hike（hike 搜索）

Write a narrative explaining what happened in this coding session.

Phase 1: Gather context

Build a clear picture of what changed and why. Use whatever sources are available — pick the ones that fill in the gaps:

• Conversation history — the richest source when present. Extract the intent, pivots, decisions, and aha moments. Skip noise (retries, dead ends, routine tool output).
• Git — git diff, git status, git log. Essential when the conversation is thin or absent.
• File reads — read key changed files when the diff alone doesn't make the change legible.

By the end of this phase you should know:

• Session goal — what was the intent?
• Files changed — which files were touched?
• Behavior changes — what's different now?
• Notable decisions — what tradeoffs were made?

Discard dead ends. Only the final state matters.

Phase 2: Decide Explanation Order

Look at all the files changed and think about the best order to present them. This is a teaching order, not the chronological order of the session.

• Foundation before usage — if file A defines a helper that file B uses, explain A first.
• Setup before dependent logic — config, types, schemas before the code that relies on them.
• Core mechanism before edge cases — the happy path first, error handling after.
• Split a file across multiple <Walk> blocks when it helps — if a file has many changes or changes that serve different purposes, explain each group in a separate <Walk> placed where it fits in the narrative. Don't force everything into one block just because it's the same file.

The goal is that at every point in the narrative, the reader already has the context they need for what comes next. If a file change depends on understanding another file's change, the dependency comes first.

Phase 3: Build the Steps for Each File

The goal is to split each file's diff into multiple small steps that make the changes easy to follow. The UI animates transitions between steps, so the reader only needs to notice what changed — not re-read the whole block.

Go through the ordered files one at a time. For each file, think about how to break the diff into steps.

How to split

• Hard limit: no more than 5 lines changed between consecutive steps. "Changed" means lines that are different when comparing step N to step N-1 (for the first step in a Walk, all lines count). Collapsing previously-introduced code into a comment doesn't count as a changed line. If a step exceeds 5, split it — no exceptions. Count literally.
• If you're wondering whether to split — split.
• If it's a new file, start very small.
• Keep some shared code between consecutive steps so the reader has an anchor for the transition. If two steps share no code at all, they probably belong in separate <Walk> blocks.
• When stepping through sibling functions in the same file, collapse the previous function and show the new function below it. This gives the reader a visual anchor for the transition. If the functions are truly unrelated, use separate <Walk> blocks instead.
• Freeze context code between steps. Code carried over from a previous step for orientation must not change — don't silently add, modify, or reorder lines in it. (Collapsing multiple lines into a comment is fine — that reduces noise.) If the reader sees the context shift, their attention splits between the context change and the new code you're actually introducing. If you need to change that code too, do it in its own step first.

What to include in each step

• Only the code that matters. Filler code or unimportant classnames distract the reader from the important parts. Trim Tailwind class strings to only the classes relevant to the explanation (e.g., if the callout is about print:hidden, show className="... print:hidden" — don't include the full "fixed bottom-0 left-0 right-0 bg-white/90 dark:bg-zinc-900/90 backdrop-blur..." string). Omit the className prop entirely when no class is relevant.
• Always include parent scope. Every snippet must show all enclosing scopes up to the top-level declaration (function, class, component) so the reader knows *where* in the file the code lives. Never show floating lines without their surrounding structure.
• If some code isn't used until a later step, wait until that step to introduce it. Don't preload imports, helpers, or variables.
• If needed, use comments to signal collapsed code, code that will be introduced later, or code from earlier steps that is no longer relevant to the current step.
• Add a !callout to every step. Callouts guide the reader's attention and help them parse the step faster. Carefully pick the most important line in the step to place the callout — it should point to the key change.
• Max one !callout per step. If a step needs more than one callout, it's probably better to split into more steps.
• Mark changes. Wrap contiguous new/changed lines with !mark-start and !mark-end. The callout target can be inside the range — that's fine. Marks are mandatory whenever a step mixes context code with new/changed code — they're the only way the reader knows what changed. Skip marks only when most lines in the step would be marked (marks add no signal in that case).
• For extra information, use !tooltip — supplementary info shown on demand.

Walk syntax

Each group of related changes gets one <Walk> block with a filename attribute (a file may have multiple <Walk> blocks if its changes are unrelated — see Phase 2). Inside it, each step is a code fence with "!!" in the metastring. Tooltip content goes in ##!id headings after the steps.

