Audit Project

Project Audit

When the user invokes /audit-project, perform a thorough health check of project documentation and conventions. This is a periodic maintenance task — not needed every session, but valuable before milestones, after intensive work periods, or when things feel out of sync.

0. Detect Project Type

Check the project's .claude/CLAUDE.md for a project-type: field:

• data-science — Full audit including all steps
• general (or no field found) — Skip steps marked [Data Science only]

0b. Mechanical pre-pass (deterministic — gather, don't judge)

Several checks below are pure structure, not judgment — gather them with bash up front and feed the results into the interactive review, so your attention goes to the hard calls rather than to what a regex settles:

wc -l .claude/CLAUDE.md .claude/*.md 2>/dev/null                 # §6 CLAUDE.md size + planning-doc sizes (§2)
ls .claude/*.md                                                  # §3 unregistered docs (diff against the Registry)
ls scripts/[0-9]*_* 2>/dev/null | grep -v '\.qmd$'              # §4b numbered non-.qmd violations
ls scripts/scratch/ 2>/dev/null                                 # §4b scratch leftovers
# §4b correspondence: numbered scripts/*.qmd vs outs/<num>_*  (orphans either way)

Hand these to Steps 2–6 as evidence. The cross-document judgment stays solo — drift, duplication, and convention pruning need one context holding CLAUDE.md + every planning doc at once; fanning that read out across isolated agents would destroy the cross-doc signal the audit exists to find. (Only the optional gated verification in §7b — refuting findings already formed — may fan out.)

0c. Report-only / scheduled mode

When invoked non-interactively — e.g. a headless weekly cron (claude -p) or when the user asks for "report only" — run the detection half only and change nothing:

• Do: §0b mechanical pre-pass, §1 drift detection, §2 automated checks, §2c session-log health, §3 unregistered docs/scripts, §4/§4b script conventions (data-science), §5 prune candidates, §6/§7 size checks → compile the §8 Summary and save it to .claude/audit_reports/.
• Skip: §2 interactive review, §2b STATUS_SUMMARY rebuild, and every other step that asks a question or edits a file. Make no file changes except writing the report; ask nothing.

The saved report is a worklist: the user later opens the project interactively and says "work through the audit findings," at which point the interactive/decision steps run with them. This is the mode the bundled jobs/weekly-audit.sh digest job uses across recently-active projects.

1. Cross-Check CLAUDE.md Against Planning Documents

The main project CLAUDE.md should not maintain script or file lists that duplicate planning documents. Check for staleness:

1. Read the Project Document Registry in the project CLAUDE.md
2. For each registered planning document, read it and compare against the main CLAUDE.md:
   • Are any scripts referenced in CLAUDE.md marked as legacy/inactive in planning docs?
   • Does CLAUDE.md have stale lists that duplicate planning doc tables?
3. For stale/duplicated lists, propose:
   • Replace with a pointer to the planning document (preferred — avoids future drift)
   • Update to match (only if a summary is genuinely useful in CLAUDE.md)

2. Planning Document Health (Interactive)

For each registered planning document, perform both automated checks and an interactive review with the user.

Automated checks

1. Uncollapsed completed phases — Are there completed phases with detailed content (>10 lines) that should be collapsed to 3-5 line summaries? Report line counts per phase.
2. Stale "Next Session" sections — Does a "Next Session" or "Open Tasks" section contain items that appear to have been done already (based on status tables or completed phases)?
3. Status table accuracy — Are phases marked "Not started" that have actually been worked on, or "In progress" for work that's clearly complete?
4. Overall size — Report line count and KB for each planning doc. Flag any over 300 lines as candidates for collapsing.

Interactive review

After automated checks, walk through each planning document with the user:

5. Relevance check — "Is [document] still relevant, or should we close/archive it?" Propose closing planning docs whose work is complete or permanently parked.
6. Priority review — For each active work stream or phase, confirm the current priority is still accurate. Update priorities based on user answers.
7. Cleanup — Mark completed phases, close out TODO items that are no longer planned, update status tables to reflect reality.

