consolidate

name	consolidate
description	Audit knowledge directories (.memory/, docs/, staging/, archive/) and Claude Code auto memory for stale, misplaced, or duplicated content. Generates an editable cleanup plan file, then executes approved actions after review. Use when: periodic knowledge cleanup, after project milestones, files feel scattered or disorganized, output staging feels cluttered, moving deliverables to permanent homes. Trigger phrases: consolidate knowledge, knowledge cleanup, audit knowledge directories, find stray docs, tidy up knowledge, triage outputs, review staging, clean up outputs.
disable-model-invocation	false
metadata	{"author":"Backchain","version":"1.0.0"}

Consolidate

Audit knowledge directories and Claude Code auto memory for stale, misplaced, or duplicated content. Produce an editable plan file the user reviews, then execute approved actions. Designed to complement, not replace, native auto-memory: this skill never modifies MEMORY.md directly and never deletes auto-memory entries — it focuses on graduation (promoting stable entries to permanent rules) and on knowledge directories outside auto-memory's scope.

Relationship with auto memory

Native Claude Code auto memory writes implicit learnings to ~/.claude/projects/<project>/memory/. This skill complements it:

Concern	Native auto memory	consolidate (this skill)
Capture implicit learnings	Yes	No — defer to native
Maintain MEMORY.md index	Yes (manual edits via /memory)	No — never modify directly
Graduate feedback → rules / CLAUDE.md	No	Yes — primary responsibility
Audit project knowledge directories	No	Yes — user-declared scan roots
Manage staging-directory lifecycle	No	Yes — age-based classification
Detect stray operational knowledge	No	Yes

Do not:

• Recommend DELETE or ARCHIVE for files inside ~/.claude/projects/<project>/memory/. Use the RULES or CLAUDEMD actions to graduate stable entries; deletion of the source file then happens as part of graduation, not as cleanup.
• Modify MEMORY.md.
• Touch any lock files in the auto-memory directory.

Scan roots

Before scanning, the skill needs a list of knowledge directories to inspect. If the user has not declared scan roots in their project's CLAUDE.md or in conversation, ask them. Sensible defaults to suggest:

.memory/
docs/
outputs/                                    # or whatever the project uses for staging
archive/                                    # or whatever the project uses for archive
~/.claude/projects/<project>/memory/

Skip these directory patterns unconditionally:

.git/                node_modules/          .venv/
.claude/             dist/                  build/
*.pdf  *.docx  *.xlsx                       # binary files — no content analysis

If a directory the user named does not exist, surface that in the report rather than failing silently.

Commands

Command	Purpose
/engram:consolidate or /engram:consolidate scan	Full audit: scan all declared knowledge dirs, generate plan file
/engram:consolidate staging	Quick triage of one designated staging directory with age classification + destination prompts (inline report, no plan file)
/engram:consolidate execute	Read the user-edited plan file and execute approved actions
/engram:consolidate status	Inline health summary: file count and age per declared directory, stray count

The skill is invoked as /engram:consolidate (the plugin-qualified skill name) followed by the optional subcommand. Natural-language phrasings like "consolidate knowledge" or "audit knowledge directories" trigger the same dispatch.

Scan algorithm

Step 1 — Discover

Walk each declared scan root recursively. Collect: path, age, size, content summary. Skip the unconditional-skip patterns above.

Step 2 — Classify each file

Auto-memory files (~/.claude/projects/<project>/memory/) use graduation-only classification:

1. MEMORY.md or any lock file → KEEP (auto-memory infrastructure — never touch)
2. Feedback memory, ≥14 days old, stable across sessions → RULES or CLAUDEMD (graduate)
3. Project memory, now obsolete (completed work, resolved decision) → DELETE (only after asking the user — graduating these to a finished-work archive is often better)
4. Otherwise → KEEP

All other files use the general classification (evaluate in order; first match wins):

1. System artifact (.DS_Store, .gitkeep) → DELETE
2. Superseded (newer version exists, or content captured in rules / CLAUDE.md / skills) → DELETE or ARCHIVE
3. Implementation note for completed work → DELETE
4. Point-in-time snapshot → DELETE if >30 days, ARCHIVE if ≤30 days
5. Misplaced (operational knowledge in wrong knowledge directory) → MOVE
6. Promotable (in-flight work ready for permanent home) → DOCS, RULES, CLAUDEMD, or SKILL
7. Otherwise → KEEP

Step 3 — Detect strays

A "stray" is operational knowledge (methodology, decisions, planning, standards, patterns, workflows) found outside the declared scan roots. Heuristics for "operational":

• Filename contains: guide, plan, strategy, standard, pattern, workflow, decision, adr
• Content contains methodology, process, architectural decisions, design rationale, planning, conventions, or reusable patterns

Skip directories that legitimately hold non-operational content (project deliverables, raw input data, execution logs, meeting records, the .claude/ configuration tree). When in doubt, surface in the report and let the user classify rather than guess.

Step 4 — Generate plan file

Write a plan to a path the user has designated for staging output (or <scan-root>/plans/consolidate-plan-YYYY-MM-DD.md if undeclared) using the template at ${CLAUDE_SKILL_DIR}/templates/plan-template.md. The plan is editable: the user changes the Action column to approve, override, or skip each row before running execute.

Action codes

Code	Effect
DELETE	Remove permanently
ARCHIVE	Move to the project's archive directory with mirrored source path
DOCS	Move to the project's docs/
MEMORY	Move to the project's .memory/
RULES	Move to .claude/rules/
CLAUDEMD	Incorporate into CLAUDE.md (present diff, require explicit approval)
SKILL	Move into a skill directory under .claude/skills/
MOVE	Relocate (destination given in the Reason column)
KEEP	No action
SKIP	Defer decision

Execution safety rules (/engram:consolidate execute)

1. Parse the edited plan; extract all rows where Action is not KEEP or SKIP.
2. Create destination directories first.
3. Execute moves before deletes (MOVE, ARCHIVE, DOCS, RULES, SKILL). Prevents data loss when a path appears in both a move and a delete row.
4. Execute deletes last.
5. CLAUDEMD rows present a proposed diff and require an explicit yes before applying.
6. Clean empty directories produced by moves and deletes — empty directories mislead future scans.
7. Verify final counts match the plan summary; report discrepancies.

Staging triage (/engram:consolidate staging)

A focused, quick-check variant for a single designated staging directory. Produces an inline report — no plan file. Age thresholds:

Age	Status	Default action
<7 days	Current	None
7–30 days	Aging	Suggest a permanent destination
>30 days	Stale	Suggest archive or move

For each file the user wants to act on, prompt for the destination. Do not guess destinations unless the user has declared a destination map in conversation or CLAUDE.md.

Status output

/engram:consolidate status produces inline text only: file count and age per declared directory, stray count with up to three example paths.

Error states

Condition	Response
No scan roots declared and user declines to declare	Print the request, exit cleanly
A declared scan root does not exist	Note in report, continue with the rest
Plan file already exists for today	Overwrite; note "overwrote earlier plan from today"
Designated staging directory empty	"No files found. Nothing to triage."
Designated staging directory missing	"Staging directory does not exist."

References

• ${CLAUDE_SKILL_DIR}/references/directory-registry.md — common knowledge-directory patterns and exclusion rules
• ${CLAUDE_SKILL_DIR}/templates/plan-template.md — plan file scaffold