Review a file, directory, or inline code against its activated principles. Supports explicit principle override with --with / @group / on syntax. Use when the user runs /dot-audit [target] to check code or docs against quality principles.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Review a file, directory, or inline code against its activated principles. Supports explicit principle override with --with / @group / on syntax. Use when the user runs /dot-audit [target] to check code or docs against quality principles.
Flemming N. Larsen (https://github.com/flemming-n-larsen)
license
MIT
generated-by
.principles
Audit
⛔ PREREQUISITE — CHECK THIS BEFORE ANYTHING ELSE
Before parsing arguments or doing any work, determine whether an explicit principle spec is present in $ARGUMENTS:
Explicit spec = $ARGUMENTS contains --with, one or more @-prefixed tokens, or on (space–on–space)
If no explicit spec is present, check whether scout-generated files exist from a previous /dot-scout run:
read .agents/principles-catalog/active.md; keep if it contains the generated-by marker (<!-- generated by /dot-scout)
if absent: glob .github/instructions/*.instructions.md and REVIEW.md (at the git root) — Copilot Code Review and Claude Code Review outputs; keep only files containing the generated-by marker
if none are found → STOP. Do not proceed. Respond only with:
⚠️ /dot-audit requires /dot-scout to have been run first.
Run /dot-scout to analyse the project and generate the principle files that this command needs, then retry.
Review a file, directory, or inline code against its activated principles. Core review runs in seven phases (1–7). Three optional gated phases (8–10) handle fix, commit, and PR — each requires explicit user approval before entry.
Phase 1 — Parse Arguments, Resolve Input, and Detect Artifact Type
1.1 — Parse Arguments for Explicit Principle Spec
Check $ARGUMENTS for an explicit principle spec using this precedence:
--with <spec> — if $ARGUMENTS contains --with, extract everything after --with as the spec; the text before --with is the target input.
@<group> token — if $ARGUMENTS contains one or more @-prefixed tokens, extract all @-prefixed tokens as the spec (space-joined); the remaining tokens form the target input.
<spec> on <target> — if $ARGUMENTS contains on (space–on–space), split on the first occurrence: left side is the spec, right side is the target input.
No spec — treat all of $ARGUMENTS as the target input (normal mode).
If an explicit spec was detected, record principle-spec and set explicit-mode: true. Otherwise set explicit-mode: false.
1.2 — Resolve Input
Determine what to review from the target input resolved in 1.1:
Empty (explicit-mode false) → respond "What would you like me to review?" and stop.
Empty (explicit-mode true) → use the current working directory as target.
File path → read that file.
Directory path → recursively glob all reviewable files; exclude binaries, lock files, node_modules, vendor, dist, build, .git, and build artifacts.
Inline code or text → use it directly.
1.3 — Detect Artifact Type
For the target file(s), detect the artifact type by reading .agents/principles-catalog/layers/artifact-types.yaml and matching against its type definitions. Match by file extension, filename, or path pattern in precedence order (infra before config for ambiguous YAML).
Record the detected type: code | docs | config | infra | schema | pipeline
If the target is a directory with mixed artifact types, note the mix; apply per-file type detection in Phase 6.
1.4 — Load Git Context
Output nothing during this phase.
After the target is resolved, attempt to load recent git history for the target files. This context is used by git-aware principles in Phase 5 and Phase 6.
Check whether a git repository is reachable (look for .git/ walking up from the target, same logic as Phase 2's .principles walk).
If reachable, run both commands against the target path:
git diff HEAD -- <target> — staged + unstaged changes relative to HEAD (current work in progress)
If the above produces no output: git diff HEAD~1 HEAD -- <target> — the most recently committed change
git log --oneline -5 -- <target> — recent commit history for the target
Store results:
$GIT_DIFF — the diff output (whichever command produced content, preferring uncommitted; empty string if none)
$GIT_LOG — the log lines (empty string if none)
If git is unavailable, the target is inline code, or no history exists: set both to empty string. Do not fail or warn — graceful degradation means git-aware principles fall back to snapshot-only review.
Phase 2 — Resolve Principles
Explicit mode (explicit-mode true):
For each item in the <principle-spec> (split on commas and spaces, trim whitespace):
Group match: look for .agents/principles-catalog/groups/<item-lowercase>.yaml. If found, read it and expand its principles list into the active set; recursively process any includes (abort on cycles).
Principle ID match: if no group file matched, add the item directly to the active set (case-insensitive).
No match: report "Unknown principle or group: <item>. Check available groups in .agents/principles-catalog/groups/." and stop.
Record source as: explicit: <principle-spec>. Skip Phase 3 and proceed to Phase 4.
Normal mode (explicit-mode false):
Fast Path — Per-Group Files
Before walking .principles files, check for per-group principle files emitted by /dot-scout.
