| name | karpathy-guidelines-and-skill-creator |
| description | Behavioral guidelines to reduce common LLM coding mistakes AND a complete system for creating, testing, and optimizing skills. Use this skill WHENEVER the user asks you to write, review, debug, refactor, or discuss code — or whenever you are about to generate any code yourself. Also use when the user wants to create a skill from scratch, edit or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy. Apply to every coding task including fixing bugs, adding features, creating new files, refactoring, code reviews, architecture discussions, writing tests, and skill development. |
Karpathy Guidelines
Behavioral guidelines to reduce common LLM coding mistakes, derived from Andrej Karpathy's observations on LLM coding pitfalls.
Tradeoff: These guidelines bias toward caution over speed. For trivial tasks (simple typo fixes, obvious one-liners), use judgment — not every change needs the full rigor. The goal is reducing costly mistakes on non-trivial work.
The Problems
From Andrej Karpathy:
"The models make wrong assumptions on your behalf and just run along with them without checking. They don't manage their confusion, don't seek clarifications, don't surface inconsistencies, don't present tradeoffs, don't push back when they should."
"They really like to overcomplicate code and APIs, bloat abstractions, don't clean up dead code... implement a bloated construction over 1000 lines when 100 would do."
"They still sometimes change/remove comments and code they don't sufficiently understand as side effects, even if orthogonal to the task."
Four principles in one file that directly address these issues:
| Principle | Addresses |
|---|
| Think Before Coding | Wrong assumptions, hidden confusion, missing tradeoffs |
| Simplicity First | Overcomplication, bloated abstractions |
| Surgical Changes | Orthogonal edits, touching code you shouldn't |
| Goal-Driven Execution | Leverage through tests-first, verifiable success criteria |
Principle 1: Think Before Coding
Don't assume. Don't hide confusion. Surface tradeoffs.
Before implementing anything, follow these steps:
- State your assumptions explicitly. If you are uncertain about scope, format, fields, or approach — ask instead of guessing.
- If multiple interpretations exist, present all of them with tradeoffs. Never pick one silently.
- If a simpler approach exists than what the user described, say so. Push back when warranted.
- If something is unclear, stop immediately. Name exactly what is confusing and ask for clarification.
The reason this matters: LLMs naturally want to be helpful and will confidently run with wrong assumptions rather than appear uncertain. This creates wasted work and bugs. Surfacing confusion early is always cheaper than fixing wrong code later.
Example 1: Hidden Assumptions
User Request: "Add a feature to export user data"
❌ What LLMs Do (Wrong Assumptions)
def export_users(format='json'):
"""Export all users to a file."""
users = User.query.all()
if format == 'json':
with open('users.json', 'w') as f:
json.dump([u.to_dict() for u in users], f)
elif format == 'csv':
with open('users.csv', 'w') as f:
writer = csv.DictWriter(f, fieldnames=['id', 'email', 'name'])
writer.writeheader()
writer.writerows([u.to_dict() for u in users])
return f"Exported {len(users)} users"
Problems:
- Assumed it should export ALL users (what about pagination? privacy?)
- Assumed file location without asking
- Assumed which fields to include
- Assumed CSV fieldnames without checking actual data structure
✅ What Should Happen (Surface Assumptions)
Before implementing, I need to clarify:
1. **Scope**: Export all users or filtered subset? (privacy implications)
2. **Format**: You mentioned "export" - did you mean:
- Download file in browser?
- Background job with email notification?
- API endpoint returning data?
3. **Fields**: Which user fields? (some might be sensitive)
4. **Volume**: How many users typically? (affects approach)
Simplest approach: Add an API endpoint that returns paginated JSON.
Would need more info for file-based exports.
What's your preference?
Example 2: Multiple Interpretations
User Request: "Make the search faster"
❌ What LLMs Do (Pick Silently)
@lru_cache(maxsize=1000)
async def search(query: str) -> List[Result]:
✅ What Should Happen
"Make the search faster" could mean:
1. **Faster response time** (< 100ms instead of 500ms)
- Add database indexes
- Cache frequent queries
- Estimated effort: 2 hours
2. **Handle more concurrent searches** (throughput)
- Add async/await
- Connection pooling
- Estimated effort: 4 hours
3. **Faster perceived speed** (UX)
- Show partial results immediately
- Progressive loading
- Estimated effort: 3 hours
Current search takes ~500ms for typical queries.
Which aspect matters most for your use case?
Principle 2: Simplicity First
Write the minimum code that solves the problem. Nothing speculative.
Follow these rules strictly:
- Do not add features beyond what was explicitly asked for.
- Do not create abstractions (base classes, strategy patterns, factory methods) for code that is only used once.
- Do not add "flexibility" or "configurability" that was not requested.
- Do not add error handling for scenarios that cannot realistically occur.
- If you write 200 lines and it could be 50, rewrite it as 50.
Ask yourself: "Would a senior engineer say this is overcomplicated?" If the answer is yes, simplify.
The reason this matters: overengineered code is harder to understand, introduces more bugs, takes longer to implement, and is harder to test. Simple code can always be refactored later when complexity is actually needed. Premature abstraction is worse than no abstraction.