2b. Rebuild STATUS_SUMMARY.md

Check if STATUS_SUMMARY.md exists at the repo root.

STATUS_SUMMARY.md uses a unified format shared between Claude Code (/done) and Cowork (wrapup), so executive-assistant skills can scan all projects uniformly. See the /done skill's Step 2b for the exact format template.

If it doesn't exist — offer to create one

1. Ask: "This project doesn't have a STATUS_SUMMARY.md yet. Want me to create one?"
2. If yes, follow the initial creation workflow:
   • Scan all planning documents registered in the Project Document Registry
   • Build the Active Plans table interactively — walk through each registered plan with the user, confirm status (Active/Blocked/Paused/Complete) and next action
   • Populate People section (obligations in both directions) — ask the user
   • Populate Upcoming Tasks (NOW/THIS WEEK/SOON) from plans and user input
   • Populate Flags for Executive Assistant — ask the user
   • Start Recent Activity fresh (one entry for this audit)
   • Write the Current state as one specific sentence about where the project stands right now
3. If no, skip this step

If it exists — rebuild from scratch

1. Full scan of all planning documents registered in the Project Document Registry
2. Rebuild all sections based on the planning documents and the interactive review from Step 2
3. Flag drift — report any discrepancies found:
   • Plans marked "Active" that haven't been touched in weeks
   • Plans marked "Paused" that have new activity
   • Planning docs whose internal status disagrees with the STATUS_SUMMARY
4. Ask the user about ambiguous items before assigning statuses — "Is [plan] still Active, or should we mark it Paused?"
5. Preserve Recent Activity entries — don't overwrite historical entries. Add a new entry for this audit, trim to last 5.
6. Regenerate all other sections (Current state, Active Plans, People, Upcoming Tasks, Flags) from scratch based on current state

2c. Check Session Log Health

Check the Session Log section at the bottom of the project's .claude/CLAUDE.md:

1. Missing section — If no Session Log exists, flag it and offer to create one with a single entry summarizing the current audit session.
2. Stale "Next" items — Look at the "Next" bullets from recent entries. Are any clearly done already (based on planning docs, completed phases, or files that now exist)? Flag these as candidates for removal or note that they've been addressed.
3. Entry count — If more than 5 entries, trim the oldest. If fewer than 2, note that the log is thin (not an error, just informational).
4. Stale dates — If the most recent entry is more than 4 weeks old and the project is supposedly active, flag the gap.

3. Check for Unregistered Documents and Scripts

Documents

Scan .claude/ for .md files not listed in the Project Document Registry. Propose adding any missing ones to the appropriate category table.

Scripts

Check if any scripts in scripts/ are not tracked in any planning document. For untracked scripts that relate to ongoing work, suggest either:

• Adding them to an existing planning document
• Creating a new planning document with /new-plan

Data files [Data Science only]

Check if any significant data files were created but not added to the Canonical Data Files Registry.

4. Script Lifecycle Status [Data Science only]

For projects using the script organization conventions:

1. Missing status fields: Check .qmd scripts for missing status: YAML frontmatter
2. Status/reality mismatch: Flag scripts marked finalized that have been modified since their last commit, or development scripts that appear complete
3. Planning doc sync: Verify YAML frontmatter status matches the relevant planning document's tracking table

4b. Script Convention Compliance [Data Science only]

Check the scripts/ directory for structural convention violations:

Numbered non-.qmd files

Scan scripts/ (not subdirectories) for files matching [0-9]*_* that are NOT .qmd:

• .R, .py, .Rmd files with number prefixes are violations
• Skip: scripts/old/, scripts/scratch/, scripts/exploratory/, unnumbered files (legacy)
• For each violation, report it and suggest: convert to .qmd, move to R//python/ as helper, or archive to scripts/old/

Scratch folder

Check if scripts/scratch/ exists and has files:

• If non-empty, warn — these are working files that should have been consolidated
• List the files and suggest consolidation targets