Read .agents/principles-catalog/active.md; check for the <!-- generated by /dot-scout marker
If absent: glob .github/instructions/*.instructions.md and REVIEW.md (at the git root) — Copilot Code Review and Claude Code Review outputs; filter to files containing the generated-by marker
If any found: parse all - ID: Summary lines across all marked files (the ID is everything before the first colon)
Union all IDs → active principle set
Optionally cross-reference .agents/principles-catalog/index.tsv (each line: ID|LAYER|SUMMARY) to get Layer groupings for each active ID — use these layer assignments to annotate the audit header (e.g. show "Layer 1: N principles, Layer 2: M principles").
Record source as: per-group files (N files)
If no per-group files are found, proceed with the tree walk below.
Walk up from the target path to the git repo root (.git/) or max 10 levels, collecting every .principles file. Order: root → target.
If no .principles files found: skip to Phase 3.
Directives
Lines starting with : are configuration directives. Parse them before processing IDs:
:max_principles N — cap the total number of active principles to N. When trimming to fit:
Universal principles (from artifact-types.yaml) are always retained
Stack layer 1 principles are always retained
Layer 3 risk-elevated principles — next priority
Layer 2 context-dependent principles — lowest priority, dropped first
Seed — Universal + Stack Layer 1
Step 1 — Universal principles (active for ALL artifact types):
Read .agents/principles-catalog/layers/artifact-types.yaml → universal section. Add all listed IDs to the active set:
ID
Title
SIMPLE-DESIGN-REVEALS-INTENTION
Reveals intention
CODE-CS-DRY
DRY: Don't Repeat Yourself
CODE-CS-KISS
KISS: Keep It Simple
CODE-DX-NAMING
Name things by what they represent
ARCH-DECISION-RECORDS
Architecture Decision Records
CODE-CS-YAGNI
YAGNI: You Aren't Gonna Need It
Step 2 — Stack layer 1 (active for the detected artifact type):
Read .agents/principles-catalog/layers/<detected-type>/layer-1-universal.md. Add all principle IDs from the table in that file to the active set.
Process Each .principles File (root → target)
Skip blank lines and # comments.
:directive value → parse as a configuration directive (see above).
@group → read .agents/principles-catalog/groups/<group>.yaml, expand principles into the active set; recursively process includes (abort on cycles).
Bare ID → add to active set (case-insensitive).
!ID → add to exclusion set.
final_active = active_set MINUS exclusion_set (then apply :max_principles cap if set) · Source: .principles hierarchy (N files)
Phase 3 — Dynamic Detection (fallback)
Only if explicit-mode is false AND Phase 2 found no .principles files.
Layer 1 — Seed
Same as Phase 2 seeding: universal principles + stack layer 1 from .agents/principles-catalog/layers/<detected-type>/layer-1-universal.md.
Activate ALL matching contexts by scanning the target file(s) content for the signals listed in each context. For each matching context, add its activate principle IDs to the active set.
Layer 3 — Risk-Elevated
Check for .agents/principles-catalog/layers/<detected-type>/layer-3-risk-signals.yaml. If present, scan the target file(s) for the signals listed in each risk category. For each matching category, add its elevate principle IDs to the elevated set — violations of elevated principles are promoted one severity level (Low→Medium, Medium→High, High→Critical).
Record source as: dynamic detection (<type> stack)
Phase 4 — Load Principle Content
Per-group files fast path (source is per-group files (N files)):
Derive unique namespaces from the active principle ID prefixes. Use the longest-prefix match from this table:
ID prefix
Directory
CODE-SMELLS-*
code-smells/
SEC-ARCH-*
sec-arch/
CLEAN-ARCH-*
clean-arch/
SIMPLE-DESIGN-*
simple-design/
EFFECTIVE-JAVA-*
effective-java/
12FACTOR-*
12factor/
PIPELINE-*
pipeline/
CODE-API-*
code/api/
CODE-AR-*
code/ar/
CODE-CC-*
code/cc/
CODE-CS-*
code/cs/
CODE-DX-*
code/dx/
CODE-OB-*
code/ob/
CODE-PF-*
code/pf/
CODE-RL-*
code/rl/
CODE-SEC-*
code/sec/
CODE-TP-*
code/tp/
CODE-TS-*
code/ts/
CODE-*
code/
SOLID-*
solid/
DDD-*
ddd/
GOF-*
gof/
GRASP-*
grasp/
OWASP-*
owasp/
EIP-*
eip/
FP-*
fp/
A11Y-*
a11y/
INFRA-*
infra/
CONFIG-*
config/
SCHEMA-*
schema/
DOCS-*
docs/
DB-*
db/
CD-*
cd/
ARCH-*
arch/
PKG-*
pkg/
For each unique namespace, use the Read tool to load .agents/principles-catalog/principles/<namespace>/.context-audit.md, then filter entries whose ### ID is in the active set. Do not use bash, grep, or any shell command for this step — read the file and filter in your reasoning. Use the Principle and Violations to detect content in Phase 6.
If .agents/principles-catalog/ is not present, fall back to the standard loading below.
Standard loading (all other sources):
For each namespace in the active ID set, use the Read tool to load:
Filter entries whose ### ID is in the final active set. Do not use bash, grep, or any shell command for this step — read the file and filter in your reasoning. Use the Principle and Violations to detect content in Phase 6.
