Teach your coding agent to build with VAKYN.
claude plugin marketplace add vakyn-ai/skillsclaude plugin install vakyn@vakyn-ai- Claude Code: then run
/reload-pluginsin an open session, or start a new one. - Codex, Cursor, OpenCode: run it in your project folder. The skills land in
.agents/skills/; add-gfor every project.
Four steps from first call to go-live.
/vakyn:research finds where VAKYN fits, a Build skill writes it, and /vakyn:check proves it on your data before you switch. No server yet? Start with /vakyn:start.
Step 01
Start
Set up and first call
/vakyn:startSet up
Checks whether VAKYN fits, picks a model file for your machine, sets up the server and makes a first call.
Set up VAKYN on this machine.What it covers
- IdeasWhat VAKYN could do in your app, before you pick a step.
- Fit checkWhether VAKYN fits your problem.
- Model choiceWhich model file suits your machine.
- SetupInstall and start the server, and get an API key.
- First callSend a first request and read the answer.
- API referenceWhat the API accepts, returns and refuses.
- SDK guideCall it from Python or TypeScript.
- FixesErrors such as 401, 422, 429 or 503, out of GPU memory, or a server that won't start.
- Use it when
- You haven't run VAKYN yet, the server won't start, or a call comes back with an error.
- What you get back
- A running server, an API key kept in VAKYN_API_KEY, and a first answer.
- Call it directly
/vakyn:startin Claude Code, or just ask.- Another skill fits better when
- Your code already calls a hosted decision API: /vakyn:migrate
Step 02
Find
Where VAKYN fits
/vakyn:researchResearch
Finds the decisions a chat model or a person makes in your app or process, and ranks them.
Where could VAKYN help in this repo?What it covers
- Codebase scanEvery chat-model call in your code, followed through shared wrappers.
- Process reviewA process described in plain words, when there is no code.
- Logs and tracesReal volume, cost and wait time from your call logs.
- SortingEach call marked as a decision, writing to keep, or no fit.
- RankingGain against risk and effort for each spot.
- ReportVakyn-research.md with the top 3, a ranked table and open questions.
- Next skillWhich skill to run for each spot.
- Use it when
- You want to know where VAKYN would help before you change anything.
- What you get back
- vakyn-research.md: each decision with its file and line, ranked by gain against risk and effort. No code changed, no calls made.
- Call it directly
/vakyn:researchin Claude Code, or just ask.- Another skill fits better when
- You already picked the prompt or tool list: /vakyn:convert
Step 03
Build
Write it
/vakyn:questionsQuestions
Picks noul, choice or score, writes the options and levels, and fixes a question that answers badly.
Why did it pick billing for this message?What it covers
- Question typeNoul, choice or score for each judgment.
- Options and levelsThe options, an escape option, and score levels.
- DescriptionsFor options that get mixed up or need long definitions.
- The stateWhat to send as the document.
- Exact work in codeRules, sums, dates and thresholds moved out of the question.
- One request or twoWhen some questions depend on others.
- Diagnose and rewriteA request whose answer looks wrong.
- ConfidenceWhat the numbers mean.
- Use it when
- You're writing a request, or an answer looks wrong, unsure or sure of the wrong thing.
- What you get back
- The cause, and a rewritten request that passes the pre-send checks.
- Call it directly
/vakyn:questionsin Claude Code, or just ask.- Another skill fits better when
- The wording is fine and you need a threshold: /vakyn:check
/vakyn:convertConvert
Turns a chat-model prompt, an output schema or a tool list into VAKYN calls, with Python or TypeScript code.
Make VAKYN pick the tool in agent.py.What it covers
- Prompt sortEvery line of a prompt that sorts, routes, flags or rates, put in its place.
- Output schemasA structured output schema turned into questions.
- Tool listsThe tool lists of the common chat-model APIs, so VAKYN picks the tool.
- MCP serversA server's tool list.
- Two-stage pickMore than 255 tools.
- Code shapeHow to run side by side, log both answers and switch.
- Use it when
- A prompt in your app sorts, routes, flags or rates, or a chat model picks which tool to call.
- What you get back
- Every line of the prompt sorted, the request as a .json file, and the code change. Your chat model keeps running side by side until you switch.