Example 1: Over-abstraction
User Request: "Add a function to calculate discount"
❌ What LLMs Do (Overengineered)
from abc import ABC, abstractmethod
from enum import Enum
from typing import Protocol, Union
from dataclasses import dataclass
class DiscountStrategy(ABC):
@abstractmethod
def calculate(self, amount: float) -> float:
pass
class PercentageDiscount(DiscountStrategy):
def __init__(self, percentage: float):
self.percentage = percentage
def calculate(self, amount: float) -> float:
return amount * (self.percentage / 100)
class FixedDiscount(DiscountStrategy):
def __init__(self, fixed_amount: float):
self.fixed_amount = fixed_amount
def calculate(self, amount: float) -> float:
return min(self.fixed_amount, amount)
@dataclass
class DiscountConfig:
strategy: DiscountStrategy
min_purchase: float = 0.0
max_discount: float = float('inf')
class DiscountCalculator:
def __init__(self, config: DiscountConfig):
self.config = config
def apply_discount(self, amount: float) -> float:
if amount < self.config.min_purchase:
return 0
discount = self.config.strategy.calculate(amount)
return min(discount, self.config.max_discount)
✅ What Should Happen (Simple)
def calculate_discount(amount: float, percent: float) -> float:
"""Calculate discount amount. percent should be 0-100."""
return amount * (percent / 100)
discount = calculate_discount(100.0, 10.0)
When to add complexity: Only when you actually need multiple discount types. If that requirement comes later, refactor then.
Example 2: Speculative Features
User Request: "Save user preferences to database"
❌ What LLMs Do (Unnecessary Features)
class PreferenceManager:
def __init__(self, db, cache=None, validator=None):
self.db = db
self.cache = cache or InMemoryCache()
self.validator = validator or DefaultValidator()
def save(self, user_id: int, prefs: dict,
merge: bool = True,
validate: bool = True,
notify: bool = False) -> bool:
"""Save preferences with optional merging, validation, and notifications."""
if validate:
errors = self.validator.validate(prefs)
if errors:
raise ValidationError(errors)
existing = self.get(user_id) if merge else {}
if merge:
existing.update(prefs)
prefs = existing
self.db.execute(
"UPDATE users SET preferences = ? WHERE id = ?",
(json.dumps(prefs), user_id)
)
if self.cache:
self.cache.set(f"prefs:{user_id}", prefs)
if notify:
self.notify_preference_change(user_id, prefs)
return True
def notify_preference_change(self, user_id: int, prefs: dict):
pass
✅ What Should Happen (Just What Was Asked)
def save_preferences(db, user_id: int, preferences: dict):
"""Save user preferences to database."""
db.execute(
"UPDATE users SET preferences = ? WHERE id = ?",
(json.dumps(preferences), user_id)
)
Add later if needed: Caching (when performance matters), validation (when bad data appears), merging (when requirement emerges).
Principle 3: Surgical Changes
Touch only what you must. Clean up only your own mess.
When editing existing code, follow these rules:
- Do not "improve" adjacent code, comments, or formatting that is unrelated to the task.
- Do not refactor things that are not broken.
- Match the existing code style (quote style, spacing, naming conventions), even if you would do it differently in a new project.
- If you notice unrelated dead code or issues, mention them in your response — but do not modify them.
When your changes create orphaned code:
- Remove imports, variables, and functions that YOUR changes made unused.
- Do not remove pre-existing dead code unless the user specifically asks you to.
The test for every change: every modified line should trace directly back to the user's request. If a line does not connect to the request, revert it.
The reason this matters: "drive-by refactoring" creates noisy diffs, introduces unexpected regressions, and makes code review harder. The user asked for one thing — deliver exactly that.
Example 1: Drive-by Refactoring
User Request: "Fix the bug where empty emails crash the validator"
❌ What LLMs Do (Too Much)
def validate_user(user_data):
- # Check email format
- if not user_data.get('email'):
+ """Validate user data."""
+ email = user_data.get('email', '').strip()
+
+ # Validate email
+ if not email:
raise ValueError("Email required")
-
- # Basic email validation
- if '@' not in user_data['email']:
+ if '@' not in email or '.' not in email.split('@')[1]:
raise ValueError("Invalid email")
- # Check username
- if not user_data.get('username'):
+ # Validate username
+ username = user_data.get('username', '').strip()
+ if not username:
raise ValueError("Username required")
+ if len(username) < 3:
+ raise ValueError("Username too short")
+ if not username.isalnum():
+ raise ValueError("Username must be alphanumeric")
return True
Problems:
- "Improved" email validation beyond the bug fix
- Added username validation nobody asked for
- Changed comments
- Added docstring
✅ What Should Happen (Surgical)
def validate_user(user_data):
# Check email format
- if not user_data.get('email'):
+ email = user_data.get('email', '')
+ if not email or not email.strip():
raise ValueError("Email required")
# Basic email validation
- if '@' not in user_data['email']:
+ if '@' not in email:
raise ValueError("Invalid email")
# Check username
if not user_data.get('username'):
raise ValueError("Username required")
return True
Only changed: The specific lines that fix empty email handling.
Example 2: Style Drift
User Request: "Add logging to the upload function"
❌ What LLMs Do (Reformat Everything)
- def upload_file(file_path, destination):
+ def upload_file(file_path: str, destination: str) -> bool:
+ """Upload file to destination with logging."""