Letter suffix consistency

For scripts with letter suffixes (e.g., 15a_, 15b_, 15c_):

• Verify all scripts in a lettered set share a single output directory (outs/XX_topic/, not outs/XXa_topic/)
• Check that the a script exists (letters imply sequence)

Script-output directory correspondence

For each numbered .qmd:

• Extract the number prefix
• Verify outs/{number}_{topic}/ exists (lettered scripts share one dir by number)
• Flag scripts with no output dir (may be unrun status: development — note, don't error)

For each output directory in outs/ with a number prefix:

• Verify a corresponding script exists
• Flag orphaned output directories

5. Prune Self-Evident Conventions

Review existing convention entries in the project's .claude/CLAUDE.md. For each, ask: Is this now self-evident from the codebase?

A convention is self-evident when:

• The pattern is consistently established across the project's files
• Claude would infer it from reading any few files
• Removing the entry wouldn't cause Claude to do the wrong thing

Propose removing redundant entries. Show which entries would be pruned and why.

Be conservative — only prune conventions that are truly obvious from the code. When in doubt, keep the entry.

6. Project CLAUDE.md Size Check

Report the current size of the project CLAUDE.md (line count, approximate KB). If it's growing large (>300 lines), suggest:

• Moving detailed sections into planning documents
• Extracting complex conventions into skills
• Replacing inline tables with pointers to planning docs

7. User-Level Health Check

Quick checks on user-level configuration (lightweight — skip silently if everything is fine):

1. MEMORY.md size — Check ~/.claude/projects/*/memory/MEMORY.md for the current project. If over 200 lines, flag for pruning.
2. User CLAUDE.md size — Check ~/.claude/CLAUDE.md line count. If over 150 lines, suggest extracting content into skills.
3. Skills audit — For a deep review of skill quality, structure, and redundancy, run /audit-skills. Skip skill-level checks here.

7b. Optional: Adversarial Verification & Completeness (gated — requires the Workflow tool)

Gated — optional enhancement, not a requirement. The audit is complete and valid without this phase. Run it ONLY when both hold:

• the Workflow tool is available in this environment — it is not present in every Claude Code setup (lab-plugin users on a standard install will not have it), and
• at least one opt-in signal: ultracode is on, the user explicitly asked for thorough / adversarial / "double-check" verification, or you are producing a saved handoff report for another session.

If the Workflow tool is absent or no opt-in signal is present, skip this entire phase and proceed with your solo findings. Never block, delay, or weaken the audit waiting on a capability the environment may not have, and don't mention the workflow to users who can't run it.

When the conditions hold, fan out a short verification workflow over the drift / staleness findings you've already gathered solo (Steps 1–7). This is well suited to a large project with many planning docs, where one context can't hold every doc at once:

1. Adversarially verify each finding — one agent per finding (or per planning doc), each re-reading the actual CLAUDE.md / planning doc / file and prompted to refute: confirm the stale reference / duplication / orphaned file / status mismatch is real with exact locations, or reject it. Catches "looks stale but isn't" before you propose edits.
2. Completeness critic — one agent re-scanning the Project Document Registry, .claude/, and scripts/ fresh, told what you already found, hunting only for what you missed: unregistered docs/scripts, orphaned output dirs, status/reality mismatches, stale "Next" items already done.
3. (Handoff mode) draft concrete edits — literal replacement text per confirmed finding.

Fold the verified / refuted / newly-found findings into the Summary, and note that an adversarial pass was run. Do the read and cross-referencing of Steps 1–7 solo; only this verification pass over already-formed findings may fan out. This is the "adversarial verify + completeness critic" pattern from the Workflow tool's quality-patterns guidance.

8. Summary

Report findings organized as:

• Issues found — things that need fixing (stale references, unregistered files, uncollapsed phases, etc.)
• Suggestions — optional improvements (convention pruning, size reduction)
• All clear — areas that checked out fine

Ask the user which findings they'd like to act on before making any changes.