- Call it directly
/vakyn:convertin Claude Code, or just ask.- Another skill fits better when
- You're building a new app: /vakyn:workflow
- An agent works through a written procedure: /vakyn:distill
/vakyn:workflowWorkflow
Designs a new decision workflow from a task in plain words, such as routing, triage, picking a tool or checking a claim.
Build a router for our returns emails.What it covers
- RouteSend each item to one team, queue or person.
- TriageSort a batch by type and urgency.
- Tool pickChoose an assistant's next tool or step.
- Check a claimA statement against a document.
- RubricGrade on a scale, on one or several criteria.
- Best matchPick one item from a short list.
- Spot riskFlag what needs a person, such as refunds, legal or safety.
- Pick, don't writeGet a value, date or sentence out of text by picking it.
- ComposeSeveral of these in a row.
- Use it when
- You're building something new and know the task, not the questions.
- What you get back
- The questions, a state builder, working code from the closest pattern, and a path for unsure answers.
- Call it directly
/vakyn:workflowin Claude Code, or just ask.- Another skill fits better when
- You have a prompt or tool list already: /vakyn:convert
/vakyn:cascadeCascade
A cheap chat model drafts, VAKYN checks the draft against its sources, and only failed drafts go to the expensive model or a person.
Let a cheap model draft our support answers and send only the failed drafts to the big one.What it covers
- FitWhether your pipeline should cascade.
- Draft checkThe question, state and threshold that check a draft.
- RoutingThe code around your existing chat-model call.
- Side by sideTry it on real traffic with both answers logged.
- Cost savedA summary from the cascade log.
- EscalationsWhat to look at when many drafts fail.
- Use it when
- A pipeline writes answers from sources, such as help articles or contracts, and the expensive model does all of it.
- What you get back
- The draft check, the routing code around your current call, a side-by-side log and a cost-saved summary.
- Call it directly
/vakyn:cascadein Claude Code, or just ask.- Another skill fits better when
- You check one claim against one document: /vakyn:workflow
- You grade outputs offline: /vakyn:judge
/vakyn:judgeJudge
Grades chat-model or agent outputs against your rubric, lines the grades up with your experts, and runs in CI.
Grade our support bot's replies with this rubric and gate the CI on it.What it covers
- Rubric to questionsOne question per criterion.
- PairwiseCompare two outputs: model A against B, or an old prompt against a new one.
- AlignmentGrades lined up with your experts' grades.
- CI and monitoringA CI gate and a score on every trace.
- PitfallsWhether a high confidence means the grade is right.
Usual order: rubric to questions → grade 10–20 outputs → expert grades → alignment → CI gate → monitoring → alignment again when the rubric changes.
- Use it when
- You grade replies, answers or agent runs by hand, or with a chat model as the judge.
- What you get back
- One question per criterion, grades checked against your experts' grades, and a CI gate.
- Call it directly
/vakyn:judgein Claude Code, or just ask.- Another skill fits better when
- The grader sits inside a new app's flow: /vakyn:workflow
- You need a threshold for unsure answers: /vakyn:check
/vakyn:batchBatch
Runs a CSV, JSONL, Parquet or SQL table through VAKYN: a pilot first, then the full run with resume and progress.
Tag the 40,000 rows in reviews.csv by topic and sentiment.What it covers
- PlanThe data, questions, columns and server, agreed before anything runs.
- FilesCSV or JSONL.
- TablesDatabase tables, Parquet or a log dump.
- Personal dataWhich columns stay out, and emails and phone numbers removed from free text.
- Pilot30 rows spread over the data, with the projected run time.
- Full runResume, retries, a progress line and a summary.
- Script repairThe missing parts added to a slow or fragile loop.
- Use it when
- You have rows to sort, tag or score, and questions that work on a few of them.
- What you get back
- A script in your project, answer and confidence columns next to your data, and a review file for the unsure rows. Your source data is never overwritten.
- Call it directly
/vakyn:batchin Claude Code, or just ask.- Another skill fits better when
- You still need the questions: /vakyn:workflow
- You want a side-by-side test on 30 records: /vakyn:check
/vakyn:distillDistill
Turns an agent that follows a written procedure into a program. Code does the exact steps, VAKYN questions make the decisions, small agent steps do look-ups, and a person takes the hard calls.
