🛠 MAJI Skills

Stop repeating yourself to Claude.

Four patterns that make Claude listen the first time. Save ~75% tokens. Ship without rework. Free.

⬇️ Install in 30 seconds · ⭐ Star this repo · 🔗 Share on LinkedIn

🎯 Why You Should Use MAJI Skills

Every Claude Code session, you're fighting the same battles:

😩 Claude extrapolates beyond what you asked — fixes the typo AND refactors three files you didn't want touched
💸 Token bills creep up because Claude responds with paragraphs of preamble before the actual answer
🔁 You have to re-explain your working style every session — Claude forgets you want bullets, not essays
🚨 Claude treats "let's discuss this" as "let's build it now" — and you delete 11 commits to recover
😤 When you say "that's not right", Claude tries variant 2, variant 3 — instead of stopping to ask what you actually want

MAJI Skills fix all of this with a discipline layer that activates on every prompt.

What you get in 30 seconds:

✅ Claude recognises outcome-first prompts and skips the padding
✅ Decision modes — discuss / build / pause / pivot — eliminate ambiguity
✅ Frustration signals halt Claude instead of doubling down
✅ Pre-Action Gate prevents destructive surprises (force pushes, scope creep)
✅ ~75% token savings per exchange — banned padding phrases, compression modes, table-first output
✅ Works on any project, any language, any stack
✅ Free, AGPL v3 (strong copyleft), no signup, no telemetry

→ Install now · → See real before/after examples

✨ Skills in this Repo

5 core skills covering the highest-frequency dev workflows — chat, explain, debug, review, commit. Plus 5 experimental skills for less-frequent use cases.

Core Skills (5)

Skill	Use case	Frequency	Tier
maji-mode	Foundation discipline (every Claude session)	Always	🔴 Tier 1 — Install first
maji-commit	Auto-generate git commit messages	Daily	🟠 Tier 2 — High daily value
maji-explain	Fast layered explanations of code/concepts	Daily	🟠 Tier 2
maji-debug	5-step disciplined debug protocol	Daily	🟡 Tier 3 — Per task
maji-review	Severity-grouped code review	Daily	🟡 Tier 3

Experimental Skills (5)

These exist but aren't validated by real-world use yet. See experimental/ folder.

maji-test · maji-doc · maji-refactor · maji-summary · maji-todo

A skill graduates from experimental to core when daily/weekly use is proven and external feedback validates value.

Recommended Install Order

flowchart LR
    T1[🔴 Tier 1<br/>Install first] --> M[maji-mode]
    T2[🟠 Tier 2<br/>High daily value] --> C[maji-commit]
    T2 --> E[maji-explain]
    T3[🟡 Tier 3<br/>Per task] --> D[maji-debug]
    T3 --> R[maji-review]

    style T1 fill:#8B1A1A,stroke:#0A0A0A,color:#FFFFFF
    style T2 fill:#A88838,stroke:#0A0A0A,color:#0A0A0A
    style T3 fill:#0A0A0A,stroke:#0A0A0A,color:#FFFFFF
    style M fill:#F5F1EA,stroke:#0A0A0A
    style C fill:#F5F1EA,stroke:#0A0A0A
    style E fill:#F5F1EA,stroke:#0A0A0A
    style D fill:#F5F1EA,stroke:#0A0A0A
    style R fill:#F5F1EA,stroke:#0A0A0A

Quick start: Install Tier 1 first (maji-mode). Add Tier 2 within first session. Add Tier 3 within first week.

