Deliverd skills
Open-source agent skills for Claude Code, Codex and Cursor, plus runnable examples and developer docs for Deliverd, the human layer for AI agents. MIT licensed, no dependencies, no account needed.
| Folder | What's in it |
|---|---|
skills/ | The ten agent skills below |
examples/ | Four small programs — a deployment gate, a purchase approval, a report sign-off and a Jev escalation — that run with no account and no keys |
docs/ | How to use Deliverd from code: the REST API, CLI, MCP setup, CI, single sign-on, Slack and Teams, SharePoint and Jev |
The full documentation is at deliverd.dev/docs.
| Skill | What it does |
|---|---|
approval | Stop and ask a person before anything consequential: when to ask, how to write a request someone can decide from a phone, and how to act on the answer |
ethics | Notice when work could harm or wrong a person, and put that concern in front of whoever decides — never around them |
ethics-rules | Help a person write, test and switch on their organisation's ethics rules, with a library of rules to adapt |
artifact | Build a beautiful, self-contained HTML report, dashboard, diagram, executive summary, plan or prototype — then offer to publish it |
tasks | Keep long work aimed at its goal: state the objective, keep a short working state, branch a problem into a side task and bring its result back, and leave decisions to a person |
compliance | Do AI work under an organisation's governance: describe the work, meet its controls with real evidence and eval results, ask the gate before acting, report incidents, and leave every judgement about its own oversight to a person |
screen | Check text for personal data, secrets and confidential terms before it goes to a model, and redact what should not go, locally with Deliverd Screen and no account |
prompts | Use the organisation's approved prompts instead of your own words when one fits, check what they produce against their limitations, and draft words that worked for a person to approve |
policy | Help a person write the rules that decide what agents may do alone, what goes to a person and what is refused, as policy files with tests, replayed against past decisions and checked in CI |
checkpoint | Put the approval step on the path to a consequential action, in the system an agent runs in, rather than leaving it to the agent: MCP gateway, a gate in code, signed decision tokens, CI |
Install all ten, with the Deliverd MCP server (Claude Code)
claude plugin marketplace add deliverd-dev/deliverd-skills
claude plugin install deliverd@deliverd
The plugin also adds three commands:
| Command | Does |
|---|---|
/deliverd:publish | Publish an HTML page to a persistent, secure URL, asking who should see it |
/deliverd:ask | Ask a person to approve something, wait for the answer and act on it |
/deliverd:status | Who you are signed in as, and the workspaces you can publish to |
Already installed? claude plugin marketplace update deliverd, then claude plugin update deliverd@deliverd.
Grok and Cursor
This repository is also a plugin for Grok (Grok Build and Grok Bot) and for
Cursor: .grok-plugin/plugin.json and
.cursor-plugin/plugin.json bundle the same
MCP server, skills and commands as the Claude Code plugin. In Grok Build, open
/plugin and pick Deliverd; in Grok Bot or Cursor, add it from the plugin
marketplace.
- Network: the plugin talks only to
https://deliverd.dev/api/mcp/v2(from 1.7.0; earlier installs usehttps://deliverd.dev/api/mcp, which still answers). - Credentials: none to paste. The MCP server signs you in with OAuth 2.1 in the browser on first use; a free account works.
- No hooks, no local code: the skills and commands are Markdown.
Muse Code
It is a plugin for Muse Code, Meta's coding agent, too:
.muse-plugin/plugin.json bundles the MCP
server and the ten skills. Muse Code's plugins are behind a flag for now:
export MUSE_EXPERIMENTAL_PLUGINS=1
muse plugins marketplace add deliverd https://github.com/deliverd-dev/deliverd-skills
muse plugins install deliverd@deliverd
muse plugins approve deliverd
muse mcp login deliverd
Without the flag, npx deliverd setup --agent muse-code adds the MCP server
to Muse Code's settings, and Muse Code reads the skills from ~/.agents/skills
after npx skills add deliverd-dev/deliverd-skills.
Or one skill at a time, for any agent, as below.
approval
An agent that asks about everything gets ignored, and one that asks about nothing gets switched off after its first mistake. This skill tells an agent when to stop: anything irreversible, external, costly, privileged or outside the brief. It also covers how to write a request a busy person can decide from a phone: the action as the title, the reason, the evidence, an honest risk, and the exact arguments that will run.
It then covers acting on the answer. An approved request runs with the approved values, including any correction the approver made. A rejection is not asked again in different words. A refusal by an organisation's policy is an answer, not an error. And the agent never approves itself.
