Workflow skills
These entries describe the 32 workflow and style skills in this library. Each example is a chat prompt after installation. Codex uses $skill-name; Claude Code uses /skill-name. For the native plugins, use the namespaced equivalent: select zstack:how through the Codex skill picker or invoke /zstack:how in Claude Code. Replace the task details with your own files, feature, or repository.
Codex permits automatic selection for the read-only how and why skills and for python-best-practices when working with Python; every other skill, including setup-zstack, requires explicit invocation. Claude Code retains explicit-only flags for how, why, and python-best-practices; setup-zstack permits implicit invocation there. Reading a companion inside an explicitly selected workflow does not change that companion’s invocation policy.
Delegated workflows use native host agents by default, or Herdr execution when explicitly enabled. Native agents inherit model configuration; Herdr-launched CLIs use their own settings. Both retain bounded concurrency and isolated write ownership. Independent runs are not proof of provider diversity. Reports disclose unavailable agents and incomplete coverage. A workflow request does not independently authorize commits, publication, messages, or tracker writes.
Herdr workflow
You want existing workflows to run their agents and useful long-running supporting processes visibly in Herdr.
Start with the Herdr walkthrough for setup, execution behavior, pause/resume, and troubleshooting, or visit the upstream Herdr documentation.
| Host | Example invocation |
|---|---|
| Codex | $z-mode Use Herdr execution for this session, then fix the reconnect bug. |
| Claude Code | /zstack:z-mode Use Herdr execution for this session, then fix the reconnect bug. |
| Hermes | Load zstack:z-mode with skill_view and enable Herdr execution for this session. |
For execution selection without activating z-mode, invoke herdr-workflow explicitly and ask to enable it. A one-off request such as $herdr-workflow Inspect the agents assigned to this task does not persist a session preference.
How it works
Z-mode selects the normal task playbook, then loads this companion when Herdr is enabled. The playbook determines work, ownership, and acceptance; Herdr supplies agent launch, prompt submission, lifecycle waits, terminal inspection, and useful supporting panes. A small task still runs directly. The coordinator maintains task-to-agent/pane assignments and accepts artifacts only after the workflow's checks.
The session hook supplies Herdr and Native controls alongside Enable and Disable. Herdr changes execution only; Native disables Herdr while preserving z-mode. Enable selects z-mode and retains an existing execution preference; Disable clears both. These are preference writes, not process controls. Explicit pause drains or stops assigned workers; client detach leaves processes running.
Expected result
The original workflow's deliverable, visible task-owned agents/processes, verified results, and a reconciled roster or handoff. Reports distinguish agent readiness from task acceptance and name incomplete evidence.
Dependencies and limits
Herdr control requires HERDR_ENV=1, a compatible CLI/server, and the selected agent CLI. The installed CLI's help determines syntax. Missing capabilities require a disclosed, user-selected fallback; host restrictions and combined worker limits still apply. The upstream agent guide, concepts, working guide, and automation docs describe Herdr itself.
New CLI sessions do not inherit the coordinator's context, model settings, permissions, or activation. Panes alone do not isolate filesystem writes. Use existing task authority for exclusive files or requested worktrees, and supply consolidated briefs. Server restart is different from detach: conversation restoration depends on native integrations and does not prove work continued.
The preference is a marker beside mode state in each native host's plugin data directory and restored only for the same session ID. Stopping or switching away from active z-mode and clear/reset remove both; a Herdr-only selection survives a style switch; new children and rotated IDs start without either. Without the new hooks, record selection and opt-out conversationally. No installation, permission changes, server restarts, publication, or extra delegation follow merely from enabling it.
Full herdr-workflow instructions
Benchmark checklist
Use this before reporting or acting on a measured speedup or regression.
| Host | Example invocation |
|---|---|
| Codex | $benchmark-checklist Vet the before/after export benchmark and its measurement script. |
| Claude Code | /benchmark-checklist Vet the before/after export benchmark and its measurement script. |
How it works
Checks the limiter, production tuning, physical limits, failures, repeatability, end-to-end relevance, and whether the timed work happened. Comparative runs use at least five samples per side with counterbalanced or balanced randomized ordering, equivalent starting state, and independent warmup where needed. Record the order and state preparation, and report the median and range. A requested single-run ballpark still checks errors and completed work and is labeled as one run.
Expected result
A faster, slower, no measurable difference, or inconclusive verdict with units, run count, spread, limiter evidence, and artifact paths.
Dependencies and limits
Uses Explain the number. Z-mode's performance and hillclimb playbooks read this companion without changing its explicit-only policy. Missing evidence stays a gap; the skill does not expand the requested runtime budget or authorize publication.