How Skills Compose

maji-mode is the foundation. Other skills inherit its discipline (no preamble, banned phrases, action-verb endings).

flowchart TD
    M([maji-mode<br/>foundation])
    M --> C[maji-commit]
    M --> E[maji-explain]
    M --> D[maji-debug]
    M --> R[maji-review]
    R -.composes with.-> D

    style M fill:#8B1A1A,stroke:#0A0A0A,color:#FFFFFF
    style C fill:#A88838,stroke:#0A0A0A,color:#0A0A0A
    style E fill:#A88838,stroke:#0A0A0A,color:#0A0A0A
    style R fill:#F5F1EA,stroke:#0A0A0A
    style D fill:#F5F1EA,stroke:#0A0A0A

`maji-mode` — Claude Collaboration Discipline

A discipline layer between you and Claude. Activates 4 patterns that make Claude listen properly the first time — saving tokens, reducing rework, and preventing scope drift.

→ Read the full skill

`maji-commit` — Auto-generate Commit Messages

Reads your git diff and writes a clean commit message in conventional commits format. No manual typing, no overthinking the wording.

You stage changes → type "maji-commit" → Claude outputs:
feat(auth): add Google OAuth flow

Composes with maji-mode for terse, no-preamble output.

→ Read the full skill

`maji-explain` — Fast, Layered Explanations

Explains code, libraries, errors, or concepts in 3 layers: 1-sentence answer → 5-bullet detail → deep-dive only on request. No "let me walk you through this".

useEffect runs side effects after a React component renders,
with optional dependency tracking.

Key points:
• Fires after the DOM updates, not during render
• Dependency array controls re-runs
• Return a function to clean up
• Gotcha: stale closures from missed deps

Want deeper detail on dependency arrays or cleanup timing?

→ Read the full skill

`maji-debug` — Disciplined Debug Protocol

Turns Claude into a debug partner who slows down to think, not a guess-and-check refactor machine. 5-step protocol: read symptom → form hypothesis → verify → Pre-Action Gate → minimum-change fix.

SYMPTOM    TypeError on login.ts:14 — user.email undefined.
HYPOTHESIS API response shape changed; user nested under data.
VERIFICATION Confirmed in api/types.ts — LoginResponse wraps user.
FIX        response.user.email → response.data.user.email
WHY        Aligns with current type. No other consumers.

→ Read the full skill

`maji-review` — Disciplined Code Review

Reviews a PR, diff, or file and outputs structured feedback grouped by severity. No "Great work overall!" preamble. No trailing "let me know if you'd like to discuss". Just findings.

3 findings: 🔴 2 critical, 🟡 1 important.

🔴 Critical (2)
──────────────
src/auth.ts:3 — SQL injection via string interpolation
src/auth.ts:4 — Plaintext password comparison

🟡 Important (1)
──────────────
src/auth.ts:5 — Static token, not user-bound

Verdict: Block merge. Fix critical findings first.

Composes with maji-mode for evidence-first output.

→ Read the full skill

🧊 Status: Validation Phase

The 5 core skills are frozen for 30 days of real-world validation (started 7 May 2026). No new skills will be added during this window. Existing experimental skills stay in experimental/.

A skill graduates from experimental to core when:

Daily/weekly use proven by repo owner
1+ external contribution (PR, issue, testimonial)
30+ days of validation

After validation, decisions made based on real signal — not speculation.

🧠 How `maji-mode` Works

Click to expand flow diagram

flowchart TD
    A[User Prompt] --> B{maji-mode<br/>active?}
    B -->|No| Z[Standard Claude<br/>response]
    B -->|Yes| C[Read prompt fully]
    C --> D{Identify<br/>Decision Mode}

    D -->|Discuss| E[Plan + ask<br/>clarifying questions]
    D -->|Build| F[Pre-Action Gate]
    D -->|Pause| G[Stop. Wait.]
    D -->|Pivot| H[Strip + redirect.<br/>NO iteration.]