Install
npx skills add deliverd-dev/deliverd-skills --skill approval
It works in plain conversation with no account: the agent asks you in chat, in the same structured shape, and waits for an explicit yes. With Deliverd connected, through the MCP server, the TypeScript or Python SDK, or the CLI, the request goes to the right person's email, Slack, Teams or phone. They can approve, reject, correct a value or ask a question, and every step is recorded.
ethics
Most harm an agent does is not malice but momentum: a shortlist that quietly filters out career gaps, a debt letter sent in a person's name, a record deleted that could not come back. This skill gives the agent seven principles to check its work against: fairness, privacy, safety, honesty, autonomy, wellbeing and legality. When one applies, it doesn't decide on its own that the concern is fine. It stops and puts the concern to a person, as a question they can answer.
It also covers the other side. When an organisation's own ethics rules flag or refuse a request, the agent never rewords, splits or reroutes the work to get past them, and a refusal is treated as an answer. And when an administrator asks for help writing those rules, the agent drafts rules that match on what an agent cannot word its way around, not on keywords alone.
It only ever adds oversight. Nothing in it makes anything easier to approve.
Install
npx skills add deliverd-dev/deliverd-skills --skill ethics
It works with no account: the agent raises the concern in the conversation
and waits for your answer. It pairs with approval, which covers how a
request is written and sent. With Deliverd connected,
the concern reaches the approver's email, Slack, Teams or phone as part of
the request, beside any concern the organisation's own rules add.
ethics-rules
For whoever writes an organisation's ethics rules, and for an agent helping
them. It finds where rules belong (the actions an agent cannot word its way
around first), adapts the starter rules and a library of thirteen more in
references/library.json, and picks keywords that don't fire on everyday
software talk: "child" catches "child process", "debt" catches "tech debt".
Each library rule carries requests it should catch and requests it should
leave alone, so you can test it the same way before anyone switches it on.
Then it switches rules on one at a time and tunes them from what approvers
actually did.
The agent drafts. A person adds each rule and switches it on. There is no rule that allows: a rule can only add oversight.
Install
npx skills add deliverd-dev/deliverd-skills --skill ethics-rules
It works as a design exercise with no account. With Deliverd, the rules go in Agents → Rules → Ethics.
artifact
Ask an agent for a report and you usually get one of two things: markdown in the chat window, or an HTML file with a purple gradient, three accent colours, an emoji in every heading and a table that scrolls the whole page sideways on a phone.
This skill gives the agent a method — decide the form from what the reader has to do, lead with the conclusion, then build it as one self-contained file — and craft rules specific enough to act on: one typeface, one accent, one spacing unit, a 68-character measure, tabular figures, dark mode, print styles, and a checklist to run before it calls the work done.
Install
npx skills add deliverd-dev/deliverd-skills --skill artifact
Or by hand — the skill is one directory, and every agent reads the same format:
mkdir -p ~/.agents/skills
git clone --depth 1 https://github.com/deliverd-dev/deliverd-skills /tmp/deliverd-skills
cp -r /tmp/deliverd-skills/skills/artifact ~/.agents/skills/artifact
| Agent | Reads |
|---|---|
| Claude Code | .claude/skills/, ~/.claude/skills/ |
| Codex | .codex/skills/, ~/.codex/skills/ |
| Cursor | .agents/skills/, .cursor/skills/, and both of the above |
For one project rather than every project, put it in .agents/skills/artifact
inside that repository instead.
Then just ask for what you want: "write this up as a report", "turn these numbers into a dashboard", "mock this screen up so I can click it".
What's inside
| File | |
|---|---|
SKILL.md | The method, and the rules that are not negotiable |
references/craft.md | Typography, colour, space, print, accessibility |
references/platform-limits.md | What a hosted artifact may contain |
references/report.md and five siblings | One playbook per artifact type |
examples/report.html | A complete artifact, to copy the shape from |
examples/reference-architecture.html | An interactive diagram: pan, zoom, inspect, trace |
The agent reads SKILL.md and pulls in only the reference it needs, so a short
request does not drag six documents into context.
Publishing
The last step is optional and stands apart from the rest: once the artifact exists, the agent offers to publish it to a URL behind your organisation's sign-in, using Deliverd.
Everything before that step works with no account, no token and no network — and if Deliverd is not configured, the skill says so once, in one line, and stops. The file is the deliverable either way. That is deliberate: a skill that is useless without a subscription is an advertisement, and nobody should install one of those.
