Optimise Claude Context

What this skill does

Runs a structured audit of the current Claude Code setup, identifies what's wasting tokens or missing, then applies fixes. Based on the three-tier context architecture: only essentials are preloaded, everything else is on-demand.

---

Step 1 — Audit current state

Read and check each of the following. Do not skip any — the report depends on all of them.

1a. CLAUDE.md

• Read CLAUDE.md (if it exists)
• Count lines and estimate tokens (words × 1.3 is a good approximation)
• Identify sections that duplicate content already in other files (e.g. config files, README, inline code comments)
• Flag any section over 50 lines that could be summarised or moved to docs/
• Flag if total estimated tokens exceed 800

1b. .claudeignore

• Check if .claudeignore exists
• If it exists, check whether common high-noise paths are covered: node_modules/, dist/, *.lock, coverage/, build artefacts
• If it does not exist, flag as missing — this is the highest-impact quick win

1c. docs/QUICK_REF.md

• Check if docs/QUICK_REF.md exists
• If it does not, flag as missing — essential commands and never-do/always-do rules should live here, not inline in CLAUDE.md

1d. Session-start hook

• Check if .claude/hooks/session-start.sh exists and is executable
• If it does not exist, flag as missing — saves Claude running orientation commands manually every session

1e. Slash commands

• List files in .claude/commands/
• Note any obvious gaps for the repo's workflow

1f. MEMORY.md (auto-memory)

• Run git rev-parse --show-toplevel to get the absolute project root path
• Construct the MEMORY.md path: take the absolute path, replace every / with -, then prepend ~/.claude/projects/ and append /memory/MEMORY.md
• Example: /Users/alice/workspace/my-project → ~/.claude/projects/-Users-alice-workspace-my-project/memory/MEMORY.md
• Check if the file exists at that path
• If it does not exist: note "not found / clean" and skip the remaining checks for this step
• If it does exist:
• Read the file contents
• Estimate token count: count all words in the file and multiply by 1.3
• Scan every line for keywords that commonly duplicate plugin or hook managed context: OpenSpec, SDD, sdd-planner, INDEX.md, proposal, design.md, tasks.md
• Flag the full text of any line or bullet that contains one or more of those keywords as a candidate for removal
• If no lines match: note "exists / no overlap detected"
• If lines match: collect and record each flagged line so it can be shown in the report and fix steps

1g. openspec/INDEX.md

• Check if the directory openspec/specs/ exists in the repo root. If it does not exist, skip this step entirely — make no mention of it in the report.
• If openspec/specs/ exists:
• Check whether openspec/INDEX.md exists
• If it does not exist: flag as HIGH impact — every exploratory session in this repo loads all spec files (~9,458 tokens wasted per session). Record status as "missing".
• If it does exist:
• Run find openspec/specs -name "spec.md" -newer openspec/INDEX.md to check whether any spec file has been modified more recently than INDEX.md
• If the command returns one or more files: flag INDEX.md as stale — list the files that are newer. Record status as "stale".
• If the command returns no output: INDEX.md is up to date. Record status as "present / up to date".

---

Step 2 — Present the audit report

Show a concise report before making any changes:

## Claude context audit — 

### CLAUDE.md
- Lines:  (target: ≤ 100)
- Estimated tokens:  (target: ≤ 800)
- Status: ✅ lean / ⚠️ oversized
- Sections to trim: [list if any]

### .claudeignore
- Status: ✅ exists / ❌ missing
- Missing patterns (if exists): [list]

### docs/QUICK_REF.md
- Status: ✅ exists / ❌ missing

### .claude/hooks/session-start.sh
- Status: ✅ exists / ❌ missing

### .claude/commands/
- Found: [list]
- Missing: [list if applicable]

### MEMORY.md (auto-memory)
- Status: ✅ not found / ✅ clean / ⚠️ overlapping entries found
- Estimated tokens:  (or N/A if not found)
- Overlapping entries: [list of flagged lines, or "none"]

### openspec/INDEX.md
- Status: ✅ present / ⚠️ stale / ❌ missing / ➖ no openspec (skipped)
- Stale specs (if stale): [list]
- Estimated saving if fixed: ~9,458 tokens per exploratory session

### Estimated token saving from fixes: ~ tokens per session

Ask the developer: "Should I apply all recommended fixes, or do you want to review each one first?"

