Prompts

What to type once the connector is attached

Every prompt here was selected, not written. 2,305 candidates were drafted across 22 use cases, and each one ran a four-stage funnel: a deterministic prefilter, a scored rubric, a routing gate, and a head-to-head tournament. What is shown came out the other end.

Three use cases had their candidates drafted by three model tiers: 25 each from Haiku, Sonnet and Opus. The question was whether the drafting model matters. It does, at the cheapest stage. Every candidate the deterministic prefilter rejected came from the Haiku arm, most for describing a job without naming what to run it on. Past that point the stages are judged, and judged stages move between runs, so treat the rest as directional.

An example prompt makes a claim — type this and you reach that tool. A hand-written example asserts the claim. A selected one has been measured against it. New here? Start at Connect.

Workflows

Multi-step runs that chain the tools

Each of these drives several verifiers to a single verdict. The same prompts appear on the matching cards in How it works.

audit-and-bridge

Audit a collection of skills for overlap and gaps, then wire a hand-off bridge per edge.

Type thisRun onSonnet

Audit the skill collection in skills/ for overlap, then propose bridges that let each skill delegate to the right downstream skill.

The bridges get written for you; your session judges whether each rewired hand-off holds.

best-practices-pass

Audit any repository for engineering best-practice readiness and get a ranked, cheapest-first adoption plan.

Type thisRun onSonnet

Audit the monorepo in /repos/backend-services for engineering practices and prioritize adoption by implementation cost.

The ranked plan is written by the workflow; your session reads it and decides what to act on.

build-hermes-agent

Scaffold a Hermes agent project on Track A or Track B and validate it to clean.

Type thisRun onSonnet

Scaffold a Track A Hermes agent at /projects/new-agent. Validate the project structure to ensure all dependencies resolve cleanly.

You pick the track up front and diagnose any findings the fix loop could not clear.

build-skill

Author a new skill end-to-end from an approved design plan.

Type thisRun onOpus

Create the skill from .claude/plans/data-sync.md into skills/connectors/data-sync/. Implement all stages, error recovery, and retry logic outlined in the locked design document.

Your session writes the skill itself, and decides what to do when the build stalls.

extend-rule

Add or tighten a numbered rule in a validator skill, patched across all five touchpoints at once.

Type thisRun onSonnet

Rule-303 in api-consistency-validator is too loose on trailing slashes. Add stricter logic and propagate through linter, formatter, CI, git hooks, and the web interface.

Rule bounded edits across known files; the session shapes the rule and reads a pass or fail.

friction-postmortem

Local Claude Code

Diagnose why a skill underperformed in real use and apply one bounded, evidenced fix.

Type thisRun onOpus

Can you trace through session.jsonl to identify why skills/api-refactor didn't meet expectations, then propose one specific, bounded fix?

It edits your skill from transcript evidence, so the session must judge whether the fix is right.

harvest-repo

Evaluate an external repository for what is worth adopting, ending in a durable practices-analysis record.

Type thisRun onSonnet

Evaluate https://github.com/torvalds/linux and create a durable record of the practices most worth emulating.

The run gathers evidence; deciding what is worth adopting is judgment against what you already have.

prospect-to-design

Local Claude Code

Mine past sessions for recurring workflows worth turning into a skill, rank them, and draft a design plan.

Type thisRun onOpus

My past sessions are in ~/.claude/transcripts. Find recurring workflows, score by time-saving potential, design the top three as skills.

You judge which mined candidate deserves a skill and whether the draft plan holds up.

release-new-skill

Check that a newly built skill is wired into every cross-cutting registry, then emit a release checklist.

Type thisRun onSonnet

New skill: doc-coauthoring at ~/.claude/skills/doc-coauthoring/. Is it registered in all required registries? Give me the release checklist.

The audit finds the gaps; your session applies the checklist edits across each registry file.

select-skill-prompts

Choose the example prompt to publish for each of a repository's skills, by measuring which one reaches the skill it claims.

Type thisRun onSonnet

Pick the example prompt for every skill in skills/ by testing, not vibes. Generate options, check which reach the intended skill, keep the best. I want evidence behind each published line.

The funnel ranks deterministically; picking which candidate to publish is judgment your session makes.

ship-skill

Run the pre-publish GO/NO-GO gate over one finished skill before uploading it.

Type thisRun onSonnet

Run the GO/NO-GO check on skills/skill-mcp-server-builder before I upload it. Tell me pass or fail, nothing else.