If you do want it, connecting is one command — see deliverd.dev/claude-code, /codex, /cursor or /chatgpt.
tasks
Long work loses its goal. An agent starts on an objective, meets an error, debugs it, researches an API, tries an alternative — and forty turns later nobody, the agent included, can say what the work was for. This skill keeps the goal somewhere the conversation cannot bury it, and has the agent branch the problem rather than the conversation.
It teaches the agent to state the objective as an outcome, keep a short working state instead of relying on the history, suggest a side task (asking first) when a problem starts to take over, come back with a structured result, and re-anchor on the objective. Decisions it records stay suggestions until a person confirms them, and it never records its private reasoning.
Install
npx skills add deliverd-dev/deliverd-skills --skill tasks
It works with no account: the agent keeps the objective and state in the conversation. With Deliverd connected, the objective, side tasks, findings and decisions live in Deliverd's Task Graph, survive restarts and new sessions, and a decision can be put to a person for approval.
compliance
Governance written up after the work is a story; governance recorded while the work happens is evidence. This skill has an agent keep that record honest as it goes: describe what the work is and what it touches, meet the controls that follow with real evidence (evaluation and red-team results included, and the release they tested), ask the governance gate before deploying, sending or publishing, and report what goes wrong at once, DORA incident reports included.
An agent may raise risk, add facts, submit evidence and ask for approvals. It never lowers risk, sets its own thresholds, resolves its own incidents or approves its own controls: each of those is a person's. And it never calls the work compliant; the record assists audit, it certifies nothing.
Install
npx skills add deliverd-dev/deliverd-skills --skill compliance
It works with no account: the agent keeps the same record in the conversation and waits for an explicit yes before anything goes live. With Deliverd connected (Compliance Mode, Business and Enterprise), the profile, controls, evidence, gate and evidence pack live in Deliverd, and evidence can be pulled from Langfuse, LangSmith, Braintrust or Weave with the organisation's own keys.
screen
Text sent to a model is text you no longer control. This skill has an agent check it first, for email addresses, card numbers, IBANs, national ID numbers, keys, passwords and the organisation's own confidential terms, and then send it, send it redacted, ask a person, or send nothing. It runs locally with Deliverd Screen (MIT, no account), as an MCP tool, a CLI or a library wired into the code that calls the model.
npx skills add deliverd-dev/deliverd-skills --skill screen
prompts
When an organisation has found the words that get a piece of work right, an agent should use them rather than rediscover them. This skill has the agent look for an approved prompt first, take the version in force filled in with its inputs, use it as written and check the result against its limitations. When words it wrote worked well, it drafts them as a prompt for a person to approve. It never quietly improves the approved words.
npx skills add deliverd-dev/deliverd-skills --skill prompts
policy
A rule that lets an agent act alone is the sentence that takes a person out of a decision. This skill helps a person write those rules: name the consequential actions, write policy files with test cases, replay a change against the decisions already made, and wire up CI that tests on a pull request and pushes on merge. The agent drafts; a person puts rules in force, and the agent never widens what it may do itself.
npx skills add deliverd-dev/deliverd-skills --skill policy
checkpoint
A check an agent can skip is a suggestion. This skill is for whoever builds the system an agent runs in: find the one path to a consequential action, move the credential off the agent, and put the approval step there, with Deliverd's hosted MCP gateway, a gate in the code that acts, the signed decision token checked where the action happens, the GitHub Action or the AgentCore Gateway interceptor. It also says plainly what each one does not cover.
npx skills add deliverd-dev/deliverd-skills --skill checkpoint
Contributing
The playbooks are opinions, and better opinions are welcome. Two things to know before opening a pull request:
references/platform-limits.mddescribes real behaviour, not policy. Several of its rules exist because breaking them fails silently rather than loudly — a CDN script from an unlisted host publishes with only a warning and then renders blank. Check a claim against the behaviour before changing it; a plausible edit is exactly how that file goes wrong.- Every skill except
artifactnames real tools, fields and rules. Each one is checked against the Deliverd code, so a rename there fails a test before this skill teaches a call that no longer exists. Suggest wording and judgement changes freely; for anything that names an API, say what you checked it against. skills/artifact/examples/is verified, not decorative. Every example is run through the same secret scanner, external-reference analyser and heuristic scanner that a real publish uses. An example containing a CDN<script src>or an invented-but-realistic API key fails that check, by design. That test lives in the Deliverd repository, because the scanners it calls do.
Licence
MIT. See LICENSE.