<Walk filename="app/api/mtime/route.ts">

// !tooltip[/force-dynamic/m] dynamic export const dynamic = "force-dynamic";

export async function GET(request: NextRequest) { const fileName = request.nextUrl.searchParams.get("file"); const filePath = getFilePath(fileName); // !callout[/statSync/m] Read modified time from file system const stat = fs.statSync(filePath); return NextResponse.json({ mtime: stat.mtimeMs }); }

// !tooltip[/force-dynamic/m] dynamic export const dynamic = "force-dynamic";

export async function GET(request: NextRequest) { const fileName = request.nextUrl.searchParams.get("file");

// !mark-start if (!fileName) { // !callout[/NextResponse/m] Handle missing file return NextResponse.json({ error: "Missing file" }, { status: 400 }); } // !mark-end

const filePath = getFilePath(fileName); const stat = fs.statSync(filePath); return NextResponse.json({ mtime: stat.mtimeMs }); }

## !dynamic

Next.js caches route handlers by default; `force-dynamic` ensures every poll gets a fresh stat.

</Walk>

Annotation reference:

• !callout[/regex/m] message — explains intent at a regex match. The m flag searches forward through the step to find the match.
• !tooltip[/regex/m] id — on-demand detail at a regex match.
• !mark-start / !mark-end — highlights all lines between the two markers (visual emphasis, no message).
• !diff - — marks the next line as removed (struck-through / red). Use when highlighting a deletion is important for the explanation.
• Match comment style to language, for example:

- TS/JS: //!callout[/regex/m] - JSX/TSX blocks: {/*!callout[/regex/m] */}

Common Pitfalls

Parent scope

Every snippet must show all enclosing scopes up to the top-level declaration. Never show floating lines.

Bad — shows the inner function but not where it lives:

function handleDelete() {
  // ...
}

Good — full nesting visible:

function UserCard() {
  function handleDelete() {
    // ...
  }
  // ...
}

Bad — floating JSX, no enclosing component:

<button onClick={handleDelete}>
  <Trash2 className="size-4" />
</button>

Good — component scope visible:

function UserCard() {
  return (
    <div>
      <button onClick={handleDelete}>
        <Trash2 className="size-4" />
      </button>
    </div>
  );
}

Mark usage

Use !mark-start before the first changed line and !mark-end after the last changed line to bracket contiguous ranges. For non-contiguous changes, use separate pairs.

Bad — no marks, changed lines buried in context:

export async function GET(request: NextRequest) {
  const fileName = request.nextUrl.searchParams.get("file");
  const filePath = getFilePath(fileName);
  if (!fileName) {
    return NextResponse.json({ error: "Missing param" }, { status: 400 });
  }
  const normalized = path.normalize(filePath);
  // !callout[/startsWith/m] Prevent path traversal
  if (!normalized.startsWith(BASE_DIR)) {
    return NextResponse.json({ error: "Forbidden" }, { status: 403 });
  }
  const stat = fs.statSync(filePath);
  return NextResponse.json({ mtime: stat.mtimeMs });
}

Good — contiguous changed lines bracketed with !mark-start/!mark-end:

export async function GET(request: NextRequest) {
  const fileName = request.nextUrl.searchParams.get("file");
  const filePath = getFilePath(fileName);
  // !mark-start
  if (!fileName) {
    return NextResponse.json({ error: "Missing param" }, { status: 400 });
  }
  const normalized = path.normalize(filePath);
  // !callout[/startsWith/m] Prevent path traversal
  if (!normalized.startsWith(BASE_DIR)) {
    return NextResponse.json({ error: "Forbidden" }, { status: 403 });
  }
  // !mark-end
  const stat = fs.statSync(filePath);
  return NextResponse.json({ mtime: stat.mtimeMs });
}

Collapsing code

When trimming unimportant code, always use valid syntax. Use comments (//...) to collapse lines. Never use bare ... as code — it's not valid in most languages and breaks JSX/TSX parsing.

Bad — bare ... is not valid syntax:

<text x={x} ...>x</text>

function foo(...) { }

Good — just omit unimportant props, no placeholder needed:

<text x={x}>x</text>

For function args or other expressions, use a comment:

function foo(/* ... */) {}

When collapsing multiple lines inside a block, use a single-line comment:

function UserCard() {
  // ...
}

Comment syntax

Annotation comments must use the comment syntax of the surrounding context AND appear in a valid position for that syntax.

Bad — // comments inside JSX:

function UserCard() {
  // !mark-start
  const name = "Alice";
  return (
    <div>
      // !mark-end
      <button onClick={handleClick}>Save</button>
    </div>
  );
}

Good — // in JS context, {/* */} in JSX context:

function UserCard() {
  // !mark-start
  const name = "Alice";
  return (
    <div>
      {/* !mark-end */}
      <button onClick={handleClick}>Save</button>
    </div>
  );
}

Bad — {/* */} between JSX element attributes (not a valid position):

<iframe
  {/* !callout[/print:hidden/m] Hidden on print */}
  className="... print:hidden"
/>

Good — annotation on its own line before the element:

{
  /* !callout[/print:hidden/m] Hidden on print */
}
<iframe className="... print:hidden" />;

Note: {/* */} annotations are only valid as JSX children (inside an element's opening/closing tags), never between attributes, and never as bare expressions in ternary branches like condition? ({/*!mark-start */} <div>...</div>):.... Place them inside the nearest enclosing element instead.

