Captures project-specific patterns and anti-patterns into the project's CLAUDE.md. Loaded by other skills (bug-killer, feature-dev, etc.) when they discover project-specific knowledge worth encoding for future sessions.
Installs into .claude/skills of the current project.
Are you the author of Project Learnings?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/sequenzia-project-learnings)
---
name: project-learnings
description: >-
Captures project-specific patterns and anti-patterns into the project's CLAUDE.md.
Loaded by other skills (bug-killer, feature-dev, etc.) when they discover
project-specific knowledge worth encoding for future sessions.
user-invocable: false
disable-model-invocation: false
allowed-tools: Read, Edit, Glob, AskUserQuestion
---
# Project Learnings
Capture project-specific patterns and anti-patterns into the project's CLAUDE.md. This creates a self-improving feedback loop where discoveries from debugging, development, and review make future Claude sessions smarter.
**CRITICAL:** Only project-specific knowledge qualifies. Generic programming advice does not belong in CLAUDE.md.
---
## Step 1: Evaluate Discovery
Determine if the finding qualifies as project-specific. The finding must pass at least ONE of these criteria:
| Criteria | Example That Qualifies | Example That Doesn't |
|----------|------------------------|----------------------|
| Would a developer unfamiliar with this project likely hit this issue? | "The `processOrder()` function expects amounts in cents, not dollars" | "Always validate function inputs" |
| Is this pattern specific to this codebase's architecture, APIs, or conventions? | "The `UserProfile` type has an optional `metadata` field that is always present at runtime" | "Use TypeScript strict mode" |
| Is it something Claude's training data wouldn't cover? | "Never call `db.query()` without the `timeout` option — the default is infinite" | "Use async/await instead of callbacks" |
**If NO to all criteria → STOP.** Do not add generic programming knowledge to CLAUDE.md. Return to the calling skill and report that no project-specific learning was found.
**If YES to any → proceed to Step 2.**
---
## Step 2: Read Existing CLAUDE.md
1. **Find the project's CLAUDE.md:**
- Check the repository root first
- If not found, check if there's a project-level `.claude/` directory
2. **Parse existing content:**
- Understand the existing structure, headings, and conventions
- Look for sections where this learning would fit (e.g., "Known Gotchas", "Bug Patterns", "Conventions", "Known Challenges")
- Check for duplicate or similar entries already present
3. **If a similar entry already exists → STOP.** Report to the calling skill that this knowledge is already captured. Do not create duplicates.
4. **Identify placement:**
- If an appropriate section exists, plan to add the entry there
- If no appropriate section exists, plan to propose a new section (e.g., `## Known Gotchas` or `## Project-Specific Patterns`)
- New sections should be placed after the main documentation sections but before appendices or settings
---
## Step 3: Format the Learning
Write a concise, actionable instruction following these rules:
**Format:**
- Use imperative form: "Always validate X before calling Y"
- Include the WHY: "...because the API returns dates as strings, not Date objects"
- Keep it to 1-3 lines
- Follow the existing CLAUDE.md style and conventions
**Templates:**
For bug patterns:
```
- **[Area/Component]**: [What to do/avoid] — [why, with specific details]
```
For API gotchas:
```
- `functionName()` in `path/to/file`: [What's surprising about it] — [consequence if ignored]
```
For architectural constraints:
```
- [Constraint description] — [why it exists and what breaks if violated]
```
**Examples of well-formatted learnings:**
- **Order processing**: Always multiply amounts by 100 before passing to `processOrder()` — it expects cents, not dollars
- `db.query()` in `src/database.ts`: Always pass the `timeout` option — the default is infinite and has caused production hangs (30 second timeout recommended)
- Never import from `internal/` directories in `src/api/` — the build system treats these as separate compilation units and circular dependencies will silently break HMR
---
## Step 4: Confirm with User
Present the proposed addition via AskUserQuestion:
Show:
1. The exact text to be added
2. Where it will be placed in CLAUDE.md (section name, after which line/entry)
3. Why this qualifies as project-specific
Options:
- **"Add this"** — Write the entry as proposed
- **"Edit before adding"** — User provides modified text, then write that instead
- **"Skip"** — Do not add anything, return to calling skill
---
## Step 5: Write Update
If the user confirmed (or provided edited text):
1. Use the Edit tool to add the entry to CLAUDE.md at the identified location
2. If a new section was needed, create the section heading first
3. Verify the edit was applied correctly by reading the modified area
4. Report success to the calling skill with a summary of what was added
If the user chose "Skip":
- Report to the calling skill that the learning was declined
- Do not modify any files