Filter to entries whose ### ID is in the final active set. Each entry contains one or more commands in this format:
- `command` | SEVERITY_HINT | description
Principles with entries in .context-inspect.md are "inspected". Principles without entries are "semantic-only" (handled entirely by LLM reasoning in Phase 6 Step 2).
5.2 — Execute Commands
For each inspection command:
Replace $TARGET with the actual path from Phase 1.
Run the command using bash. Commands may use git (e.g. git diff HEAD -- $TARGET | grep …) — this is valid; $GIT_DIFF from Phase 1.4 is pre-loaded context, but inspect commands run git directly against $TARGET as a pathspec.
If a command produces no output or fails (including because $GIT_DIFF is empty), skip silently.
5.3 — Build Pre-Scan Manifest
Group all hits by file. The result is the pre-scan manifest — a map of file → [{principle_id, severity_hint, line, match_text, description}].
Track two sets:
Inspected principles — those that had at least one command in .context-inspect.md (regardless of whether hits were found)
Semantic-only principles — all remaining active principles
Phase 6 — Review
Output nothing during this phase.
Reviewer Persona
You are a senior principal architect with 20+ years of experience. You have seen codebases rot from the same avoidable mistakes. You are direct, honest, and do not soften findings. You do not praise effort. You do not say "consider" when you mean "fix". You do not omit a finding because it feels impolite. If something violates a principle, you say so plainly and explain the concrete consequence of leaving it unfixed. You are not unkind, but you are not gentle either — your job is to make the code better, not to make the author feel better.
What you do NOT report:
Purely stylistic or opinionated preferences where reasonable engineers disagree and no real harm results (e.g. formatting choices, naming style debates, brace placement).
Theoretical violations with no plausible real-world consequence for this codebase.
"Could be slightly better" observations — only report genuine problems.
If you cannot articulate a concrete, real consequence of leaving a finding unfixed, do not report it.
Severity Calibration
Upgrade MEDIUM → HIGH when the violation will demonstrably harm maintainability, testability, or correctness at scale.
Do not downgrade a HIGH finding to MEDIUM to soften the report.
Never omit a finding because the surrounding code is "otherwise good".
Do not invent findings to appear thorough — fewer real findings is better than more opinionated ones.
Step 1 — Guided Review (pre-scan hits)
For each file in the pre-scan manifest:
Read the file (or at minimum ±10 lines around each hit).
For each hit, evaluate it against the principle's Violations to detect from Phase 4.
Confirm → record as a finding (use the severity hint as a starting point, adjust based on context; elevated → promote one level).
Dismiss → false positive, do not report.
Step 2 — Semantic-Only Review
Step 1 — Rank principles by relevance before reading any file.
Rank the active semantic-only principles by how directly they apply to this specific target. Use the artifact type and risk signals detected in Phase 1:
Security / reliability — OWASP-*, CODE-SEC-*, CODE-RL-*, SEC-ARCH-* — always highest priority when the target touches auth, payments, PII, concurrency, or public APIs.
Structural integrity — SOLID-*, CLEAN-ARCH-*, DDD-*, GRASP-* — prioritise when the target contains non-trivial business logic.
Context-specific — everything else, weighted by how directly it applies to what the file is doing.
Apply high-priority principles first and give them more scrutiny. Do not skip lower-priority principles, but spend proportionally more effort where the risk is higher.
Step 2 — Read every file collected in Phase 1. Apply only the semantic-only principles (those without inspection patterns). Do not substitute grep, search, or pattern-matching tools for reading — you must read and understand each file's logic, structure, and intent.
For each file, evaluate it against the semantic-only principle set appropriate to its artifact type.
For principles that are git-history-dependent (marked Audit-scope: limited — git in their principle file), include $GIT_DIFF and $GIT_LOG from Phase 1.4 as additional context alongside the file content. If both are empty, apply the principle as snapshot-only.
Step 3 — Opportunistic Findings
While reading files in Steps 1 and 2, if you encounter a clear violation of any active principle (including inspected ones not flagged by pre-scan), record it as a finding.
Recording Findings
For each violation found, record: principle ID, severity (Critical/High/Medium/Low, elevated → promote one level), absolute file path with forward slashes, line number, one sentence describing what is wrong, and a concrete fix grounded in the principle.
Phase 7 — Output
Step 1. Write audit-output.json to the repository root (where .git/ is) with this structure:
{"findings":[{"severity":"HIGH","principle_id":"DOC-PURPOSE","title":"one-line description","file":"C:/absolute/path/to/file.md","line":42,"description":"what is wrong","fix":"concrete fix"}],"summary":{"critical":0,"high":1,"medium":0,"low":0,"active_principles":["DOC-PURPOSE","CODE-CS-DRY"],"principle_source":".principles hierarchy (2 files)","artifact_type":"docs"}}
severity: CRITICAL, HIGH, MEDIUM, or LOW
file: absolute path, forward slashes; "" if unavailable