The gate is deterministic, but reading a NO GO and deciding what to fix needs judgment.

triage-and-fix

Sweep a skill's findings cheapest-first: run a gate, classify Tier 1/2/3, fix, re-gate.

Type thisRun onSonnet

Run the gate against skills/skill-bridge-patcher, sort findings cheapest-first, fix what's fixable, then re-gate until it's clean.

The sweep fixes itself; your session picks the gate and reads back what the loop could not fix.

Single asks

One question, one verifier

Substitute your own path, repo, or skill for the target named in the prompt. Everything else works unchanged.

deslop_lint

Strip the statistical tells of AI-generated prose from a document without flattening its voice.

Type thisRun onSonnet

Can you de-AI my /blog/opinion-piece.md? Keep the personality but remove the hedging patterns and transition-word padding.

The linter only flags; your session does the rewriting that keeps the voice intact.

ecosystem_audit

Map overlap and gaps across a whole directory of skills.

Type thisRun onSonnet

Walk through /home/dev/.claude/skills/director and map where skills overlap versus where we lack coverage.

Scored overlap map against fixed thresholds; the session reads the graph and picks what to wire.

eval_testing_audit

Audit a project's eval and measurement discipline against ET001-ET014.

Type thisRun onSonnet

Is /project/evals/ compliant with ET001-ET014? Break down all gaps by standard.

The scan is automatic; you weigh which gaps in the eval setup are worth fixing.

quality_grade

Score one skill against the Q001-Q010 quality rubric.

Type thisRun onSonnet

Check skills/artifact-diagramming for quality against the rubric and report.

The script leaves six rubric questions for you to answer by reading the skill itself.

llm_security_audit

Review an agent or harness for prompt-injection defense, trust boundaries, and tool-risk classification.

Type thisRun onSonnet

Audit the harness at config/agent-harness.yaml for trust boundaries. Which tools could be abused if a user crafts a malicious prompt?

The scan lists the gaps; ranking real risk and picking the right defense is the session's call.

mcp_security_audit

Audit an MCP server project's security posture against the OWASP-anchored MP001-MP011 rules.

Type thisRun onSonnet

Start a full MP001-MP011 security audit of /server. What are the gaps?

The scan is automatic; judging which security gaps actually matter is the session's job.

preflight_check

Validate one skill against Claude.ai's upload rules before uploading it.

Type thisRun onHaiku

Validate skills/workflow-orchestrator against the Claude.ai skill upload rules. Tell me what needs fixing before I submit it.

Deterministic rule check; the session points at the skill and relays the pass or fail list.

repo_clarity_audit

Check whether a repository explains itself to a new contributor (RC001-RC010).

Type thisRun onSonnet

Evaluate whether /README.md and related documentation together guide a newcomer through understanding the project and making a contribution.

The scan finds the gaps; your session writes the missing README, glossary or map entry.

practices_audit

Check a repository against the BP001-BP015 software-development readiness rules.

Type thisRun onSonnet

Check /packages/lib against BP001 through BP015. Flag blockers and show which standards we're missing.

Fixed checklist over the repo; the session decides which gaps apply and what to fix first.

doc_drift

Find documentation prose still referencing files, services, or commands that no longer exist.

Type thisRun onSonnet

The service-registry repo moved from /services/catalog to /registry/services. Search the documentation for hardcoded paths pointing to the old location.

The scan is mechanical; telling real drift from a deliberate mention is the part that needs care.

The model badge

Which model to run, and why it is usually not the biggest one

The badge is a recommendation for the model in your own session, not a model we run on your behalf. Almost every verifier here is a deterministic Python script that invokes no model at all, so the tier is sized by the work around the call: composing it, reading the findings envelope back, and deciding what to do about what it found.

That is why the pre-publish gate says Haiku. It chains three deterministic validators and returns one verdict, so a bigger model has nothing extra to think about. The jobs that ask for Opus are the ones where a model has to design something or weigh a trade-off, not the ones that run the most checks.

Tiers are derived from the same records the toolkit uses internally, not written for this page, so re-tiering a skill updates the badge here. Where a job is genuinely borderline the recommendation rounds up.

What this does not prove

A judge is not the router

The routing gate is a model reasoning about which tool fits, not the router your client actually runs. Your client has its own system prompt and its own tool-selection behaviour, so treat the score as strong evidence, not a guarantee.

The measurement is versioned against the models that produced it. A new model release rebases the numbers, so the corpus, the per-candidate scores, and the model ids are committed together.