Full benchmark-checklist instructions
Architect
You need types, ownership, or module boundaries before a non-trivial implementation.
| Host | Example invocation |
|---|---|
| Codex | $architect Design the queue admission API. Stop after the design for review. |
| Claude Code | /architect Design the queue admission API. Stop after the design for review. |
How it works
The agent traces callers and runtime flow, sketches caller usage, derives types and invariants, and compares distinct designs when the choice warrants an arena. It presents tradeoffs before implementation. Authorized implementation tests the sketch against real caller behavior and revisits designs that require repeated workarounds.
Expected result
A grounded design with signatures, ownership, invariants, and tradeoffs. Code follows only when requested.
Dependencies and limits
Uses how for flow, why for historical constraints, and arena for contested designs. A design request does not authorize scaffolding or implementation.
Arena
Several plausible approaches could produce materially different designs or artifacts.
| Host | Example invocation |
|---|---|
| Codex | $arena Compare three designs for queue coalescing and synthesize the strongest result. |
| Claude Code | /arena Compare three designs for queue coalescing and synthesize the strongest result. |
How it works
The agent defines the artifact and 3–6 criteria, then assigns the same brief to 2–3 independent native candidates unless you request another useful count. Each candidate has isolated output. After every candidate finishes, a fresh read-only judge from a model family other than the coordinator's scores the artifacts. Under native execution, an installed Codex or Claude CLI from a family other than the coordinator's can supply that judge. When no other-family judge can run, the coordinator skips judging and picks the base alone. The coordinator chooses a base, adapts useful ideas, and verifies the synthesis.
Expected result
A verified artifact plus the chosen base, adopted ideas, rejected ideas, the judge's identity and verdict or the skipped judging, and coverage gaps.
Dependencies and limits
Requires native delegation for independent candidates. Concurrency stays within host limits. Missing candidates remain gaps; a rejected model drops its seat and the run tops up to two candidates. An inherited or same-family model never judges. Same-model runs do not establish cross-provider comparison.
Automate me
You want a personal mode that records stated conventions or patterns from explicitly scoped history.
| Host | Example invocation |
|---|---|
| Codex | $automate-me Draft a project mode named review-first from my stated preferences in this chat. |
| Claude Code | /automate-me Draft a project mode named review-first from my stated preferences in this chat. |
How it works
The agent locates an existing named mode, preserves confirmed rules, and mines history only when requested. It writes concise SKILL.md metadata and instructions, references companions, and validates a local draft. For both hosts, one canonical directory has a collision-safe link into the other discovery directory.
Expected result
A reviewable personal-mode skill, with its evidence basis and registration status. Subjective rules remain open to your review.
Dependencies and limits
Defaults to explicit-only invocation. Project paths are .agents/skills for Codex and .claude/skills for Claude Code. Personal paths require that scope in your request. History access can be unavailable. Authoring does not authorize publication.
Blast radius
You need to know what a change can break beyond its immediate callers or diff.
| Host | Example invocation |
|---|---|
| Codex | $blast-radius Review the blast radius of the cache eviction diff and prove its key safety assumption. |
| Claude Code | /blast-radius Review the blast radius of the cache eviction diff and prove its key safety assumption. |
How it works
The agent reads the diff and history, identifies the fact that makes the change safe, and follows indirect contracts through pinned dependencies, wire formats, database state, and timing. It separates confirmed risks from cleared cases and runs the cheapest real-code proof available. Wide reviews can use independent read-only reviewers.
Expected result
A cited account of the change, its safety fact and proof level, confirmed risks, cleared cases, and the cheapest pre-merge check. Unproven facts are labeled.
Dependencies and limits
Uses how, why, and unslop. A read-only request creates no files. Missing execution permission remains a proof gap. Source inspection is weaker evidence than a runnable check or live reproduction.
Full blast-radius instructions
Bro
The previous assistant reply is too dense or full of jargon.
| Host | Example invocation |
|---|---|
| Codex | $bro |
| Claude Code | /bro |
How it works
The agent restates its last message in concise, ordinary language.
Expected result
A clearer version of the previous reply.
Dependencies and limits
This is a rewrite of the last reply, not a new investigation or implementation. To restate your goals and problem instead, use in-your-own-words.
Correct
You keep correcting agents for the same repository mistakes and want enforceable prevention.
| Host | Example invocation |
|---|---|
| Codex | $correct Find repeated mistakes in this repo's recent history and implement prevention. Leave changes uncommitted. |
| Claude Code | /correct Find repeated mistakes in this repo's recent history and implement prevention. Leave changes uncommitted. |
How it works
Groups cited, independent incidents into classes with at least two occurrences, then addresses the most frequent within scope. Prefers architecture, types, lint or CI, and behavioral tests, with docs reserved for judgment calls. Reproduces past mistakes in isolation to demonstrate failing checks and corrected cases, and pairs remaining rules with enforcement in the repository's agent instruction file.