+ logger.info(f"Uploading {file_path} to {destination}")
+
try:
- with open(file_path, 'rb') as f:
- data = f.read()
-
- response = requests.post(destination, files={'file': data})
-
- if response.status_code == 200:
- return True
- else:
- return False
+ with open(file_path, "rb") as f:
+ data = f.read()
+ response = requests.post(destination, files={"file": data})
+ success = response.status_code == 200
+
+ if success:
+ logger.info(f"Upload successful: {file_path}")
+ else:
+ logger.error(f"Upload failed: {file_path}, status={response.status_code}")
+
+ return success
except Exception as e:
- print(f"Error: {e}")
+ logger.exception(f"Upload error for {file_path}: {e}")
return False
Problems:
- Changed quote style ('' to "")
- Added type hints nobody asked for
- Added docstring
- Reformatted whitespace
- Changed boolean return logic
✅ What Should Happen (Match Existing Style)
+ import logging
+
+ logger = logging.getLogger(__name__)
+
def upload_file(file_path, destination):
+ logger.info(f'Starting upload: {file_path}')
try:
with open(file_path, 'rb') as f:
data = f.read()
response = requests.post(destination, files={'file': data})
if response.status_code == 200:
+ logger.info(f'Upload successful: {file_path}')
return True
else:
+ logger.error(f'Upload failed: {file_path}, status={response.status_code}')
return False
except Exception as e:
- print(f"Error: {e}")
+ logger.exception(f'Upload error: {file_path}')
return False
Matched: Single quotes, no type hints, existing boolean pattern, spacing style.
Principle 4: Goal-Driven Execution
Define success criteria. Loop until verified.
Transform vague tasks into verifiable goals before writing code:
- "Add validation" → "Write tests for invalid inputs, then make them pass"
- "Fix the bug" → "Write a test that reproduces the bug, then make it pass"
- "Refactor X" → "Ensure all existing tests pass before and after the refactor"
For multi-step tasks, state a brief plan with verification at each step:
1. [Step] → verify: [how you will check this worked]
2. [Step] → verify: [how you will check this worked]
3. [Step] → verify: [how you will check this worked]
Strong success criteria let you loop independently and make confident progress. Weak criteria ("make it work") require constant clarification and lead to vague implementations.
The reason this matters: LLMs are exceptionally good at looping until they meet specific goals. Giving declarative success criteria instead of imperative instructions produces dramatically better results.
Example 1: Vague vs. Verifiable
User Request: "Fix the authentication system"
❌ What LLMs Do (Vague Approach)
I'll fix the authentication system by:
1. Reviewing the code
2. Identifying issues
3. Making improvements
4. Testing the changes
[Proceeds to make changes without clear success criteria]
✅ What Should Happen (Verifiable Goals)
To "fix authentication," I need to define success criteria.
What specific issue are we solving?
If the issue is "users stay logged in after password change":
Plan:
1. Write test: Change password → verify old session invalidated
Verify: Test fails (reproduces bug)
2. Implement: Invalidate sessions on password change
Verify: Test passes
3. Check edge cases: Multiple active sessions, concurrent changes
Verify: Additional tests pass
4. Verify no regression: Existing auth tests still pass
Verify: Full test suite green
Current test coverage for auth: [shows coverage]
What's the specific auth issue you're seeing?
Example 2: Multi-Step with Verification
User Request: "Add rate limiting to the API"
❌ What LLMs Do (All at Once)
✅ What Should Happen (Incremental with Verification)
Plan for rate limiting:
1. Add basic in-memory rate limiting (single endpoint)
Verify:
- Test: 100 requests → first 10 succeed, rest get 429
- Manual: curl endpoint 11 times, see rate limit error
2. Extract to middleware (apply to all endpoints)
Verify:
- Test: Rate limits apply to /users and /posts
- Existing endpoint tests still pass
3. Add Redis backend (for multi-server)
Verify:
- Test: Rate limit persists across app restarts
- Test: Two app instances share rate limit counter
4. Add configuration (rates per endpoint)
Verify:
- Test: /search allows 10/min, /users allows 100/min
- Config file parsed correctly
Each step is independently verifiable and deployable.
Start with step 1?
Example 3: Test-First Verification
User Request: "The sorting breaks when there are duplicate scores"
❌ What LLMs Do (Fix Without Reproducing)
def sort_scores(scores):
return sorted(scores, key=lambda x: (-x['score'], x['name']))
✅ What Should Happen (Reproduce First)
def test_sort_with_duplicate_scores():
"""Test sorting when multiple items have same score."""
scores = [
{'name': 'Alice', 'score': 100},
{'name': 'Bob', 'score': 100},
{'name': 'Charlie', 'score': 90},
]
result = sort_scores(scores)
assert result[0]['score'] == 100
assert result[1]['score'] == 100
assert result[2]['score'] == 90
def sort_scores(scores):
"""Sort by score descending, then name ascending for ties."""
return sorted(scores, key=lambda x: (-x['score'], x['name']))
Quick Reference Checklist
Before writing or modifying any code, run through this checklist:
- Assumptions — Have I stated what I am assuming? Could the user mean something different?
- Simplicity — Is this the simplest solution? Am I adding anything that was not asked for?
- Scope — Am I only changing what was requested? Am I touching unrelated code?
- Verification — Do I have a clear way to verify this works? Can I write a test first?