Phase 3.1: Self-review

After building all steps for a file, re-read each step and verify:

1. Does it show the enclosing function/class/block? If not, add parent scope.
2. Is there filler code that doesn't serve the explanation? Long Tailwind class strings with irrelevant classes? If so, trim it.
3. Does it introduce code that isn't relevant until a later step? If so, move it.
4. Count lines changed from the previous step (or all lines for the first step). More than 5? Split it.
5. Do consecutive steps share enough code for the reader to stay oriented? If not, add shared context.
6. If the step isn't self-evident, does it have a !callout? If not, add one.
7. Are contiguous new/changed lines bracketed with !mark-start/!mark-end? If not, add them. Only skip marks with good reason (e.g., most lines in the step are changed).
8. Compare each step's context code to the previous step. Is any carried-over code different (other than collapsing multiple lines into a comment)? If so, freeze it or move the change to its own step.
9. Are all annotation comments (!callout, !mark-start, !mark-end, !tooltip, !diff) using the correct comment syntax for their context (// in JS/TS, {/* */} in JSX) AND placed in valid positions? (No {/* */} between element attributes or as bare ternary expressions.)
10. Is all code valid syntax? No bare ... as expressions, props, or arguments — use comments to collapse code instead.

Fix any violations before moving to the next file.

Phase 4: Structure the Narrative Around the <Walk>

Now assemble the full MDX file. The <Walk> blocks are the backbone — the prose exists to connect and contextualize them.

Start with the problem statement: One sentence. What's broken or missing, where it shows up, why the change is needed. This comes before any code.

Then alternate between <Walk> and prose transitions:

• Between <Walk> blocks (different files), add a short prose transition if the connection isn't obvious. State the intent ("now wire the limiter into the route") not the implementation details the reader is about to see.
• If a non-trivial flow connects two files, consider a Mermaid diagram (``` `mermaid ```) — but only if multiple actors or branches make prose hard to follow. For linear flows, prose is better.

Prose style: Short sentences. Front-load the key point. Bold sparingly for scanning. Inline code as *symbol*. End with a colon when directly introducing a code block. Don't repeat what's visible in the snippet.

Comment syntax for annotations: TS/JS → //!callout[/regex/m], JSX → {/*!callout[/regex/m] */}. Annotations (!mark-start/!mark-end, !callout, !diff, !tooltip) only work inside <Walk>-managed code blocks. Outside them, use standard fenced code blocks with short prose.

Close with completeness: The final steps should show the complete working mechanism.

Finishing touches

Add frontmatter at the top of the file. Get the current timestamp by running date +%Y-%m-%dT%H:%M:%S:

---
title: "Human-readable session title"
date: "YYYY-MM-DDTHH:MM:SS"
version: "<version from this skill's frontmatter>"
---

Check for missing context. Re-read the map and the narrative side by side. If there's any important information from the map that didn't make it into the narrative — a key decision, a tradeoff, a non-obvious reason for a change — find the right place and way to include it.

Phase 5: Append Raw Thought Process

Important: do this as you work, not after. During Phases 1–4, whenever you're thinking through a decision — ordering files, splitting steps, deciding what to cut — write those thoughts down immediately in a scratch area. Don't wait until the end to reconstruct your reasoning; that produces rationalization, not the real process. Log when you transition between phases (e.g., "--- moving to Phase 3 ---") so the decision trail is easy to follow.

After the hike file is complete, append all of those raw notes to the end of the file inside a fenced code block so raw content doesn't break MDX parsing:

## Thought Process

(raw notes here)

This should be unedited stream-of-consciousness — the messy real-time reasoning, not a clean retrospective. Include false starts, uncertainties, things you almost did differently. The goal is to preserve the actual decision trail so the author can later understand what really happened during generation.

Phase 6: Store and Display

Before writing, check for secrets. Scan the full output for API keys, tokens, passwords, credentials, or private URLs. Redact or omit any that appear.

Write to .hike/<descriptive-slug>.mdx. Derive the slug from session content (e.g., add-dark-mode-toggle.mdx, fix-auth-redirect-loop.mdx, refactor-api-client.mdx). Create .hike/ if it doesn't exist. If the file already exists, append -2, -3, etc. (e.g., add-dark-mode-toggle-2.mdx).

Skip this step in non-interactive mode (Claude Code's -p/--print flag) — there's no browser to open.

Otherwise, run npx @code-hike/hike@^<version> <filename> in the background to open a browser preview. <version> is the version from this skill's frontmatter, prefixed with ^ for compatible versions. Use run_in_background: true so the user can review at their own pace without blocking the conversation. Run the command in the user's current working directory (not inside .hike/).

npx @code-hike/hike@^<version> <slug>.mdx