Expected result
Reviewable prevention changes and a report of each class, evidence, selected level, reasons higher levels were unsuitable, actual checks, and proof gaps. A review-only request produces findings and proposals.
Dependencies and limits
Uses scoped history access and the repository's existing checks. Missing history or unexecuted CI is disclosed. Explicit-only on both hosts; future corrections do not automatically activate it. Commits, publication, personal memory, messages, and tracker writes require their own authorization.
Create a formal verification skill
A race, retry, parser, or state-machine invariant needs verification beyond the cases reached by ordinary tests.
| Host | Example invocation |
|---|---|
| Codex | $create-formal-verification-skill Verify that retries cannot apply this operation twice and leave a reusable project-local check. |
| Claude Code | /create-formal-verification-skill Verify that retries cannot apply this operation twice and leave a reusable project-local check. |
How it works
Selects a property from the requested code and a suitable model checker, verifier, or proof tool. Maps the verification to real code and atomic steps, states bounds and assumptions, and creates a project-local skill with reproducible commands. Tests each property against an isolated broken variant and rechecks the baseline afterward. Model mutation alone does not establish detection of implementation regressions; code-to-model correspondence must also be checked.
Expected result
A reusable skill, retained verification evidence, and a plain-language report of findings, counterexamples, limits, and how to rerun. Pre-existing findings remain visible and cannot substitute for detecting a new mutation.
Dependencies and limits
Explicit-only. Uses a distinct verify-formal-<area> name and checks canonical destinations before writing; existing runtime skills remain separate. Uses Codex or Claude Code project discovery and preserves the user's invocation policy. Hermes, unavailable discovery contracts, and restricted discovery writes use an ordinary project source folder with registration pending. Missing tools, timeouts, inconclusive checks, or undetected mutations leave a draft or blocked result. Product fixes require their own request, and checks must not be weakened to hide findings. This verifies stated properties within the tool's scope, not the whole system, and is separate from the launch/drive feature-map maintenance workflow.
Full create-formal-verification-skill instructions
Maintain a formal verification skill
Covered code changed without rerunning its formal checks, so a passing result may describe an old implementation.
| Host | Example invocation |
|---|---|
| Codex | $maintain-formal-verification-skill Refresh the lease checks after the refactor and distinguish drift from bugs. |
| Claude Code | /maintain-formal-verification-skill Refresh the lease checks after the refactor and distinguish drift from bugs. |
How it works
Finds checks by behavior, baselines them, and reviews history from each check's last verified commit, including moved functions, new state writers, and uncommitted changes. Updates correspondence from code, reruns, and demonstrates detection using isolated mutations. Changes a property only with evidence that the promise intentionally changed.
Expected result
Verification-only changes and per-check outcomes: in sync, updated, bug, or blocked. Records covered files and functions, evidence, and each check's checkpoint. Blocked checks retain their old checkpoint; dirty snapshots are recorded separately. A checkpoint confirms correspondence, not that the code is bug-free.
Dependencies and limits
Explicit-only. Uses existing project tooling and preserves invocation policy. Missing history requires disclosed full correspondence review before a new checkpoint; unavailable tools or incomplete detection leave affected checks blocked. Product fixes and publication need separate authorization. Generated-input fallbacks remain testing rather than formal proof.
Full maintain-formal-verification-skill instructions
Statechart
Make a TLA+ model and its counterexamples readable to someone deciding what the app should do.
| Host | Example invocation |
|---|---|
| Codex | $statechart Draw the sync model and its counterexamples with Pele; reproduce findings in the code before calling them bugs. |
| Claude Code | /statechart Check the upload model and show the shortest path to each broken rule. |
| Hermes | Load zstack:statechart with skill_view and provide the model. |
How it works
scripts/statechart SPEC.tla [--config FILE] [--view FILE] [--out DIR] checks the model with TLC on the config as written, rechecks after each broken rule, and draws the state graph as Mermaid. A view module labels states and steps in the reader's words and can supply nested states, parallel regions, phase colors, and a title. Above 500 states, the chart draws only paths to broken rules. Rules that a bounded or failed rerun leaves unchecked are listed explicitly rather than treated as passing. A TLC error on the config as written stops the command with exit 2 and no output. Temporal properties that break on a loop show the path into the loop and the repeating states.