Anti-Patterns Summary
| Principle | Anti-Pattern | Fix |
|---|
| Think Before Coding | Silently assumes file format, fields, scope | List assumptions explicitly, ask for clarification |
| Simplicity First | Strategy pattern for single discount calculation | One function until complexity is actually needed |
| Surgical Changes | Reformats quotes, adds type hints while fixing bug | Only change lines that fix the reported issue |
| Goal-Driven | "I'll review and improve the code" | "Write test for bug X → make it pass → verify no regressions" |
Key Insight
The "overcomplicated" examples aren't obviously wrong—they follow design patterns and best practices. The problem is timing: they add complexity before it's needed, which:
- Makes code harder to understand
- Introduces more bugs
- Takes longer to implement
- Harder to test
The "simple" versions are:
- Easier to understand
- Faster to implement
- Easier to test
- Can be refactored later when complexity is actually needed
How to Know These Guidelines Are Working
You are applying these guidelines correctly when you see:
- Fewer unnecessary changes in diffs — only the requested changes appear
- Fewer rewrites due to overcomplication — code is simple the first time
- Clarifying questions come before implementation — not after mistakes
- Clean, minimal PRs — no drive-by refactoring or unsolicited "improvements"
Core Karpathy Insight
"LLMs are exceptionally good at looping until they meet specific goals... Don't tell it what to do, give it success criteria and watch it go."
The "Goal-Driven Execution" principle captures this: transform imperative instructions into declarative goals with verification loops.
Customization
These guidelines are designed to be merged with project-specific instructions. Add them to your existing instruction file or create a new one.
For project-specific rules, add sections like:
## Project-Specific Guidelines
- Use TypeScript strict mode
- All API endpoints must have tests
- Follow the existing error handling patterns in `src/utils/errors.ts`
Good code is code that solves today's problem simply, not tomorrow's problem prematurely.
Part 2: Skill Creator
A skill for creating new skills and iteratively improving them.
At a high level, the process of creating a skill goes like this:
- Decide what you want the skill to do and roughly how it should do it
- Write a draft of the skill
- Create a few test prompts and run claude-with-access-to-the-skill on them
- Help the user evaluate the results both qualitatively and quantitatively
- While the runs happen in the background, draft some quantitative evals if there aren't any (if there are some, you can either use as is or modify if you feel something needs to change about them). Then explain them to the user (or if they already existed, explain the ones that already exist)
- Use the
eval-viewer/generate_review.py script to show the user the results for them to look at, and also let them look at the quantitative metrics
- Rewrite the skill based on feedback from the user's evaluation of the results (and also if there are any glaring flaws that become apparent from the quantitative benchmarks)
- Repeat until you're satisfied
- Expand the test set and try again at larger scale
Your job when using this skill is to figure out where the user is in this process and then jump in and help them progress through these stages. So for instance, maybe they're like "I want to make a skill for X". You can help narrow down what they mean, write a draft, write the test cases, figure out how they want to evaluate, run all the prompts, and repeat.
On the other hand, maybe they already have a draft of the skill. In this case you can go straight to the eval/iterate part of the loop.
Of course, you should always be flexible and if the user is like "I don't need to run a bunch of evaluations, just vibe with me", you can do that instead.
Then after the skill is done (but again, the order is flexible), you can also run the skill description improver, which we have a whole separate script for, to optimize the triggering of the skill.
Communicating with the User
The skill creator is liable to be used by people across a wide range of familiarity with coding jargon. There's a trend now where the power of AI is inspiring plumbers to open up their terminals, parents and grandparents to google "how to install npm". On the other hand, the bulk of users are probably fairly computer-literate.
So please pay attention to context cues to understand how to phrase your communication! In the default case, just to give you some idea:
- "evaluation" and "benchmark" are borderline, but OK
- for "JSON" and "assertion" you want to see serious cues from the user that they know what those things are before using them without explaining them
It's OK to briefly explain terms if you're in doubt, and feel free to clarify terms with a short definition if you're unsure if the user will get it.
Creating a Skill
Capture Intent
Start by understanding the user's intent. The current conversation might already contain a workflow the user wants to capture (e.g., they say "turn this into a skill"). If so, extract answers from the conversation history first — the tools used, the sequence of steps, corrections the user made, input/output formats observed. The user may need to fill the gaps, and should confirm before proceeding to the next step.
- What should this skill enable Claude to do?
- When should this skill trigger? (what user phrases/contexts)
- What's the expected output format?
- Should we set up test cases to verify the skill works? Skills with objectively verifiable outputs (file transforms, data extraction, code generation, fixed workflow steps) benefit from test cases. Skills with subjective outputs (writing style, art) often don't need them. Suggest the appropriate default based on the skill type, but let the user decide.
Interview and Research
Proactively ask questions about edge cases, input/output formats, example files, success criteria, and dependencies. Wait to write test prompts until you've got this part ironed out.
Check available MCPs - if useful for research (searching docs, finding similar skills, looking up best practices), research in parallel via subagents if available, otherwise inline. Come prepared with context to reduce burden on the user.
Write the SKILL.md
Based on the user interview, fill in these components:
- name: Skill identifier
- description: When to trigger, what it does. This is the primary triggering mechanism - include both what the skill does AND specific contexts for when to use it. All "when to use" info goes here, not in the body. Note: currently Claude has a tendency to "undertrigger" skills -- to not use them when they'd be useful. To combat this, make the skill descriptions a little bit "pushy". So for instance, instead of "How to build a simple fast dashboard to display internal data.", you might write "How to build a simple fast dashboard to display internal data. Make sure to use this skill whenever the user mentions dashboards, data visualization, internal metrics, or wants to display any kind of company data, even if they don't explicitly ask for a 'dashboard.'"
