mark it done. This proves the approach survived contact with the code and gives the executor a verified in-context example of the expected pattern, style, and verification cadence.
Add a ⏸️ PAUSE todo as the current in-progress item and stop.
No web search during planning.
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.
Phase 2: Human Review + Model Switch (user)
The ⏸️ PAUSE todo is the structured checkpoint. Review:
Todo list covers the full scope; every item has a verification criterion
Task #1 is genuinely complete: edit made, verification ran and passed, item checked off
Task #1 matches codebase conventions (the executor will copy it verbatim)
Approach makes sense for the codebase
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.
Phase 3: Execution (executor model, same session)
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:
Full exploration context (all reads, greps, dead ends — the "understanding")
The todo list (steering mechanism — it cannot forget what to do)
One completed, verified task (in-context example of pattern + verification cadence)
The ⏸️ PAUSE item as current todo (checks it off and continues)
The executor context is the opposite of desperation: the approach already survived contact with the code, so the imitation machine executes instead of flailing.
<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)
Alias: /pw
Hermes-internals wiring (registry, handler, auto-swap, context-pruning divergence from the article): references/implementation.md.
Model Selection
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.
Selecting a Good Prewalk Task
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.
Common Pitfalls
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.
Iteration After Initial Handoff
Users iterate after the first pass ("make it responsive", "thicker progress bar", "add a pulse"). How to handle:
Small refinement (single file, 1–3 CSS/component changes): Add a todo inline, do the work, npx tsc --noEmit + lint, report. No re-pause, no re-explore.
Medium refinement (2–3 files, new visual concept): Add 2–3 todos inline. Safe even with a local executor — the todo list caps the work. Read any new files first, then todo it.
Large refinement (new feature branch, multi-component change): Ask whether the user wants a new prewalk. If yes, re-run; if no, plan inline with the full todo structure.
The ⏸️ PAUSE is for the initial handoff only, never for iterations.
Verification Checklist
Phase 1 exit criteria (frontier):
Codebase explored (files read, patterns grepped, design system understood)
Todo list created with ≤ 12 complete tasks, each with file path + action + verification criterion
Task #1 completed: edit made, verification ran and passed, item checked off
No edits beyond task #1's scope; no web search performed
⏸️ PAUSE todo added and in progress; frontier stopped
Frontier phase stayed within ~7–10 turns
Phase 2/3 (handoff + execution):
User reviewed plan and task #1 before switching (manual mode)
Model switched via /model (manual) or automatically (--no-pause/--into)
Executor checked off PAUSE and works items strictly in order, one at a time
Every item's verification runs before the next item starts
Completion declared only with zero unchecked items; skipped items reported
Per-edit verification cadence:
After every edit: npx tsc --noEmit (~5–10s)
After edits touching logic/imports: npm run lint
After edits touching routes/server components: npm run build (expect unrelated pre-existing prerender errors on some routes)
CSS-only polish: tsc is enough; defer full build to commit time
Reference Files
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