Expected result
A report.md with state counts and numbered bug paths, chart.mmd (when the chart fits or a rule breaks), tlc.out, and per-rule TLC output for paths the chart cannot show. Pele SVG rendering is optional. Findings remain model counterexamples until an isolated test or live reproduction establishes the actual code behavior.
Dependencies and limits
Explicit-only on all three hosts. Requires Node 22.18–22.x or 23.6+, Bun 1.1.5+, or Deno 2+, and Java 11+. Set TLA2TOOLS (and optionally JAVA) to use a local TLC; otherwise the first run installs checksum-pinned tla2tools 1.7.4 and Temurin 25 through mise, which needs the network and approval. The output folder must be new or empty, and only exits 0 and 1 write output. Pele is installed in an owned scratch folder and is not a plugin dependency. The from-chart skeletons, --fixed/--against comparisons, and Mermaid PNG flags were removed with upstream's compiler replacement. Review view modules as executable code. Product edits and publication need their own authorization.
Create a verification skill
The project lacks a reusable way to launch, drive, and prove real user behavior.
| Host | Example invocation |
|---|---|
| Codex | $create-verification-skill Create a verification skill for this CLI using its existing PTY tests. |
| Claude Code | /create-verification-skill Create a verification skill for this CLI using its existing PTY tests. |
How it works
The agent inspects the project’s real launch commands, interaction tools, observables, and isolation requirements. It creates a project-local verify-<app> skill with Launch, Doctor, Drive, Evidence, and Cleanup instructions, plus a feature index and initial feature files. Launch must reach a usable state. For stateful targets, prove launch from empty isolated state (an empty data directory where applicable), including any required migrations and seed data. Missing seed prerequisites fail the journey and are recorded in the feature’s Gotchas without weakening the journey. It runs one mapped feature end to end and invokes bundled helpers from a different working directory to check portability.
Expected result
A verification skill and feature map with a demonstrated launch, health check, user action, evidence capture, and cleanup. Evidence survives teardown.
Dependencies and limits
Prefers existing project tooling. If the checkout cannot build or start, the agent reports the blocker before generating executable instructions. A draft can remain unverified until that blocker is resolved. Product fixes need authorization. Discovery-directory restrictions leave registration pending. An unexecuted generated skill is a draft. When proof uses a stand-in, setup steps supplied by that stand-in but absent on the real target remain unverified until exercised on the real target.
Full create-verification-skill instructions
Figure it out
A large migration or multi-part task has no narrower playbook and needs an auditable workflow.
| Host | Example invocation |
|---|---|
| Codex | $figure-it-out Plan and carry out this API migration in independently verifiable units, preserving existing behavior. |
| Claude Code | /figure-it-out Plan and carry out this API migration in independently verifiable units, preserving existing behavior. |
How it works
The agent frames a falsifiable definition of done, scope, blockers, and rigor. It writes a phase list, captures a baseline, and builds verification before features. Each unit follows a hypothesis, small change, measurement, and keep-or-revert loop. Decisions enter a show-me-your-work trail as they occur, followed by a whole-product check.
Expected result
The designed playbook, rigor rationale, decision-trail path, verified outcomes, and remaining work.
Dependencies and limits
Starts from z-mode principles. Uses architect for consequential unresolved designs and native workers only across isolated seams. Explicit planning checkpoints remain binding. INCONCLUSIVE is not a passing result. Trail publication requires authorization.
Full figure-it-out instructions
How
You need runtime behavior, an onboarding explanation, or a decision about where code belongs.
| Host | Example invocation |
|---|---|
| Codex | $how Explain how webhook admission reaches the execution queue, with entry points and ownership. |
| Claude Code | /how Explain how webhook admission reaches the execution queue, with entry points and ownership. |
How it works
The agent anchors the question in entry points, callers, data structures, and boundaries. Narrow questions trace directly. Broad questions use 2–3 independent read-only angles, then reconcile the returned slices against code.
Expected result
A cited overview, key concepts, runtime flow, relevant files, and gotchas at the question’s scale.
Dependencies and limits
Read-only. Missing delegated slices remain gaps. Uses why for historical motivation because code alone does not prove intent.
In your own words
Your request is long or messy, and you want to see whether the agent understood your goals and the problem.
| Host | Example invocation |
|---|---|
| Codex | $in-your-own-words <long or messy ask> |
| Claude Code | /in-your-own-words <long or messy ask> |
| Hermes | Load zstack:in-your-own-words with skill_view and restate my goals: <long or messy ask> |
For native plugins, select zstack:in-your-own-words through the Codex skill picker or invoke /zstack:in-your-own-words in Claude Code.