    F --> I{All 4 gate<br/>checks pass?}
    I -->|No| J[ASK before acting]
    I -->|Yes| K[Execute]

    E --> L[Output]
    K --> L
    G --> L
    H --> L
    J --> L

    L --> M{Frustration<br/>signal detected?}
    M -->|Yes| N[STOP. Reconsider.<br/>Don't iterate.]
    M -->|No| O[End with action verb.<br/>Offer next step.]
    N --> C

    style A fill:#F5F1EA,stroke:#0A0A0A,color:#0A0A0A
    style B fill:#A88838,stroke:#0A0A0A,color:#0A0A0A
    style D fill:#A88838,stroke:#0A0A0A,color:#0A0A0A
    style F fill:#8B1A1A,stroke:#0A0A0A,color:#FFFFFF
    style I fill:#8B1A1A,stroke:#0A0A0A,color:#FFFFFF
    style M fill:#8B1A1A,stroke:#0A0A0A,color:#FFFFFF
    style N fill:#8B1A1A,stroke:#0A0A0A,color:#FFFFFF
    style J fill:#A88838,stroke:#0A0A0A,color:#0A0A0A

The 4 patterns:

Outcome before context — say what you want, Claude responds; rationale optional
Mode-tagged prompts — discuss / build / pause / pivot — Claude treats them differently
Frustration recognition — when you signal "wrong direction", Claude halts instead of doubling down
Pre-action confirmation — destructive or scope-expanding actions need explicit OK

📊 Decision Modes Visualised

Click to expand decision modes diagram

flowchart LR
    P[User Prompt] --> M{Decision Mode}
    M -->|"discuss first<br/>propose plan<br/>what do you think"| D[💬 DISCUSS<br/>Plan + clarify]
    M -->|"proceed<br/>ship<br/>approve"| B[⚙️ BUILD<br/>Execute]
    M -->|"hold on<br/>wait<br/>pause"| P2[⏸️ PAUSE<br/>Stop & wait]
    M -->|"drop X<br/>start over<br/>change direction"| PV[🔄 PIVOT<br/>Strip + redirect]

    style P fill:#F5F1EA,stroke:#0A0A0A
    style M fill:#0A0A0A,stroke:#0A0A0A,color:#FFFFFF
    style D fill:#A88838,stroke:#0A0A0A,color:#0A0A0A
    style B fill:#8B1A1A,stroke:#0A0A0A,color:#FFFFFF
    style P2 fill:#3A3A3A,stroke:#0A0A0A,color:#FFFFFF
    style PV fill:#A88838,stroke:#0A0A0A,color:#0A0A0A

🎬 Real Before/After Examples

📌 Example 1 — Building a Feature

sequenceDiagram
    actor U as User
    participant C as Claude
    rect rgb(245, 220, 220)
    Note over U,C: WITHOUT maji-mode
    U->>C: "Hey, I'm thinking maybe we could add a<br/>search feature, what do you think?"
    C->>U: "Great idea! Several considerations:<br/>algorithms, indexing... should I explore first?"
    U->>C: "Sure, I guess..."
    C->>U: "I found these options... which sounds best?"
    Note over U,C: 3 messages, nothing shipped
    end

    rect rgb(220, 245, 220)
    Note over U,C: WITH maji-mode
    U->>C: "Add search to dashboard. Discuss first."
    C->>U: "Client-side filter or full-text DB?<br/>Recommend client-side. Approve?"
    U->>C: "Approve"
    C->>U: ✅ Implementation done
    end

Saved: ~1500 tokens, 3 round-trips, mental overhead.

📌 Example 2 — The "Proceed" Trap

Without maji-mode

User: "Let's proceed phase by phase, plan stories and epics"
Claude: [scaffolds 10 stories, runs npm install, commits 11 times]
User: "Wait, what are you doing?" → has to delete everything.

With maji-mode

User: "Let's proceed phase by phase, plan stories and epics"
Claude: "Reading 'proceed' in discussion context — continuing to plan, not building yet."
No surprise commits.

Saved: ~3 hours and a deleted repo.

📌 Example 3 — Aesthetic Mismatch

Without maji-mode

User: "That doesn't look right"
Claude: "Let me try a different color scheme..." [makes 5 changes]
User: "Still wrong"
Claude: "Let me try yet another approach..." [user rage-quits]

With maji-mode

User: "Not the right vibe"
Claude: "STOP — pivoting completely instead of iterating. Show me 1-2 reference sites with the look you want, or describe in 5 words."
User: "Linear style — minimal, dark, generous spacing" → clean pivot.

