Project Terminus · Onboarding

Claude skills

Skills = recipes you trigger with /

Think of a skill as a saved checklist Claude Code follows for you. You type a slash command. It does the long work. You still need to understand the output.

Plain English: what happens when you type a skill?

You open Claude Code in the project-terminus repo.
You type something like /optmodel-discover.
Claude follows that track’s skill file — browse domain, avoid duplicates, write proposals.
You skim the result, pick an idea, then run convert to build the real folder.
You validate locally, then package with /snorkel-zip.

The 3-phase pattern (most tracks)

/<track>-discover

1 · Discover

Browse the domain. Deduplicate against tasks that already exist. Propose 3 candidate ideas with short slugs. You skim and pick the idea with the most interesting trap (why a naive attempt fails).

/datalake-discover

/<track>-research <slug>

2 · Research

Optional on some tracks. Deep-dive one idea into a design dossier: feasibility, trap, data, reference solution, Go or No-Go. If No-Go, go back to discover.

/optmodel-research nurse-shift-coverage

/<track>-convert <slug>

3 · Convert

Scaffold tasks/<track>/<slug>/ with Dockerfile, instruction, oracle, tests. Runs quality gates (oracle=1, nop=0). Often creates a task/<slug> git branch / worktree for you.

/vulns-convert buffer-overflow-auth

Utility skills (every track)

These are not tied to optmodel/datalake/etc. Run them before you upload.

Command	What it does	When
/local-checks <slug>	Full pre-upload self-check: schema, reward contract, Dockerfile, leaks, AI-slop.	Before upload
/human-review <slug>	Reviewer’s-eye verdict (Accept / Needs Revision) before Snorkel’s human sees it.	Before upload
/task-debug	Diagnose failing or mis-calibrated tasks (0/10, 10/10, reward errors).	Something failed
/snorkel-zip	Package the task on your branch into a submission ZIP for the web UI.	Ready to upload
/ai-reporter	Scan for AI-slop patterns reviewers dislike (em-dashes, banner comments, etc.).	Before upload
/task-format-review	Formatting and hygiene review of a branch or PR (comments, prose style).	Polish pass

Pick a track (match the thinking)

Your track lead tells you which track you’re on. In many tracks the agent does not write Python — Python is for our tests.

If the task is about…	Track	Agent writes in
Optimization / scheduling / routing	optmodel	MiniZinc
Messy multi-file data pipelines	datalake	R
Enterprise SQL over dirty data	entsql	SQL (DuckDB)
Causal inference / econometrics	causal	R / Julia
Formal proofs	proofgen	Dafny / Lean
Security bugs to find & patch	vulns	C/C++/Go/Rust/…
Finance / accounting papers	finpaper	R
Hardware description	verilog	Verilog
Language implementations	cipher	Rust/Go/C++/Java
Performance budgets	quicksilver	C/C++/Rust/Go
Concurrency correctness	riptide	Rust/Java/C/C++
Distributed systems	tempest	Go/Rust/Java
Cloud / infra / IaC	cloud-convert	varies

Worked example (copy the pattern)

You only type a few lines. The assistant does the hundreds of steps in between.

You: /optmodel-discover
     → 3 optimization ideas with slugs

You: /optmodel-convert nurse-shift-coverage-min-cost
     → builds folder + oracle + tests + gates

You: /local-checks nurse-shift-coverage-min-cost
You: /human-review nurse-shift-coverage-min-cost

You: /snorkel-zip
     → you upload the ZIP on Snorkel’s website

Next: follow the daily workflow and the flow diagram below.

Pipeline

How a task moves from idea → accepted

Read in order. Loops mean go back and fix. Step 08 is done.

See skills

Pick a track

Your lead assigns the kind of reasoning (optmodel, datalake, vulns…).

Discover ideas

Type /<track>-discover — get 3 ideas. Pick the one with the best trap.

Research

/<track>-research turns one idea into a Go / No-Go plan.

Convert

/<track>-convert builds the real task folder (Docker, oracle, tests).

Validate locally

oracle = 1 · nop = 0 · ruff clean · /local-checks · /human-review

Upload to Snorkel

/snorkel-zip → upload ZIP in the web UI → wait for AI eval.

Read & revise

Too easy? Harden. Unsolvable? Add fair disclosure. Re-upload same ID.

Accepted

HARD (or MEDIUM), Solvable, ≤5/10 passes, reviewer OK → stop.