How it works
The agent restates, in its own words, what it thinks your goals are and what problem you are trying to solve.
Expected result
Your goals and the problem, restated in the agent's words. If it got something wrong, correct it in your next message.
Dependencies and limits
Explicit-only on all hosts. The restatement covers your goals and problem, not the agent's own reply; use bro to simplify the agent's last message. The skill asks only for the restatement; it does not wait for confirmation.
Full in-your-own-words instructions
Interrogate
You want an adversarial review of a defined diff, set of files, or design.
| Host | Example invocation |
|---|---|
| Codex | $interrogate Review the working-tree queue changes for correctness. Return findings only. |
| Claude Code | /interrogate Review the working-tree queue changes for correctness. Return findings only. |
How it works
The agent pins scope and restates the intent, then gives 2–3 independent adversaries their own worktrees outside your checkout. When a Codex or Claude CLI from a family other than the coordinator's is installed, one adversary runs through it for a second model family, unless that CLI's sandbox blocks the project's checks. Each must prove every finding with a run against the real code. The agent reruns each finding's command in a fresh isolated worktree, keeps only failures that break the intent, and merges shared root causes.
Expected result
The intent, each verified finding with its command, output, broken intent, and the adversaries that found it, and a Dropped list counting each adversary's unproven findings and why.
Dependencies and limits
Review does not apply fixes. Your checkout and unrelated dirty work stay untouched; the agent removes the worktrees it created. Adversaries run code, so findings need a runnable target. Style, naming, and hypothetical risks are dropped. If native agents are unavailable, the review is labeled parent-only.
Maintain a verification skill
An existing verification skill or feature map may have drifted from the application.
| Host | Example invocation |
|---|---|
| Codex | $maintain-verification-skill Audit this project’s verify-admin skill and exercise every mapped feature. |
| Claude Code | /maintain-verification-skill Audit this project’s verify-admin skill and exercise every mapped feature. |
How it works
The agent locates the canonical verification directory, reconciles its feature index, and assigns bounded read-only source readers per feature. The coordinator merges recipes and owns all live driving. It health-checks instances, captures evidence for every feature, cleans failed attempts, and re-drives every harness correction.
Expected result
A clean, changed, or blocked outcome, with source and live coverage, proven local corrections, unreachable prerequisites, and product gaps.
Dependencies and limits
Edits stay inside the verification skill directory. Product regressions are reported, not repaired or hidden in docs. Every feature needs a live pass. Several candidate skills require target selection. No target routes to create-verification-skill.
Full maintain-verification-skill instructions
No comments
You want to review redundant comments, correctness suppressions, or workarounds.
| Host | Example invocation |
|---|---|
| Codex | $no-comments Review unnecessary comments in the parser. Do not edit files. |
| Claude Code | /no-comments Review unnecessary comments in the parser. Do not edit files. |
How it works
An available comment-sicko role or equivalent read-only native reviewer returns locations, proposed removals, refactor targets, evidence, and exceptions. Role selection, scoped briefs, and reporting capability gaps follow the native contract. The parent checks claims and investigates ambiguous constraints through how or why. Authorized edits fix root causes before removing suppressions and then verify replacements.
Expected result
Accepted and rejected findings, any explicitly authorized edits, proof, and unresolved constraints.
Dependencies and limits
Protects legal headers, public contracts, external constraints, and necessary suppression explanations. Review alone does not authorize deletion or unrelated refactors.
Z mode
You want evidence-driven engineering for the current conversation.
| Host | Example invocation |
|---|---|
| Codex | $z-mode Fix the reconnect bug using the project’s existing harness. |
| Claude Code | /z-mode Fix the reconnect bug using the project’s existing harness. |
How it works
The mode grounds work in runtime flow and callers, selects only the relevant playbook, and loads principles when they change a decision. Simple tasks run directly. Requested delegation uses the selected execution method with exclusive write scopes, terminal results, and artifact inspection. Enable Herdr once for the session to use it alongside existing playbooks; native execution remains the default. Checks target the real outcome.
Expected result
Task-sized changes or a scoped investigation, with actual evidence, meaningful tradeoffs, and explicit proof gaps.
Dependencies and limits
Persists until you say stop z-mode or choose another style, subject to retained conversation context. Plan-only requests remain read-only. Commits, pushes, messages, merges, and deployments remain separate requested actions. Optional plugins are not assumed.
Ketchup
You want to catch up on this chat since your last message and decide anything still awaiting you.