Saved: Massive frustration + 5+ wasted attempts.

📌 Example 4 — Destructive Action

Without maji-mode

User: "Clean up old branches"
Claude: [force-deletes branches, resets remote tracking, some unmerged work disappears]

With maji-mode

User: "Clean up old branches"
Claude: "Pre-Action Gate: destructive op detected. Found 12 branches. Show list before delete?"
User reviews → surgical cleanup, no surprise loss.

Saved: Lost work prevention. Trust preserved.

📌 Example 5 — Token Burn

Without maji-mode (~200 tokens of padding alone)

"Great question! That's a really thoughtful approach. Let me think through this carefully and consider the various angles. I appreciate you bringing this up. Here's what I'm thinking..."

With maji-mode

"3 options. Recommend (b). Reason: ships fastest. [table]"

Saved: Across 50 messages = ~10,000 tokens = real $.

💸 Token Economy Math

xychart-beta
    title "Token Cost Per Session"
    x-axis ["Without maji-mode", "With maji-mode", "+ /jimat mode"]
    y-axis "Tokens per exchange" 0 --> 700
    bar [600, 150, 80]

Item	Without `maji-mode`	With `maji-mode`	+ `/jimat` mode
Avg prompt+response	~600 tokens	~150 tokens	~80 tokens
Savings per exchange	—	~75%	~87%
100-message session	—	~45,000 tokens saved	~52,000 tokens saved
Power user (5 sessions/day)	—	~$20/day saved on Pro tier	~$23/day saved

📦 Installation

Option 1 — User-level (recommended, applies to all projects)

mkdir -p ~/.claude/skills
git clone https://github.com/Ijam18/MAJI-Skills.git
cp -r MAJI-Skills/maji-mode ~/.claude/skills/

Option 2 — Project-level (single project only)

mkdir -p ./.claude/skills
git clone https://github.com/Ijam18/MAJI-Skills.git
cp -r MAJI-Skills/maji-mode ./.claude/skills/

Option 3 — Always-on (loaded every session, every project)

Append the contents of maji-mode/SKILL.md to your ~/.claude/CLAUDE.md file:

git clone https://github.com/Ijam18/MAJI-Skills.git
cat MAJI-Skills/maji-mode/SKILL.md >> ~/.claude/CLAUDE.md

▶️ Usage

After installation, invoke at the start of any Claude Code session:

/maji-mode

Or it auto-activates via Claude's skill discovery if installed at user-level.

To verify it's active, your prompts should feel:

More direct (less preamble)
Better at scope (Claude asks before extrapolating)
Token-aware (shorter responses by default)

👥 Who benefits most

User type	Benefit fit
🚀 Solo founders + indie hackers	High
🌊 Vibe-coders (no formal CS)	Highest
👨‍💻 Senior devs using AI as pair programmer	High
🎨 Designers using Claude for code	High
🎓 Teaching contexts	High
🛠 Casual / weekend tinkerers	Lower (overhead may exceed benefit)

❓ FAQ

How do I verify maji-mode is actually active?

After install, type verify maji-mode in your Claude Code session. You should get a one-line confirmation showing 4 active patterns + current compression mode. If Claude responds with generic text instead, the skill isn't loaded — re-check install path.

Does this slow Claude down?

No. The patterns reduce response length, which means Claude streams faster. You'll feel it especially on /dry and /jimat modes.

Will this work with Cursor / Continue / other AI tools?

The skill format is Claude Code-specific, but the patterns are universal. Copy maji-mode/SKILL.md content into Cursor's .cursor/rules or Continue's system prompt and most patterns transfer. Decision modes, frustration signals, and banned phrases work with any LLM.

Can I customize the patterns? I want some but not others.

