think-like-me
A personal Claude Code skill that stores your curated engineering rules and actively consults them before Claude drafts code. Complements Claude Code's passive auto-memory: memory surfaces facts when relevant, rules shape the approach up front.
What it does
When you ask Claude to build, modify, debug, or design a technical feature, the skill:
- Identifies the domain(s) of the task (
mobile, backend, ui, general).
- Reads the matching
rules/<domain>.md file(s).
- Applies any rules whose
When: keywords overlap with the task.
- Acknowledges the applied rule in one line before writing code so you can catch a misapplied rule early.
If no rule matches, Claude proceeds normally — no forced matches.
Three ways to add a rule
| Invocation | When to use | Flow |
|---|
/learn-this | Right after correcting Claude | Claude analyzes the correction, drafts a rule with Don't pre-filled from the wrong approach, asks for approval. |
/new-rule [description] | Adding a rule from scratch (not tied to a correction) | Claude parses your prose into the structured block, asks if Why: is missing, shows draft. |
| Auto-detect | You correct Claude without typing a command | Claude offers once: "Caught — want me to save this as a rule? (yes to draft, skip to drop)" — never pressures twice. |
All three end the same: draft → approve → dual-write to rules/<category>.md and ~/.claude/projects/<project>/memory/feedback_<slug>.md (with a pointer in MEMORY.md).
Rule format
### Short searchable title
**When:** domain + feature keywords + specific context
**Do:** the fix, generalized
**Don't:** what not to do (optional)
**Why:** the mechanism that makes the wrong approach fail
**Example:** short snippet (optional)
When / Do / Why are required. Why in particular matters — it lets Claude judge edge cases the rule doesn't literally cover. Full spec in references/rule-format.md.
File layout
~/.claude/skills/think-like-me/
├── SKILL.md ← AI-facing entry point (trigger, flow, dual-write)
├── DESIGN.md ← full design spec
├── PLAN.md ← implementation plan
├── docs/README.md ← this file
├── references/rule-format.md ← canonical rule block spec
├── agents/
│ ├── learn-this.md ← /learn-this procedure
│ └── new-rule.md ← /new-rule procedure
└── rules/
├── mobile.md ← iOS / Android / React Native / Expo
├── backend.md ← Laravel / API / server
├── ui.md ← styling / layout / icons
└── general.md ← cross-cutting (starts empty)
Seeded rules (day one)
Migrated from cross-project entries in your auto-memory:
Mobile (7): NativeWind crash in tab FlatLists · Expo prebuild required for new plugins · Native module rebuild on dev client · EAS --local /var symlink fix · npm --legacy-peer-deps workaround · Google OAuth PKCE + IdToken clash · .env.local override trap
Backend (2): Laravel custom middleware priority · Sanctum v4.3 no isExpired()
UI (1): No emojis — use icon library
General: empty — accumulates via /new-rule and /learn-this
Project-specific memories (e.g. eSIM Access orderNo fix, processing-tab work, Paddle setup) were intentionally not promoted — they stay in project-scoped memory where they belong.
Extending
- Add a rule directly:
/new-rule [your description including When, Do, Why]
- Teach from a correction: correct Claude, then
/learn-this
- Add a new category: just mention it in
/new-rule (e.g. "this is a devops rule") — Claude will propose creating rules/devops.md and ask to confirm
- Edit a rule: open the relevant
rules/<domain>.md in any editor; format stays the same
- Remove a rule: delete its block from the category file; if it had a mirrored memory entry, delete that too
Design notes and references
- Full design:
DESIGN.md
- Implementation plan:
PLAN.md
- Rule format spec:
references/rule-format.md
- Capture procedures:
agents/learn-this.md, agents/new-rule.md