When one decision covers several items, each item retains its identity, user impact, evidence, risk, and any differing recommendation. A count or link cannot replace this context; the reply grows to cover the whole batch.
| Host | Example invocation |
|---|---|
| Codex | $ketchup |
| Claude Code | /ketchup |
| Hermes | Load zstack:ketchup with skill_view and catch me up on this chat. |
For native plugins, select zstack:ketchup through the Codex skill picker or invoke /zstack:ketchup in Claude Code.
How it works
The agent summarizes work after your preceding human message; notifications and worker reports do not move that boundary. It takes a current read-only snapshot of relevant PRs, CI, branches, and existing agents where tools and access permit. Earlier unresolved decisions are included; later answers remove resolved ones.
Expected result
A plain-language reply aiming for about 150 words, including visual and table text, with more space when needed for every unresolved decision, material risk, and evidence gap. It opens with how many things need you; if nothing changed and nothing waits, it says so and stops, retaining any coverage gap. Otherwise it summarizes consequential results, current status, and remaining agent work. Each decision includes feasible options and tradeoffs, a recommendation and reason, prerequisites or irreversible effects, and the consequence of no answer, or explicitly says it is unknown. Report an unasked irreversible step waiting on the user without inventing a no-answer consequence. Visuals are optional: use existing captures, tables, diffs, trees, or diagrams only when they clarify status or a decision better than prose. Missing captures and unsupported comparisons remain explicit gaps. Supporting links supplement a self-contained summary.
Dependencies and limits
Explicit-only on all hosts. Uses available conversation evidence, native history/status readers, and GitHub readers or an already available authenticated gh. Missing history, tools, access, captures, or rendering are disclosed. HumanLayer's show-me guidance is adapted into a bundled reference; no separate skill installation is needed. Comparisons stay in chat and use recorded evidence, with proposals and unknowns labeled. The invocation stays read-only, writes no visual files, launches no app or job, creates no new agents, and takes no proposed action. Existing workers can continue. Recall rebuilds context across sessions.
Recall
You need current context across recent sessions before starting or resuming work.
| Host | Example invocation |
|---|---|
| Codex | $recall Recall my queue work in this repository from the last week and identify the next step. |
| Claude Code | /recall Recall my queue work in this repository from the last week and identify the next step. |
How it works
The agent pins workspace, topic, and time window before reading scoped session evidence. It extracts goals, decisions, open threads, corrections, and artifact pointers, then checks live branch, SHA, dirty state, and relevant PR status.
Expected result
A short context capsule, accurate per-thread status, recurring blockers, and one next action.
Dependencies and limits
Unavailable history and digest-based evidence are labeled. Adjacent work is excluded unless it blocks the task. A single-session resume belongs to session-pickup. A catch-up on the current chat belongs to ketchup. Habit mining belongs to automate-me.
Reflect
The active session contains lessons worth proposing as durable skill guidance.
| Host | Example invocation |
|---|---|
| Codex | $reflect Reflect on this session and propose skill changes. Do not apply them yet. |
| Claude Code | /reflect Reflect on this session and propose skill changes. Do not apply them yet. |
How it works
Independent read-only judgment, tooling, and divergent reviewers examine scoped session evidence. The coordinator verifies moments, reads target skills, prefers existing homes, and separates missing guidance from missed invocation. It presents Accepted, Rejected, and Backlog proposals. Approved edits receive metadata, link, and relevant scenario validation.
Expected result
Grounded proposals or authorized local edits, with the evidence basis and remaining gaps.
Dependencies and limits
Applying edits requires approval or an existing explicit request that covers them. Filing backlog, writing memory, commits, and PRs are separate actions. Digest fallback and unavailable native reviewers remain visible.
Setup zstack
You need to inspect zstack discovery or configure requested native agent models.
| Host | Example invocation |
|---|---|
| Codex | $setup-zstack Inspect Codex zstack discovery and report installed versus loaded agents without changing settings. |
| Claude Code | /setup-zstack Inspect Claude Code zstack discovery and report installed versus loaded agents without changing settings. |
How it works
The agent checks the selected host’s discovery and installation collisions. For a requested model change, it verifies current native configuration and documentation, previews exact fields, and preserves unrelated settings. A fresh session checks explicit invocation and a delegated read-only investigation.
It also inspects project formal tooling (models, proofs, and verifier commands) and skills. When relevant races, retries, or state invariants lack suitable coverage, it offers create-formal-verification-skill once. Acceptance leads to instructions for the user to explicitly invoke the generator through the selected host; declining continues setup.
Expected result
An installed-versus-loaded report, requested configuration changes, collision reports, and actual host verification or its gap, plus a conditional formal-verification offer and user-invocation handoff. Setup does not generate the verification skill merely because the offer was accepted.