- compatibility: Required tools, dependencies (optional, rarely needed)
- the rest of the skill :)
Skill Writing Guide
Anatomy of a Skill
skill-name/
├── SKILL.md (required)
│ ├── YAML frontmatter (name, description required)
│ └── Markdown instructions
└── Bundled Resources (optional)
├── scripts/ - Executable code for deterministic/repetitive tasks
├── references/ - Docs loaded into context as needed
└── assets/ - Files used in output (templates, icons, fonts)
Progressive Disclosure
Skills use a three-level loading system:
- Metadata (name + description) - Always in context (~100 words)
- SKILL.md body - In context whenever skill triggers (<500 lines ideal)
- Bundled resources - As needed (unlimited, scripts can execute without loading)
These word counts are approximate and you can feel free to go longer if needed.
Key patterns:
- Keep SKILL.md under 500 lines; if you're approaching this limit, add an additional layer of hierarchy along with clear pointers about where the model using the skill should go next to follow up.
- Reference files clearly from SKILL.md with guidance on when to read them
- For large reference files (>300 lines), include a table of contents
Domain organization: When a skill supports multiple domains/frameworks, organize by variant:
cloud-deploy/
├── SKILL.md (workflow + selection)
└── references/
├── aws.md
├── gcp.md
└── azure.md
Claude reads only the relevant reference file.
Principle of Lack of Surprise
Skills must not contain malware, exploit code, or any content that could compromise system security. A skill's contents should not surprise the user in their intent if described. Don't go along with requests to create misleading skills or skills designed to facilitate unauthorized access, data exfiltration, or other malicious activities. Things like a "roleplay as an XYZ" are OK though.
Writing Patterns
Prefer using the imperative form in instructions.
Defining output formats - You can do it like this:
## Report structure
ALWAYS use this exact template:
# [Title]
## Executive summary
## Key findings
## Recommendations
Examples pattern - It's useful to include examples:
## Commit message format
**Example 1:**
Input: Added user authentication with JWT tokens
Output: feat(auth): implement JWT-based authentication
Writing Style
Try to explain to the model why things are important in lieu of heavy-handed musty MUSTs. Use theory of mind and try to make the skill general and not super-narrow to specific examples. Start by writing a draft and then look at it with fresh eyes and improve it.
Test Cases
After writing the skill draft, come up with 2-3 realistic test prompts — the kind of thing a real user would actually say. Share them with the user: "Here are a few test cases I'd like to try. Do these look right, or do you want to add more?" Then run them.
Save test cases to evals/evals.json. Don't write assertions yet — just the prompts. You'll draft assertions in the next step while the runs are in progress.
{
"skill_name": "example-skill",
"evals": [
{
"id": 1,
"prompt": "User's task prompt",
"expected_output": "Description of expected result",
"files": []
}
]
}
See references/schemas.md for the full schema (including the assertions field, which you'll add later).
Running and Evaluating Test Cases
This section is one continuous sequence — don't stop partway through. Do NOT use /skill-test or any other testing skill.
Put results in <skill-name>-workspace/ as a sibling to the skill directory. Within the workspace, organize results by iteration (iteration-1/, iteration-2/, etc.) and within that, each test case gets a directory (eval-0/, eval-1/, etc.). Don't create all of this upfront — just create directories as you go.
Step 1: Spawn All Runs (With-Skill AND Baseline) in the Same Turn
For each test case, spawn two subagents in the same turn — one with the skill, one without. This is important: don't spawn the with-skill runs first and then come back for baselines later. Launch everything at once so it all finishes around the same time.
With-skill run:
Execute this task:
- Skill path: <path-to-skill>
- Task: <eval prompt>
- Input files: <eval files if any, or "none">
- Save outputs to: <workspace>/iteration-<N>/eval-<ID>/with_skill/outputs/
- Outputs to save: <what the user cares about — e.g., "the .docx file", "the final CSV">
Baseline run (same prompt, but the baseline depends on context):
- Creating a new skill: no skill at all. Same prompt, no skill path, save to
without_skill/outputs/.
- Improving an existing skill: the old version. Before editing, snapshot the skill (
cp -r <skill-path> <workspace>/skill-snapshot/), then point the baseline subagent at the snapshot. Save to old_skill/outputs/.
Write an eval_metadata.json for each test case (assertions can be empty for now). Give each eval a descriptive name based on what it's testing — not just "eval-0". Use this name for the directory too. If this iteration uses new or modified eval prompts, create these files for each new eval directory — don't assume they carry over from previous iterations.
{
"eval_id": 0,
"eval_name": "descriptive-name-here",
"prompt": "The user's task prompt",
"assertions": []
}
Step 2: While Runs Are in Progress, Draft Assertions
Don't just wait for the runs to finish — you can use this time productively. Draft quantitative assertions for each test case and explain them to the user. If assertions already exist in evals/evals.json, review them and explain what they check.
Good assertions are objectively verifiable and have descriptive names — they should read clearly in the benchmark viewer so someone glancing at the results immediately understands what each one checks. Subjective skills (writing style, design quality) are better evaluated qualitatively — don't force assertions onto things that need human judgment.
Update the eval_metadata.json files and evals/evals.json with the assertions once drafted. Also explain to the user what they'll see in the viewer — both the qualitative outputs and the quantitative benchmark.
