normalize-whitespace

Normalize indentation of a large reference text document (e.g. README.md) to a consistent TAB style while keeping content byte-identical. Use when asked to fix inconsistent leading spaces, align command blocks, or clean indentation spread across hundreds of lines in a prose-heavy, single-fence document. Built from the 2026-09 instguid README.md normalization session.

Jeff Yang@j3ffyang

Install

openclaw skills install @j3ffyang/normalize-whitespace

Normalize whitespace in a text document

Removes inconsistent leading-space indentation (2/4/8-space outliers) by converting command lines and code blocks to a leading-tab style, while preserving content exactly and leaving verbatim/preformatted regions untouched. The guiding invariant: no character of content may change — only leading/trailing whitespace.

Rules (immutable)

Whitespace changes only. The final content-integrity check is git diff --ignore-all-space --ignore-blank-lines against the pre-change state: it must show only real, user-approved content edits (line merges etc.). If it shows anything else, a bug ate a character — stop and fix before continuing.
Classify every line before editing. Buckets: (a) command/code line → normalize to tab; (b) verbatim/prose/preformatted → leave byte-identical; (c) ambiguous → ask the user, never guess.
Never touch what is content, not noise: verbatim terminal captures ([root@... ~]#, mbp:~ jeff$), ASCII-art/flowcharts, GUI-navigation prose (panel > ...), and verbatim copies of config files (raidtab, XF86Config, /etc/config/network, XML). Prefer state your leave-as-is classes up front so the user can correct the list.
grep-ability is preserved. Word-repeat separators may truncate the last repeat (e.g. cachyos...cach) — intended; grep '^cachyos' still matches.

Procedure

1. Establish the file's conventions

Read the repo's AGENTS.md for its whitespace/wall rules (instguid's README: single ```bash fence, TAB indentation, 60-char walls).
State the leave-as-is classes aloud, then confirm the target style (tab vs space) before running anything that writes.

2. Scan & classify with a marker-level view

Use cat -A (^I = tab, $ = EOL) over suspicious regions, and drive bulk analysis from Python, e.g. list every line with leading spaces:

python

from pathlib import Path
lines = Path('README.md').read_text().split('\n')
for i, l in enumerate(lines, 1):
    if l.startswith(' ') and l.strip():
        print(i, repr(l[:80]))

3. Apply whitespace edits with Python read/replace/write

The built-in edit tool treats whitespace-only oldStrings as no-ops — it will not change " x" to "\tx". Do the writes in Python:

python

from pathlib import Path
p = Path('README.md')
lines = p.read_text().split('\n')
for i, l in enumerate(lines):
    if l.startswith('    '):          # exact old/new required
        lines[i] = '\t' + l[4:]
p.write_text('\n'.join(lines))

Longest-prefix-first. When stripping multiple indent depths, branch on the longest first:

text

if l.startswith('    '):  -> \t + l[4:]
elif l.startswith('  '):  -> \t + l[2:]

Matching the 2-space branch first silently leaves 2 spaces behind.

Prefer explicit exact-string maps over arithmetic slicing for anything except mechanical startswith(n) strips. A slicing-start bug (old[2:] vs old[1:]) can silently drop the first character (no → o) — dump repr() of the surrounding lines before and after any non-trivial transform, and re-run git diff after each batch.

4. Binary/UTF-8 lines break string matching

Lines containing non-ASCII bytes (e.g. openssl Salted__... demo output, OCR-corrupted art) fail literal str.replace/find. Switch to line-index transforms (enumerate) keyed on a neighbouring ASCII anchor (if 'SSH Tunnel not running' in l:) instead of matching the whole line.

5. Verify each band, then the whole file

After each section: re-scan with awk 'NR>=... && /^ +[^ ]/' README.md | cat -A until no unexpected leading-space lines remain.
When a line's destination depends on your earlier loop state (a variable reused across strides), re-verify the exact output lines — this is where content-loss bugs (e.g. wifi-menu -o overwritten) sneak in.
End of file: confirm the single code fence is intact (grep -n '^```' → exactly the opening line and the closing line) and the separator walls' reported width matches the spec.

Example: instguid README.md (2026-09)

Normalized ~600 lines across AIX/linux/security/ssh/mac/arch/pacman/nodejs/python sections:

4-space command lines → \t; nested script bodies → \t (tab + 4-space steps, matching the existing __complete_ssh_host() / createTunnel() style).
Script shebangs got tabbed with their bodies; heredocs (docker-completion), command lists, and ==> Formulae/==> Casks output headers tabbed.
Left byte-identical: [root@storage-2 ~]# / mbp:~ jeff$ / jeff@debian:~$ captures, ifconfig/glxinfo/brew output alignment, the OpenVPN MTU prose, GUI-nav prose, config copies (raidtab, XF86Config, OpenWrt network, fontconfig XML), the ASCII rubric.
Content-integrity diff stayed clean the whole way; two line-merges were the only real edits. One loop bug temporarily replaced wifi-menu -o with a stale value — caught by immediate re-scan and restored.