Three new task-style skills, each using Anthropic's context: fork + agent:
Explore so the work happens in an isolated subagent context and only a
short summary returns to the main session.
skill/claude-md-drift-audit/SKILL.md (51 lines):
Walks the last N days of git history (default 7) and flags CLAUDE.md
lines that reference deleted paths, renamed paths, or removed
dependencies from that window. Returns a punch list with file:line.
Read-only. Standalone invocation: /claude-md-drift-audit [days=7].
skill/claude-md-link-check/SKILL.md (51 lines):
Verifies every @path chain import and every relative markdown link
inside every CLAUDE.md resolves to an existing file. Returns broken
links with file:line. Read-only. Standalone: /claude-md-link-check
[path-glob].
skill/claude-md-dependency-rescan/SKILL.md (54 lines):
Re-detects the project's tech stack from package.json /
requirements.txt / pyproject.toml / go.mod / Cargo.toml and diffs
against every CLAUDE.md's Tech Stack section. Returns added /
removed / renamed per file. Read-only. Standalone:
/claude-md-dependency-rescan [manifest-path].
command/sync-claude-md.md:
- argument-hint and when_to_use mention --weekly.
- New Phase 0 (only when --weekly is passed): invoke the three skills
above in parallel via the Skill tool, aggregate findings into
## Weekly Audit Summary, then proceed to existing Phase 1.
- Without --weekly, Phase 0 is skipped entirely (no behaviour change
for normal sync runs).
.claude-plugin/plugin.json:
skills array now lists all five entries (existing two + three new).
Verified (6/6 smoke tests):
- Plugin manifest is valid JSON; 5 skills registered; all 8
referenced paths resolve on disk.
- Each new SKILL.md parses; context=fork, agent=Explore, both
description and when_to_use present, allowed-tools is a list,
body ≤ 60 lines.
- Each body is imperative (numbered steps present, not reference
material — fork-context requires explicit task instructions).
- Sync command body contains Phase 0, all three skill references,
parallel-invocation language, and Skill in allowed-tools.
- Skill name in frontmatter matches the directory name for every
new skill (so /claude-md-drift-audit etc. register correctly).
Wave 3 - adoption hardening. Patterns adapted (in original prose, with
attribution) from MIT-licensed shanraisshan/claude-code-best-practice.
Commands (command/enhance-claude-md.md, command/sync-claude-md.md):
- Add allowed-tools / disallowedTools / argument-hint / when_to_use so the
commands auto-suggest in the slash menu and avoid permission prompts.
- disallowedTools blocks WebFetch + WebSearch on both commands.
- Drop the previous broken hooks block (array-of-{matcher, commands} shape
did not match canonical schema; was never firing).
Skills:
- skill/karpathy-guidelines/SKILL.md: paths: glob over 23 code-file
extensions, so the guardrails auto-load only when editing source, not
markdown or data.
- skill/SKILL.md: model: haiku, effort: medium, paths: scoped to CLAUDE.md
+ AGENTS.md + .claude/rules/*.md so validator/generator passes run
cheaply without changing the user-facing model.
CLAUDE.local.md personal tier:
- skill/validator.py BestPracticesValidator now accepts filename=; any
*.local.md basename waives the 150-line cap.
- hooks/validate-claude-md.py reads the exempt suffix from hooks-config.
- .gitignore covers CLAUDE.local.md, **/CLAUDE.local.md,
.claude/settings.local.json, hooks/hooks-config.local.json.
Layered hook config:
- hooks/hooks-config.json: committed defaults
(validateClaudeMd.enabled/maxLines/exemptFilenameSuffix/exitCodeOnViolation,
stopAuditLine.enabled).
- hooks/validate-claude-md.py merges hooks-config.json +
hooks-config.local.json key-by-key; honours enabled=false (silent
exit 0), configurable cap, configurable exit code.