Step 3: As Runs Complete, Capture Timing Data
When each subagent task completes, you receive a notification containing total_tokens and duration_ms. Save this data immediately to timing.json in the run directory:
{
"total_tokens": 84852,
"duration_ms": 23332,
"total_duration_seconds": 23.3
}
This is the only opportunity to capture this data — it comes through the task notification and isn't persisted elsewhere. Process each notification as it arrives rather than trying to batch them.
Step 4: Grade, Aggregate, and Launch the Viewer
Once all runs are done:
-
Grade each run — spawn a grader subagent (or grade inline) that reads agents/grader.md and evaluates each assertion against the outputs. Save results to grading.json in each run directory. The grading.json expectations array must use the fields text, passed, and evidence (not name/met/details or other variants) — the viewer depends on these exact field names. For assertions that can be checked programmatically, write and run a script rather than eyeballing it — scripts are faster, more reliable, and can be reused across iterations.
-
Aggregate into benchmark — run the aggregation script from the skill-creator directory:
python -m scripts.aggregate_benchmark <workspace>/iteration-N --skill-name <name>
This produces benchmark.json and benchmark.md with pass_rate, time, and tokens for each configuration, with mean ± stddev and the delta. If generating benchmark.json manually, see references/schemas.md for the exact schema the viewer expects. Put each with_skill version before its baseline counterpart.
-
Do an analyst pass — read the benchmark data and surface patterns the aggregate stats might hide. See agents/analyzer.md (the "Analyzing Benchmark Results" section) for what to look for — things like assertions that always pass regardless of skill (non-discriminating), high-variance evals (possibly flaky), and time/token tradeoffs.
-
Launch the viewer with both qualitative outputs and quantitative data:
nohup python <skill-creator-path>/eval-viewer/generate_review.py \
<workspace>/iteration-N \
--skill-name "my-skill" \
--benchmark <workspace>/iteration-N/benchmark.json \
> /dev/null 2>&1 &
VIEWER_PID=$!
For iteration 2+, also pass --previous-workspace <workspace>/iteration-<N-1>.
Cowork / headless environments: If webbrowser.open() is not available or the environment has no display, use --static <output_path> to write a standalone HTML file instead of starting a server. Feedback will be downloaded as a feedback.json file when the user clicks "Submit All Reviews". After download, copy feedback.json into the workspace directory for the next iteration to pick up.
Note: please use generate_review.py to create the viewer; there's no need to write custom HTML.
- Tell the user something like: "I've opened the results in your browser. There are two tabs — 'Outputs' lets you click through each test case and leave feedback, 'Benchmark' shows the quantitative comparison. When you're done, come back here and let me know."
What the User Sees in the Viewer
The "Outputs" tab shows one test case at a time:
- Prompt: the task that was given
- Output: the files the skill produced, rendered inline where possible
- Previous Output (iteration 2+): collapsed section showing last iteration's output
- Formal Grades (if grading was run): collapsed section showing assertion pass/fail
- Feedback: a textbox that auto-saves as they type
- Previous Feedback (iteration 2+): their comments from last time, shown below the textbox
The "Benchmark" tab shows the stats summary: pass rates, timing, and token usage for each configuration, with per-eval breakdowns and analyst observations.
Navigation is via prev/next buttons or arrow keys. When done, they click "Submit All Reviews" which saves all feedback to feedback.json.
Step 5: Read the Feedback
When the user tells you they're done, read feedback.json:
{
"reviews": [
{"run_id": "eval-0-with_skill", "feedback": "the chart is missing axis labels", "timestamp": "..."},
{"run_id": "eval-1-with_skill", "feedback": "", "timestamp": "..."},
{"run_id": "eval-2-with_skill", "feedback": "perfect, love this", "timestamp": "..."}
],
"status": "complete"
}
Empty feedback means the user thought it was fine. Focus your improvements on the test cases where the user had specific complaints.
Kill the viewer server when you're done with it:
kill $VIEWER_PID 2>/dev/null
Improving the Skill
This is the heart of the loop. You've run the test cases, the user has reviewed the results, and now you need to make the skill better based on their feedback.
How to Think About Improvements
-
Generalize from the feedback. The big picture thing that's happening here is that we're trying to create skills that can be used a million times (maybe literally, maybe even more) across many different prompts. Here you and the user are iterating on only a few examples over and over again because it helps move faster. The user knows these examples in and out and it's quick for them to assess new outputs. But if the skill you and the user are codeveloping works only for those examples, it's useless. Rather than put in fiddly overfitty changes, or oppressively constrictive MUSTs, if there's some stubborn issue, you might try branching out and using different metaphors, or recommending different patterns of working. It's relatively cheap to try and maybe you'll land on something great.
-
Keep the prompt lean. Remove things that aren't pulling their weight. Make sure to read the transcripts, not just the final outputs — if it looks like the skill is making the model waste a bunch of time doing things that are unproductive, you can try getting rid of the parts of the skill that are making it do that and seeing what happens.
-
Explain the why. Try hard to explain the why behind everything you're asking the model to do. Today's LLMs are smart. They have good theory of mind and when given a good harness can go beyond rote instructions and really make things happen. Even if the feedback from the user is terse or frustrated, try to actually understand the task and why the user is writing what they wrote, and what they actually wrote, and then transmit this understanding into the instructions. If you find yourself writing ALWAYS or NEVER in all caps, or using super rigid structures, that's a yellow flag — if possible, reframe and explain the reasoning so that the model understands why the thing you're asking for is important. That's a more humane, powerful, and effective approach.