Dependencies and limits
Defaults to native model inheritance. Bundled roles are z-agent and comment-sicko. Modified agent copies survive installer reruns. Model entitlement, unsupported identifiers, and effort-suffixed IDs are never inferred. Delegation may be unavailable.
Full setup-zstack instructions
Sharpen
You want a stronger prompt to review before starting a task.
| Host | Example invocation |
|---|---|
| Codex | $sharpen Write a migration guide for the v2 config format; leave the rollout strategy open. |
| Claude Code | /sharpen Write a migration guide for the v2 config format; leave the rollout strategy open. |
| Hermes | Load zstack:sharpen with skill_view, then rewrite: Write a migration guide for the v2 config format; leave the rollout strategy open. |
How it works
Briefly identifies what an expert would add, the original source material, specific methods and tools, and an independent check. It then returns a complete rewritten prompt while preserving the goal, constraints, and choices that belong to you.
Expected result
A prompt ready for your review and later submission. The underlying task is not performed. If you want execution, send the reviewed rewrite separately, optionally with z-mode.
Dependencies and limits
Explicit invocation only. Unknown source contents, tool availability, and user decisions stay conditional or unresolved. Sharpen does not invoke another workflow or treat the prompt being rewritten as authorization to act.
Show me your work
A long-running task needs a durable record of decisions and their evidence.
| Host | Example invocation |
|---|---|
| Codex | $show-me-your-work Keep a local decision trail for this migration at .audit/api-migration.tsv. |
| Claude Code | /show-me-your-work Keep a local decision trail for this migration at .audit/api-migration.tsv. |
How it works
The agent maintains one append-only TSV with ts, phase, decision, why, evidence, and result. It logs decisions and checkpoints rather than every command. Each run tracks its own start boundaries. Before handoff, it audits its rows against scoped session evidence and obtains an independent read-only trail review when available.
Expected result
A canonical local log and an Attention section identifying the reviewer and flags, or no flags. Wrong entries receive superseding rows.
Dependencies and limits
Defaults to an uncommitted working artifact. The bundled log helper resolves from the loaded skill’s real directory and sanitizes tabs, newlines, and spreadsheet formula prefixes. History gaps and parent-only review are labeled. Commit requires authorization.
Full show-me-your-work instructions
Swarm
The task needs partitioned parallel coverage, independent races, or a mixture.
| Host | Example invocation |
|---|---|
| Codex | $swarm Use three read-only workers to trace API, database, and UI effects of this change. |
| Claude Code | /swarm Use three read-only workers to trace API, database, and UI effects of this change. |
How it works
The agent declares the done predicate and selection rule, then creates the smallest useful native worker set. Briefs define slices, write limits, exclusive output, checks, and PASS, ISSUES, or BLOCKED receipts. It waits for terminal results, inspects receipts, and drains or cancels remaining writers after a race.
Expected result
A consolidated table, evidenced findings, the selection rule, actual identities when exposed, and gaps.
Dependencies and limits
Workers stay within host concurrency and isolate worktrees, data, and ports where needed. Cloud placement is not assumed. Measurements require exact SHAs, workloads, samples, and methods. Missing required coverage is not a pass.
TDD
You requested a failing regression test, or the bug has an obvious cheap local test target.
| Host | Example invocation |
|---|---|
| Codex | $tdd Add a failing regression test for duplicate admission, then fix it and show the before-and-after result. |
| Claude Code | /tdd Add a failing regression test for duplicate admission, then fix it and show the before-and-after result. |
How it works
The agent identifies intended behavior and the smallest reproduction, selects an existing narrow test path, and runs the new test before changing production code. It verifies the failure cause, applies a focused fix, and reruns the regression and relevant nearby checks.
Expected result
Named failing-before and passing-after evidence, or the reason that a practical failing test was unavailable and the closest executable check used.
Dependencies and limits
Does not create broad harnesses, brittle mocks, expensive infrastructure, or unrelated fixtures to force TDD. Existing assertions are not weakened to match a bug. Deterministic checks are preferred for flaky behavior.
Teach
You want to understand a subsystem or body of work at your own pace.
| Host | Example invocation |
|---|---|
| Codex | $teach Teach me how queue coalescing works and why this project uses it. |
| Claude Code | /teach Teach me how queue coalescing works and why this project uses it. |
How it works
The agent uses how for mechanisms and scoped why research for rationale, then combines their findings in plain language. It starts with a small complete explanation and adds depth as you ask. Diagrams grow progressively when the flow or layout benefits from them.
Expected result
The explanation itself, grounded in the real code and retaining why’s confidence language.
Dependencies and limits
Read-only teaching does not change code. Narrow questions need less research. Broad questions may use bounded read-only slices. Unavailable diagram tools get a disclosed textual fallback. No quizzes or forced pacing.