Wait for their answer before proceeding.

---

Step 3 — Apply fixes

Apply only what was flagged. Do not change things that are already correct.

Fix: CLAUDE.md is oversized

The goal is a CLAUDE.md that loads fast and stays cached. It should contain:

• What this service does — 2–3 sentences max
• Essential commands — just the most-used ones, or a pointer to docs/QUICK_REF.md
• Critical rules — the 3–5 things that must always apply
• Architecture pointer — one line pointing to where the deep detail lives

Move any section over these bounds to docs/ and replace it with a one-line reference.

Never delete content — move it. If a section is being removed from CLAUDE.md, it must land somewhere (docs/QUICK_REF.md, a new docs/*.md file, or the relevant spec file).

Fix: .claudeignore missing or incomplete

Create or update .claudeignore with patterns appropriate to this repo's tech stack. Always include:

node_modules/
dist/
*.lock
*.log
coverage/

For any repo, check .gitignore — anything in .gitignore that's also a large directory should be in .claudeignore too.

Fix: docs/QUICK_REF.md missing

Create docs/QUICK_REF.md containing:

• Top 10 commands for copy-paste
• Never-do list (repo-specific rules that Claude must never break)
• Always-do list (conventions Claude must always follow)
• Key files table (the 5–8 most-referenced files in this repo)

Extract this content from CLAUDE.md if it exists there — don't duplicate it.

Fix: Session-start hook missing

Create .claude/hooks/session-start.sh that displays:

• Current git branch and last commit
• Any active in-progress state (e.g. open feature branches, active work items)
• Contextual tips based on recently changed files

Make it executable with chmod +x.

Fix: MEMORY.md has redundant content

Only apply this fix if step 1f flagged one or more overlapping entries.

Show the engineer the exact lines that will be removed:

The following lines in MEMORY.md were flagged as overlapping with context already managed by hooks or plugins:

1. 
2. 
   ...

Ask: "Should I remove these entries from MEMORY.md?"

Wait for the engineer's answer before making any changes.

• If confirmed: Remove only the flagged lines. Leave all other content exactly as it is.
• If declined: Leave MEMORY.md unchanged. Note in the Step 4 summary that the engineer chose to keep the overlapping entries.

Fix: openspec/INDEX.md missing or stale

Only apply this fix if step 1g flagged INDEX.md as missing or stale.

1. Find all spec files: find openspec/specs -name "spec.md" | sort
2. For each spec file found, read it and extract the text of the first ### Requirement: heading line (strip the ### Requirement: prefix to get just the heading text).
3. Derive the capability name from the spec file path: it is the directory name immediately under openspec/specs/ (e.g. openspec/specs/my-capability/spec.md → capability name is my-capability).
4. Build the index entries — one line per capability:

- ****: — \openspec/specs//spec.md\``

1. Write openspec/INDEX.md with the following structure:

``` # OpenSpec Index

```

1. If openspec/specs/ is empty or contains no spec.md files, write openspec/INDEX.md with the header comment and a note that no specs were found, and report zero entries generated.
2. Never delete an existing openspec/INDEX.md without immediately replacing it in the same operation.

---

Step 4 — Show before/after summary

After applying all fixes:

## Optimisation complete

| Item | Before | After |
|---|---|---|
| CLAUDE.md lines |  |  |
| CLAUDE.md est. tokens |  |  |
| .claudeignore | ❌ | ✅ |
| docs/QUICK_REF.md | ❌ | ✅ |
| session-start hook | ❌ | ✅ |
| MEMORY.md overlapping entries |  tokens /  entries |  tokens / "none" (or "kept — engineer declined") |
| openspec/INDEX.md | ❌ missing / ⚠️ stale | ✅ present (~9,458 tokens saved per exploratory session) |

Estimated saving: ~ tokens per session (~% reduction)

Content moved (not deleted):
- [section] → [destination file]

Suggested commit message:

chore(.claude): optimise context loading — trim CLAUDE.md, add .claudeignore, QUICK_REF and session hook

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jawadmjn
• Source: jawadmjn/claude-skills
• License: MIT
• Homepage: https://medium.com/@jawadnawaz/how-i-cut-claude-code-token-costs-by-66-per-session-ac525589b7bb

Install and usage instructions live in the source repository linked above.