-
Look for repeated work across test cases. Read the transcripts from the test runs and notice if the subagents all independently wrote similar helper scripts or took the same multi-step approach to something. If all 3 test cases resulted in the subagent writing a create_docx.py or a build_chart.py, that's a strong signal the skill should bundle that script. Write it once, put it in scripts/, and tell the skill to use it. This saves every future invocation from reinventing the wheel.
This task is pretty important — we are trying to create billions a year in economic value here! — and your thinking time is not the blocker; take your time and really mull things over. Write a draft revision and then look at it anew and make improvements. Really do your best to get into the head of the user and understand what they want and need.
The Iteration Loop
After improving the skill:
- Apply your improvements to the skill
- Rerun all test cases into a new
iteration-<N+1>/ directory, including baseline runs. If you're creating a new skill, the baseline is always without_skill (no skill) — that stays the same across iterations. If you're improving an existing skill, use your judgment on what makes sense as the baseline: the original version the user came in with, or the previous iteration.
- Launch the reviewer with
--previous-workspace pointing at the previous iteration
- Wait for the user to review and tell you they're done
- Read the new feedback, improve again, repeat
Keep going until:
- The user says they're happy
- The feedback is all empty (everything looks good)
- You're not making meaningful progress
Advanced: Blind Comparison
For situations where you want a more rigorous comparison between two versions of a skill (e.g., the user asks "is the new version actually better?"), there's a blind comparison system. Read agents/comparator.md and agents/analyzer.md for the details. The basic idea is: give two outputs to an independent agent without telling it which is which, and let it judge quality. Then analyze why the winner won.
This is optional, requires subagents, and most users won't need it. The human review loop is usually sufficient.
Description Optimization
The description field in SKILL.md frontmatter is the primary mechanism that determines whether Claude invokes a skill. After creating or improving a skill, offer to optimize the description for better triggering accuracy.
Step 1: Generate Trigger Eval Queries
Create 20 eval queries — a mix of should-trigger and should-not-trigger. Save as JSON:
[
{"query": "the user prompt", "should_trigger": true},
{"query": "another prompt", "should_trigger": false}
]
The queries must be realistic and something a user would actually type. Not abstract requests, but requests that are concrete and specific and have a good amount of detail. For instance, file paths, personal context about the user's job or situation, column names and values, company names, URLs. A little bit of backstory. Some might be in lowercase or contain abbreviations or typos or casual speech. Use a mix of different lengths, and focus on edge cases rather than making them clear-cut (the user will get a chance to sign off on them).
Bad: "Format this data", "Extract text from PDF", "Create a chart"
Good: "ok so my boss just sent me this xlsx file (its in my downloads, called something like 'Q4 sales final FINAL v2.xlsx') and she wants me to add a column that shows the profit margin as a percentage. The revenue is in column C and costs are in column D i think"
For the should-trigger queries (8-10), think about coverage. You want different phrasings of the same intent — some formal, some casual. Include cases where the user doesn't explicitly name the skill or file type but clearly needs it. Throw in some uncommon use cases and cases where this skill competes with another but should win.
For the should-not-trigger queries (8-10), the most valuable ones are the near-misses — queries that share keywords or concepts with the skill but actually need something different. Think adjacent domains, ambiguous phrasing where a naive keyword match would trigger but shouldn't, and cases where the query touches on something the skill does but in a context where another tool is more appropriate.
The key thing to avoid: don't make should-not-trigger queries obviously irrelevant. "Write a fibonacci function" as a negative test for a PDF skill is too easy — it doesn't test anything. The negative cases should be genuinely tricky.
Step 2: Review with User
Present the eval set to the user for review using the HTML template:
- Read the template from
assets/eval_review.html
- Replace the placeholders:
__EVAL_DATA_PLACEHOLDER__ → the JSON array of eval items (no quotes around it — it's a JS variable assignment)__SKILL_NAME_PLACEHOLDER__ → the skill's name__SKILL_DESCRIPTION_PLACEHOLDER__ → the skill's current description
- Write to a temp file (e.g.,
/tmp/eval_review_<skill-name>.html) and open it
- The user can edit queries, toggle should-trigger, add/remove entries, then click "Export Eval Set"
- The file downloads to
~/Downloads/eval_set.json — check the Downloads folder for the most recent version in case there are multiple
This step matters — bad eval queries lead to bad descriptions.
Step 3: Run the Optimization Loop
Tell the user: "This will take some time — I'll run the optimization loop in the background and check on it periodically."
Save the eval set to the workspace, then run in the background:
python -m scripts.run_loop \
--eval-set <path-to-trigger-eval.json> \
--skill-path <path-to-skill> \
--model <model-id-powering-this-session> \
--max-iterations 5 \
--verbose
Use the model ID from your system prompt (the one powering the current session) so the triggering test matches what the user actually experiences.
While it runs, periodically tail the output to give the user updates on which iteration it's on and what the scores look like.
This handles the full optimization loop automatically. It splits the eval set into 60% train and 40% held-out test, evaluates the current description (running each query 3 times to get a reliable trigger rate), then calls Claude to propose improvements based on what failed. It re-evaluates each new description on both train and test, iterating up to 5 times. When it's done, it opens an HTML report in the browser showing the results per iteration and returns JSON with best_description — selected by test score rather than train score to avoid overfitting.