Technical writing
You need documentation, an RFC, a README, or a review of technical prose.
| Host | Example invocation |
|---|---|
| Codex | $technical-writing Review docs/install.md for clear reference structure and unambiguous instructions. |
| Claude Code | /technical-writing Review docs/install.md for clear reference structure and unambiguous instructions. |
How it works
The agent selects a Diátaxis mode, applies Google developer style, limits each sentence to a clear thought or instruction, and removes ambiguous syntax with Global English rules. It uses actual repository symbols and applies unslop to the writing.
Expected result
Clearer technical writing with a consistent document purpose, concrete names, readable sentences, and preserved meaning.
Dependencies and limits
The rules serve clarity rather than mechanical compliance. Product UI copy follows product guidelines. New jargon offenders are proposed with the diff, not silently added to unslop. Counts and trees must match the repository state.
Full technical-writing instructions
Python best practices
You want Python scripts, packages, or reviews guided by uv, Ruff, mypy, extensive type hints, and Pydantic validation.
| Host | Example invocation |
|---|---|
| Codex | $python-best-practices Add a typed JSON importer with Pydantic boundary validation. |
| Claude Code | /python-best-practices Add a typed JSON importer with Pydantic boundary validation. |
How it works
Uses uv for Python operations wherever supported, Ruff for both formatting and linting, and mypy for type checking. External data is parsed into typed Pydantic models or through TypeAdapter. Examples cover standalone scripts, packages, checking configuration, boundary tests, and internal domain variants.
Expected result
Scoped Python changes with precise annotations, validated inputs, and reported formatting, lint, type-check, and runtime-test evidence.
Dependencies and limits
Codex agents may select this skill automatically for Python work; Claude Code invocation remains explicit. Reuses type-system and boundary principles. Examples target Python 3.11+ and Pydantic 2; adapt to the project's supported versions. Existing packaging is not silently migrated, and static checks do not prove runtime behavior.
Full python-best-practices instructions
TypeScript best practices
You want TypeScript changes or review guided by explicit domain and boundary types.
| Host | Example invocation |
|---|---|
| Codex | $typescript-best-practices Review src/admission.ts for invalid states, unsafe casts, and boundary validation. |
| Claude Code | /typescript-best-practices Review src/admission.ts for invalid states, unsafe casts, and boundary validation. |
How it works
The agent applies discriminated unions, validated brands, constructive types, and the simplest total representation where they improve correctness. It treats external data as unknown, prefers installed schemas and derived types, uses narrowing and exhaustiveness, and verifies behavior through real tests.
Expected result
Scoped TypeScript findings or requested changes that encode valid states and validate data at entry points.
Dependencies and limits
Uses principle-type-system-discipline when it changes a decision. New stronger types are justified where loose types force casts or impossible-state failures. Object arguments have hot-path exceptions. Shipped telemetry uses structured logging.
Full typescript-best-practices instructions
Unslop
Requested writing contains filler, jargon, repetitive phrasing, or other AI patterns.
| Host | Example invocation |
|---|---|
| Codex | $unslop Rewrite this PR description in plain English while preserving its technical claims. |
| Claude Code | /unslop Rewrite this PR description in plain English while preserving its technical claims. |
How it works
The agent scans the stable pattern catalog and rewrites wording, structure, punctuation, and tone. It removes vague attribution and generic conclusions, preserves meaning, and keeps concrete mechanisms and necessary confidence qualifiers.
Expected result
Edited text that matches the intended tone without changing its claims.
Dependencies and limits
A style rewrite does not establish facts or authorize editing unrelated files. Pattern rules are stable identifiers used by companion skills.
Why
You need design rationale, regression history, postmortem context, or evidence for a threshold.
| Host | Example invocation |
|---|---|
| Codex | $why Why does the queue retain tied timestamps? Scope research to Git history and relevant PRs. |
| Claude Code | /why Why does the queue retain tied timestamps? Scope research to Git history and relevant PRs. |
How it works
The agent traces blame, renamed files, substantive commits, and PR discussion. It maps available evidence sources to relevant categories within the requested topic and time window. Broad independent sources may use read-only investigators. It reconciles citations and contradictions while separating direct evidence, conclusions, inference, speculation, and unknowns.
Expected result
Cited facts, inferences, competing hypotheses when useful, source gaps, and actual sources searched. Before a change, it derives Preserve, Change, Avoid, and Risk constraints.
Dependencies and limits
Uses how for runtime behavior. Authentication and connectors are not assumed. Unavailable and unsearched categories are reported. Research does not authorize contacting people or editing records.