Stop audit hook:
- hooks/audit-claude-md.py walks the project tree, prints one stderr
line: total tracked / OVER cap / near cap (>=80%). Respects
stopAuditLine.enabled from config.
- hooks/hooks.json registers Stop event with matcher "".
Guardian fail-closed contract:
- agent/claude-md-guardian.md Safety & Validation section now explicitly
requires Skill-tool invocation (no inline paraphrase of SKILL.md),
abort on missing validated output, never auto-commit, and respect
local hook config.
Verified (8/8 smoke tests):
- Both commands parse with new fields and no broken hooks block.
- karpathy paths: 23 globs, includes .py/.ts/.go/.rs.
- skill model=haiku effort=medium with CLAUDE.md path scope.
- Validator: *.local.md (300 lines) -> pass; CLAUDE.md (300) -> fail;
legacy ctor without filename -> default behavior preserved.
- hooks-config.json valid; validateClaudeMd.enabled=true, maxLines=150.
- Hook validator: default rc=2 on bloated, rc=0 when local override
disables it, rc=0 on *.local.md (exempt).
- Stop hook entry present; audit script: rc=0 with "5 CLAUDE.md tracked".
- Regression: large-fullstack root still 52 lines with chain imports.
Closes the gap between ClaudeForge and the Anthropic docs at
code.claude.com/docs/en/memory (and its linked hooks / skills / plugins
pages). Five tightly-scoped changes, each verified by smoke test.
A. Fix guardian hook frontmatter shape (agent/claude-md-guardian.md)
The previous array-of-{event, commands} shape did not match the
documented schema (hooks: { EventName: [{ matcher, hooks: [{ type,
command }] }] }), so the guardian's hooks silently did not fire.
Rewritten to the canonical keyed-object shape with PostToolUse,
PreToolUse, SessionStart, and the new InstructionsLoaded event.
B+C. Plugin-level hooks/hooks.json + hooks/validate-claude-md.py
New deterministic enforcement of the 150-line cap. The validator
script reads the hook payload from stdin, extracts any referenced
CLAUDE.md path (supports both PostToolUse tool_input.file_path and
InstructionsLoaded path / file shapes), and exits 2 with stderr
feedback when the file is over 150 lines. Wired to both PostToolUse
on Write|Edit and InstructionsLoaded on every load_reason
(session_start, nested_traversal, path_glob_match, include, compact).
The cap is now enforced at every load *and* write, not only when the
guardian decides to run sync.
D. ContentGenerator.generate_rules_file() (skill/generator.py)
Emits path-scoped .claude/rules/*.md instruction files with name,
description, and paths: glob frontmatter. Claude loads these
lazily — only when accessing files matching the globs — so
file-type-specific guidance no longer has to live in the root
CLAUDE.md. Validates that paths is non-empty (ValueError otherwise).
E. AGENTS.md / .cursorrules / .windsurfrules interop
command/enhance-claude-md.md Phase 1 now lists which sibling
instruction files exist. ContentGenerator.generate_root_file() reads
project_context['existing_instruction_files'] and prepends an
## External Instructions section with @AGENTS.md (etc.) imports, so
repos already using other agent tooling can adopt ClaudeForge
without losing their existing instructions.
Smoke tests (all pass):
- Guardian hooks frontmatter parses as a dict with 4 events, each
carrying matcher + nested hooks array of {type, command} entries.
- hooks.json is valid JSON; PostToolUse matcher = Write|Edit;
InstructionsLoaded matcher covers all five load_reason values.
- validate-claude-md.py: small file -> rc 0, bloated file (200 lines)
-> rc 2 with stderr referencing the 150 cap, InstructionsLoaded
payload shape also handled, non-CLAUDE.md paths ignored, no stdin
-> rc 0.
- generate_rules_file emits valid frontmatter with paths glob and
raises ValueError when paths is empty.
- generate_root_file inserts @AGENTS.md and @.cursorrules imports
when existing_instruction_files lists them; omits the section
otherwise.