How Skill Triggering Works
Understanding the triggering mechanism helps design better eval queries. Skills appear in Claude's available_skills list with their name + description, and Claude decides whether to consult a skill based on that description. The important thing to know is that Claude only consults skills for tasks it can't easily handle on its own — simple, one-step queries like "read this PDF" may not trigger a skill even if the description matches perfectly, because Claude can handle them directly with basic tools. Complex, multi-step, or specialized queries reliably trigger skills when the description matches.
This means your eval queries should be substantive enough that Claude would actually benefit from consulting a skill. Simple queries like "read file X" are poor test cases — they won't trigger skills regardless of description quality.
Step 4: Apply the Result
Take best_description from the JSON output and update the skill's SKILL.md frontmatter. Show the user before/after and report the scores.
Package and Present (only if present_files tool is available)
Check whether you have access to the present_files tool. If you don't, skip this step. If you do, package the skill and present the .skill file to the user:
python -m scripts.package_skill <path/to/skill-folder>
After packaging, direct the user to the resulting .skill file path so they can install it.
Claude.ai-Specific Instructions
In Claude.ai, the core workflow is the same (draft → test → review → improve → repeat), but because Claude.ai doesn't have subagents, some mechanics change. Here's what to adapt:
Running test cases: No subagents means no parallel execution. For each test case, read the skill's SKILL.md, then follow its instructions to accomplish the test prompt yourself. Do them one at a time. This is less rigorous than independent subagents (you wrote the skill and you're also running it, so you have full context), but it's a useful sanity check — and the human review step compensates. Skip the baseline runs — just use the skill to complete the task as requested.
Reviewing results: If you can't open a browser (e.g., the VM has no display, or you're on a remote server), skip the browser reviewer entirely. Instead, present results directly in the conversation. For each test case, show the prompt and the output. If the output is a file the user needs to see (like a .docx or .xlsx), save it to the filesystem and tell them where it is so they can download and inspect it. Ask for feedback inline: "How does this look? Anything you'd change?"
Benchmarking: Skip the quantitative benchmarking — it relies on baseline comparisons which aren't meaningful without subagents. Focus on qualitative feedback from the user.
The iteration loop: Same as before — improve the skill, rerun the test cases, ask for feedback — just without the browser reviewer in the middle. You can still organize results into iteration directories on the filesystem if you have one.
Description optimization: This section requires the claude CLI tool (specifically claude -p) which is only available in Claude Code. Skip it if you're on Claude.ai.
Blind comparison: Requires subagents. Skip it.
Packaging: The package_skill.py script works anywhere with Python and a filesystem. You can run it and the user can download the resulting .skill file.
Updating an existing skill: The user might be asking you to update an existing skill, not create a new one. In this case:
- Preserve the original name. Note the skill's directory name and
name frontmatter field -- use them unchanged. E.g., if the installed skill is research-helper, output research-helper.skill (not research-helper-v2).
- Copy to a writeable location before editing. The installed skill path may be read-only. Copy to
/tmp/skill-name/, edit there, and package from the copy.
- If packaging manually, stage in
/tmp/ first, then copy to the output directory -- direct writes may fail due to permissions.
Cowork-Specific Instructions
If you're in Cowork, the main things to know are:
- You have subagents, so the main workflow (spawn test cases in parallel, run baselines, grade, etc.) all works. (However, if you run into severe problems with timeouts, it's OK to run the test prompts in series rather than parallel.)
- You don't have a browser or display, so when generating the eval viewer, use
--static <output_path> to write a standalone HTML file instead of starting a server. Then proffer a link that the user can click to open the HTML in their browser.
- After running tests, you should always generate the eval viewer for the human to look at examples before revising the skill yourself and trying to make corrections, using
generate_review.py (not writing your own boutique html code). GENERATE THE EVAL VIEWER BEFORE evaluating inputs yourself. You want to get them in front of the human ASAP!
- Feedback works differently: since there's no running server, the viewer's "Submit All Reviews" button will download
feedback.json as a file. You can then read it from there (you may have to request access first).
- Packaging works —
package_skill.py just needs Python and a filesystem.
- Description optimization (
run_loop.py / run_eval.py) should work in Cowork just fine since it uses claude -p via subprocess, not a browser, but please save it until you've fully finished making the skill and the user agrees it's in good shape.
- Updating an existing skill: The user might be asking you to update an existing skill, not create a new one. Follow the update guidance in the Claude.ai section above.
Reference Files
The agents/ directory contains instructions for specialized subagents. Read them when you need to spawn the relevant subagent.
agents/grader.md — How to evaluate assertions against outputsagents/comparator.md — How to do blind A/B comparison between two outputsagents/analyzer.md — How to analyze why one version beat another
The references/ directory has additional documentation:
references/schemas.md — JSON structures for evals.json, grading.json, etc.
Skill Creator Core Loop (Summary)
Repeating the core loop for emphasis:
- Figure out what the skill is about
- Draft or edit the skill
- Run claude-with-access-to-the-skill on test prompts
- With the user, evaluate the outputs:
- Create benchmark.json and run
eval-viewer/generate_review.py to help the user review them
- Run quantitative evals
- Repeat until you and the user are satisfied
- Package the final skill and return it to the user
Please add steps to your TodoList, if you have such a thing, to make sure you don't forget. If you're in Cowork, please specifically put "Create evals JSON and run eval-viewer/generate_review.py so human can review test cases" in your TodoList to make sure it happens.
Good code is code that solves today's problem simply, not tomorrow's problem prematurely.