Yes — maji-mode/SKILL.md is just text. Fork the repo, edit the file, install your version. Common customizations:

Remove banned phrases you don't mind (some users like "Great question!")
Adjust compression default (some prefer slightly more verbose by default)
Add your own magic phrases vocabulary
Translate to your language

Does this work on Claude Pro / Max / API?

Works on any tier that supports Claude Code skills. Token savings matter most on metered API or hitting Pro tier limits — that's where the 75% reduction translates to real $$$.

How do I uninstall?

Delete the skill folder:

rm -rf ~/.claude/skills/maji-mode

Or for project-level: rm -rf ./.claude/skills/maji-mode. If you appended SKILL.md to your ~/.claude/CLAUDE.md, edit that file to remove the section.

Does this send any data anywhere? Telemetry?

No. The skill is a static markdown file Claude reads locally. Zero telemetry, zero phone-home, zero tracking. AGPL v3 licensed — read the file, you'll see.

What if I miss the friendly tone?

The patterns reduce padding but Claude still gives quality answers. If you want some warmth back, invoke compression less aggressively — skip /jimat mode and stick with default. Or fork and adjust the banned phrases list to allow some pleasantries.

Will this conflict with my other skills?

Generally no — maji-mode defines response posture, not domain logic. It composes well with other skills (e.g., a code-review skill, a documentation skill). If you hit conflicts, file an issue with details.

I tried it and it feels too tight. What now?

Two options: (1) skip compression modes (/dry, /jimat) — default maji-mode is already terse but still readable. (2) Fork and remove the strict token discipline section — keep just the 4 patterns (outcome-first, decision modes, frustration signals, Pre-Action Gate) without banned phrases.

🤝 Contributing

PRs welcome. Patterns to add should be:

✅ Universal (not specific to one project / language / domain)
✅ Token-positive (saves tokens, doesn't cost more)
✅ Easy to explain in 1-2 sentences
✅ Backed by a real failure mode the pattern prevents

📜 License

AGPL v3 — strong copyleft. Use, modify, redistribute freely, but:

Modifications and derivatives must also be AGPL v3
If you run a modified version on a network server (SaaS), users must be able to access the source
Cannot bundle in proprietary/closed-source products
Must keep copyright notice and license text

See LICENSE.

🌱 About MAJI · Connect

MAJI (Malaysia Artificial Joint Institute) is an Applied AI movement that builds tools, runs workshops, and ships skills for people who want to use AI as a builder — not just a consumer.

Origin: Malaysia. Audience: anyone, anywhere.
Methodology: No Codes, Only Vibes — we don't teach coding, we teach building.

Questions, feedback, or sharing your experience? DMs open:

→ LinkedIn: Zarul Izham (Ijam)
→ Threads: @_zarulijam

Tag the repo when sharing — happy to feature builders using maji-mode in the wild.

🛠 MAJI Skills

Stop repeating yourself to Claude.

🎯 Why You Should Use MAJI Skills

What you get in 30 seconds:

✨ Skills in this Repo

Core Skills (5)

Experimental Skills (5)

Recommended Install Order

How Skills Compose

maji-mode — Claude Collaboration Discipline

maji-commit — Auto-generate Commit Messages

maji-explain — Fast, Layered Explanations

maji-debug — Disciplined Debug Protocol

maji-review — Disciplined Code Review

🧊 Status: Validation Phase

🧠 How maji-mode Works

📊 Decision Modes Visualised

🎬 Real Before/After Examples

💸 Token Economy Math

📦 Installation

▶️ Usage

👥 Who benefits most

❓ FAQ

🤝 Contributing

📜 License

🌱 About MAJI · Connect

Reviews (0)

`maji-mode` — Claude Collaboration Discipline

`maji-commit` — Auto-generate Commit Messages

`maji-explain` — Fast, Layered Explanations

`maji-debug` — Disciplined Debug Protocol

`maji-review` — Disciplined Code Review

🧠 How `maji-mode` Works