Turn our supplier-review agent into a program, using its traces.What it covers
- Collect runsTraces with inputs, tool calls, outputs, cost and notes, kept on your machine.
- Step mapEach step marked as a decision, code, a look-up or a hard call.
- ProgramThe code in your stack, with a record of every decision.
- RolloutTraffic moved in stages, with the agent as fallback.
- ReportCost per case, share handled, escalations and agreement for each batch.
- Use it when
- An agent does the same job every time, such as supplier reviews or invoice disputes, and each run costs too much or can't be explained.
- What you get back
- A step map of the procedure, the program in your stack with a record of every decision, and a rollout in stages with the agent as fallback.
- Call it directly
/vakyn:distillin Claude Code, or just ask.- Another skill fits better when
- You have one prompt or tool list, no agent loop: /vakyn:convert
- It's a new task with no agent: /vakyn:workflow
/vakyn:migrateMigrate
Moves code that calls a hosted decision API to your own VAKYN server, and points framework adapters at it.
Move our decision calls to our own VAKYN server.What it covers
- Call sitesEvery call to a hosted decision API, with file and line.
- SettingsModel names, base URLs and key variables in use.
- Framework adaptersCommon AI frameworks pointed at your server.
- DifferencesWhat changes: limits, errors, retries, fields only one side has.
- Side by sideAbout 30 recorded requests compared.
- Switch and roll backA config change, with a rollback plan written first.
- Use it when
- Your code already calls a hosted decision API and you want it on your own servers.
- What you get back
- Every call site with its file and line, the config change, a side-by-side comparison and a way back.
- Call it directly
/vakyn:migratein Claude Code, or just ask.- Another skill fits better when
- You're moving chat-model prompts or tool calls: /vakyn:convert
- You're setting up for the first time: /vakyn:start
Step 04
Prove
Check it on your data
/vakyn:checkCheck
Tests VAKYN against your chat model on about 30 of your records, reads side-by-side logs, sets thresholds, checks calibration and drift, and reviews the code before go-live.
Here's a week of side-by-side logs. Can we switch?What it covers
- Smoke testVAKYN next to your chat model on about 30 records.
- Log summaryA side-by-side log from a converted app.
- ThresholdsWhere to set the cut-off for unsure answers, from records you have marked.
- CalibrationWhether 0.9 really means right 9 times in 10.
- Drift checkAfter a change to the model, server, questions or state; runs in CI.
- ReplayWhich past cases change after a new rule or threshold.
- Tuning and promotionWhether to switch to a new version.
- Audit recordWhich model, inputs and policy version decided a case.
- Go-live reviewA review of the code before you switch.
Usual order: smoke test → side by side in the app → log summary → thresholds and calibration → go-live review; then a drift check after every model or question change.
- Use it when
- You're about to switch, and again after every change to the model, the questions or the code.
- What you get back
- How often VAKYN agrees with your chat model, where to set the threshold for unsure answers, and a go-live review.
- Call it directly
/vakyn:checkin Claude Code, or just ask.- Another skill fits better when
- Answers are odd because of the wording: /vakyn:questions
Not sure which one? Ask in plain words. Your agent reads each skill's description and picks the one that fits, and the skills point to each other when another one fits better.
Read the skills on GitHubWhat the skills always do.
A skill is a set of instructions your coding agent loads when your request matches it. These hold for all eleven.
Checks every request
Right shapes and limits, described options, an escape option such as other or none, one question per proposition.
Keeps exact work in code
Arithmetic, dates, thresholds and rules stay in your code. VAKYN answers the judgment.
Gives unsure answers a path
A person, a chat model or a review queue.
Runs side by side first
A converted app keeps using your chat model and logs both answers until you switch.
Never invents answers
If your server can't be reached, the agent says so and still writes the request and the code.
Keeps your key out of the chat
It lives in the VAKYN_API_KEY environment variable, not in the conversation.
Before you start
Give the skills your key. The code they write calls VAKYN MAX.
export VAKYN_API_KEY=vk_… # from vakyn.com/console/keysexport VAKYN_BASE_URL=https://api.vakyn.com