- Regression: large-fullstack root still 52 lines with chain imports
intact; all five sub-CLAUDE.md files in this repo still pass
validator (status = pass).
Docs:
- agent/CLAUDE.md updated to show the canonical hook shape and the
hook-driven validator wiring.
- skill/CLAUDE.md notes generate_rules_file and AGENTS.md interop.
- CHANGELOG.md documents all five changes under Unreleased.
Splits the 518-line root CLAUDE.md into a 70-line root plus four chained
sub-files (skill/, command/, agent/, docs/), all under the 150-line cap,
and prunes stale references uncovered by the audit.
- CLAUDE.md (518 -> 70): project overview, navigation with @-imports,
behavioural guidelines summary, naming reference, license link
- skill/CLAUDE.md (85): component flow, Python module architecture,
modifying modules, adding templates, tech-stack detection, quality
scoring, karpathy skill
- command/CLAUDE.md (38): slash command development, top-level .md install
layout, skill <-> command integration
- agent/CLAUDE.md (70): guardian frontmatter, hook responsibilities,
skill <-> agent integration, git triggers
- docs/CLAUDE.md (109): install verification, smoke test, manual e2e
tests, common operations, release process
Stale content removed:
- 'Dual Directory Structure' section (claude-md-enhancer/ legacy dir
does not exist on disk)
- '20-300 lines' / '300-line cap' references rewritten to 150
- '~/.claude/commands/enhance-claude-md/' bundle paths rewritten to
top-level '.md' files (matching the new installer layout)
- 'Three integrated components' updated to five
- 'Current Version: 2.0.0' updated to 2.1.0
- '6 examples' in sample_input.json claim corrected (file has 1 scenario)
Each sub-file opens with a markdown back-link and @../CLAUDE.md import;
root contains @-imports for all four sub-files. Validator passes (status
pass) for every file.
Turns ClaudeForge into an installable Claude Code plugin and adds the
missing pieces for a clean lifecycle: hard line-cap enforcement, modular
chaining via @path imports, a sync/cleanup command, and Explore-agent
delegation for project discovery.
- .claude-plugin/plugin.json: plugin manifest registering both skills,
both commands, and the guardian agent (installable via /plugin marketplace
add alirezarezvani/ClaudeForge && /plugin install claudeforge)
- skill/validator.py: MAX_RECOMMENDED_LINES 300 -> 150 (warning at 120)
- skill/template_selector.py: target_lines capped at 150 across all
team sizes (solo 75 / small 100 / medium 125 / large 150) so any
single CLAUDE.md fits within the cap; bigger projects spread content
across modular sub-files
- skill/analyzer.py: length thresholds, quality scoring, and recommendations
rebased on the 150 cap (was 300/400)
- skill/generator.py: modular root now emits @path imports next to the
human-readable links; every sub-CLAUDE.md gets a back-link header
pointing to ../CLAUDE.md (markdown + @import) for bidirectional chaining
- command/sync-claude-md.md: new /sync-claude-md command that inventories
every CLAUDE.md, prunes stale references, enforces the 150 cap by
splitting into sub-files, and repairs the root <-> sub chain
- command/enhance-claude-md.md: discovery phase now delegates the deep
codebase scan to the Explore subagent to keep context lean
- install.sh / install.ps1: each command in command/ installs as its own
~/.claude/commands/<name>.md (legacy bundle backed up on upgrade)
- skill/SKILL.md, CLAUDE.md, README.md, CHANGELOG.md: docs updated for
the plugin install path, sync command, and new line cap
Verified via smoke test: validator constants, template targets, generator
output line counts across 5 presets (all <= 150), context files with
backlinks, @-import chain in modular root, idempotent merge_with_existing,
validator status transitions at the new cap, analyzer quality differential,
and plugin manifest JSON shape with all referenced paths existing on disk.
Claude
Alireza Rezvani
Alireza Rezvani