AI skills, built to order
We turn recurring work procedures into skills for AI agents. We test them on agreed tasks and provide pass counts and failure details under our own grading criteria.
What a skill is
A skill is a set of instructions an AI agent reads for a particular task. It stores recurring procedures, checks, and actions to avoid in a file.
What you receive
You receive the skill file, usage instructions, and test records. We also list untested items and a recommended date to run it again.
- Skill file
- Usage instructions
- Test records
Terms of service
- 01
Environment and costs
The standard deliverable is a Claude Code skill file (SKILL.md); support for other tools requires a compatibility check and agreement. Skills run in your account and environment, and you cover external service fees.
- 02
Materials and testing
We test with materials you are authorized to share, with personal and confidential information redacted, or with samples we create; we do not accept passwords or API keys. Results apply to the stated test conditions, and a person must review output before it is sent externally.
- 03
Delivery and setup
We provide the skill file, instructions, test records, and a method and recommended date for rerunning checks. One scheduled screen-sharing session covers installation and the first run; you enter any secrets yourself.
- 04
Corrections
We correct build defects reported within 30 days starting the day after delivery, reproducible within the agreed environment and scope, at no charge; we provide a schedule after review. Changes to work rules or external tools, and support after that period, require a separate quote; statutory rights remain unaffected.
- 05
Quote and agreement
After reviewing the work and materials, we provide the scope, completion criteria, price, and estimated timeframe. Payment, cancellation, refunds, deliverable usage rights, and data retention and deletion terms are agreed in writing before work begins.
Work you can hand us
Work with clear inputs and outputs, repeated by people to the same standard, moves well into a skill.
- 01
A first draft of a report you write every week to a fixed template
- 02
Finding blank fields and contradicting values before a document goes out
- 03
Checking title, tags and sources before a blog post is published
- 04
Bringing a code review routine that differs by team into one order
These are examples of the kind of work, not past deliveries.
What we do not take
- We require a human review step before output is sent externally.
- Work that requires handing us passwords or API keys
- Preparing public procurement or certification filings on someone's behalf
How a build goes
- 01
Understand the work
We ask how you handle the work, how often it occurs each month, and which tasks an error would affect.
- 02
Look for what exists
If a published skill already does the job, we point you to it instead of building one.
- 03
Scope and quote
We agree on the scope and test method, then provide the cost and schedule.
- 04
Build
The skill file and a how-to. Your rules and the things it must not do go inside it.
- 05
Test
Run the same tasks with and without the skill, then score it under our grading criteria.
- 06
Hand over
We provide the skill file, usage instructions, and test records, along with untested items and a recommended date to run it again.
What we have built
The citation check, instruction standard, and scoreboard sync were built for our internal work. The spec checker was developed during commissioned specification review work. Each case describes what we checked, with scoring records linked where available.
Scoring under our own criteria: not performed
Naver and Google AI citation check
Reviews Korean articles and search queries with the aim of gaining citations in Naver and Google AI search, and suggests changes. It labels evidence as official statements, measurements, or claims with unverified sources.
What we checkedWe compared responses to three tasks with and without the skill. General advice showed no difference; differences appeared on items requiring internal information such as our observations and tool paths. This test evaluated responses. It did not measure an increase in actual search citations.Scoring under our own criteria: not performed
Agent instruction standard
This standard guides the writing and revision of instructions for our internal agents. It organizes instructions into eight sections and requires four clauses and completion checks.
What we checkedWe documented procedures for writing and revising agent instructions. We have not compared results before and after applying the standard.Scoring under our own criteria: not performed
Skill scoreboard sync
Transfers confirmed skill comparison results into the task routing table. If it cannot parse the source, it writes nothing and preserves manually added exceptions.
What we checkedOn its first run, it found that two of the four skills selected in comparison tests were missing from the routing table. We checked that repeated application caused no further changes to the table.GradedSee the scoring record
Spec checker
An internal Python tool based on checks people performed by eye during actual specification reviews. It looks for missing fields and inconsistent values and lists its findings.
What we checkedWe re-scored it four times after its downgrade. Across 65 runs, the recorded build caught 22 of 25 planted defects, missed three, and wrongly flagged three clean documents. We also documented the downgrade and subsequent fixes.
What went wrong while building
Places where we got stuck or got it wrong while building and using skills, with what we changed.
2026-09-12
We found a failure when we reran a signed-off skill after 18 days
The diagram skill was signed off and last verified on 25 August. When we reran it on 12 September, a file in an external library dependency returned a 404 error. We could not establish when the failure began.
What changedWe switched to a bundled library build from another provider. One typeface still fell back to a system font.2026-09-09
The AI search citation skill we found was already installed
An AI search citation skill we found through social media and GitHub was already installed. Its 314-line seo-geo file contained zero occurrences of naver at the time.
What changedWe built a skill with additional Naver checks and Google guidance for Korean content. After comparing responses on three tasks, we reduced general advice and expanded the content based on our own measurements.2026-08-27
Our call counter overstated the unused count
The counter reported 24 entries with no calls in a month. It counted skill calls but missed agent calls, recording zero for an agent that had actually been called 22 times.
What changedCounting both invocation types reduced the zero-call count to 14. Separately from call counts, we record six statuses: selected, on hold, tested but not selected, untested, used without comparison, and excluded from operation.2026-08-20
One full-width space caused three checks to be skipped
In the third re-scoring after the downgrade, we found that one full-width space in a specification table header caused three checks to be skipped. The tool reported zero findings and exited with code 0, and its grade fell to 5.
What changedAfter the fix, the fourth re-scoring brought it to grade 4. It remained below the grade 1 required for the work. The scoring record documents the downgrade and fixes.
Is there work you keep repeating?
Tell us what the work is, how you handle it now, and how often it comes up each month. We reply first with whether it can become a skill, the scope, and a quote.
Plenty of work still needs a person to check the result after it becomes a skill. For that work, we build the check into the skill.