dispatch-authoring skill
✓ VerifiedCreate structured handoff documents that allow coding agents to execute technical plans unattended and return reviewable draft pull requests.
- GitHub stars
- 1,066
- Files
- 6
- Last commit
- Nov 8, 2026
- License
- MIT
- Allowed tools
- —
What does the dispatch-authoring skill do?
dispatch-authoring creates structured handoff documents that allow coding agents to execute technical plans unattended and return reviewable draft pull requests.
Summary generated from the skill's SKILL.md.
About the dispatch-authoring skill
Creating structured handoff documents for unattended coding agents is what this skill enables. It allows developers to transform a technical plan or specification into a highly detailed brief that an AI can execute without human intervention. By generating a dispatch, you provide the necessary context for tools like Claude Code, Codex, or Cursor agents to work through a task and return a reviewable draft pull request.
An agent should load this skill when a user wants to queue work for overnight execution, write an agent brief, or turn a static spec into an actionable instruction set. The resulting markdown file follows a strict seven-part framework including a header, Gate 0 checks, scope, specification, verification, delivery rules, and stop conditions. This structured approach ensures the agent knows exactly what to build, how to prove it works using specific commands, and when to stop if it encounters ambiguity or destructive operations.
What an agent can do with it
- 1Queue overnight coding tasksA developer can use this to prepare a task for an agent to run in the background while they are away.
- 2Standardize agent handoffsTeams can use this skill to ensure every instruction given to an AI coder is comparable and measurable.
- 3Convert specs to executable briefsYou can turn a technical ticket or specification into a detailed set of instructions for an AI agent.
- 4Prevent unauthorized code mergesThe skill enforces a workflow where the agent only produces a draft pull request for human review.
Written from the skill's SKILL.md. Every use case is checked against a quote from it.
How to install the dispatch-authoring skill
Each command copies the skill folder into the client's skills directory. Restart the client or start a new session to load it.Install the dispatch-authoring skill in Plugin
/plugin marketplace add rampstackco/claude-skills /plugin install rampstack-skills@rampstack
Install the dispatch-authoring skill in Claude Code
~/.claude/skills/dispatch-authoring/SKILL.md
$ npx degit rampstackco/claude-skills/skills/dispatch-authoring ~/.claude/skills/dispatch-authoringInstall the dispatch-authoring skill in Claude Code (project)
.claude/skills/dispatch-authoring/SKILL.md
$ npx degit rampstackco/claude-skills/skills/dispatch-authoring .claude/skills/dispatch-authoringInstall the dispatch-authoring skill in Codex
~/.agents/skills/dispatch-authoring/SKILL.md
$ npx degit rampstackco/claude-skills/skills/dispatch-authoring ~/.agents/skills/dispatch-authoringInstall the dispatch-authoring skill in Gemini CLI
~/.gemini/skills/dispatch-authoring/SKILL.md
$ npx degit rampstackco/claude-skills/skills/dispatch-authoring ~/.gemini/skills/dispatch-authoringInstall the dispatch-authoring skill in Claude.ai
npx degit rampstackco/claude-skills/skills/dispatch-authoring dispatch-authoring && zip -r dispatch-authoring.zip dispatch-authoring # claude.ai → Customize → Skills → + → Create skill → Upload a skill → dispatch-authoring.zip
Automated checks
Ran Nov 14, 2026Full report →Passed our automated checks. That is not a security audit: read the scripts before you let an agent run them. How checks work
- Source repository linked and reachablerampstackco/claude-skills/skills/dispatch-authoring is public and reachablegithub.com/rampstackco/claude-skills/tree/HEAD/skills/dispat ↗Pass
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit 2026-11-08github.com/rampstackco/claude-skills/commits ↗Pass
- No download-and-run or credential theft in scriptsNo scripts in the skill folderPass
- No hidden instructions in SKILL.mdNo override phrases, safeguard switches, credential requests or hidden charactersgithub.com/rampstackco/claude-skills/blob/HEAD/skills/dispat ↗Pass
- Pre-approved tools are scopedNo pre-approved tools; the agent asks before using onegithub.com/rampstackco/claude-skills/blob/HEAD/skills/dispat ↗Pass
SKILL.md
View on GitHubDispatch Authoring
Write the handoff that lets a coding agent execute a plan unattended and come back with something a human can review. Agent-agnostic: works for Claude Code, Codex, or your coding agent of choice.
When to use
- Handing a planned change to a coding agent to run without supervision
- Queueing work to run overnight or in the background
- Turning a spec or a ticket set into an executable brief for an agent
- Standardizing handoffs so agent runs are comparable and measurable
- Reviewing a draft dispatch before it goes out
When NOT to use
- Writing the spec or PRD the dispatch implements (use
pm-spec-writing) - Writing the tickets the work is tracked in (use
ticket-authoring) - Reviewing the PR the agent returns (use
code-review-web) - Measuring how agent lanes perform over time (use
sprint-operations) - Pairing interactively with an agent on a small change. A dispatch is for unattended runs.
Required inputs
- The plan or spec the agent will implement
- The target repository, its visibility (public or private), and its default branch
- The repo's own contribution and authoring rules, if any
- The commands that prove the work: tests, lint, build, generators
- Who merges, and who cleans up afterward
If you do not know how the work will be verified, stop and find out. A dispatch without verification is a request for a confident guess.
The framework: 7 parts of a dispatch
A dispatch is read top to bottom by an agent with no other context. Each part answers one question the agent would otherwise have to guess at.
1. Header
Date, target repo, visibility, lane, branch name, and roles (who planned, who executes, who merges). The header lets the agent and the reviewer confirm they are talking about the same work.
2. Gate 0
Checks the agent runs before any write. Each one has a stop action if it fails.
- Correct remote, matched against the exact owner and repo name
- Correct visibility. A public repo needs extra care with what goes in it.
- Fresh worktree or branch, never the canonical checkout or the default branch
- Required reading, in order: the repo's authoring guide, contribution guide, the closest existing example, the scripts the dispatch will call
- Record how generated files are produced (which script, which flag). Never hand-edit a generated file.
3. Scope
What is in, then what is out, in that order. Name the files or areas that may change. Name the ones that must not.
4. Specification
The work, in enough detail that the agent does not need to ask. Point at the source spec instead of pasting it, unless the agent cannot reach it. Give constraints, conventions, and examples to match. Where you want judgment, say so and say what to optimize for.
5. Verification
How the agent proves the work is done: the exact commands, the expected result of each, and the files to read back. Ask for the output to be pasted, not summarized.
6. Delivery
Branch name, commit structure, held draft PR, PR body format, and cleanup responsibility. The agent opens the PR as a draft and stops. It never merges.
7. Stop conditions
When the agent must stop and report rather than proceed: a Gate 0 check fails, a required script is missing, the spec is ambiguous in a way that changes the result, an operation is destructive or irreversible (force-push, history rewrite, deleting data, touching the default branch).
Principles to state in every dispatch
- The agent never merges. A human reads the diff before it lands.
- Honest stop beats fabricated completion. A clear "blocked on X" is a good outcome. A plausible result built on a guess is a bad one.
- A successful tool message is not evidence. Read the file back. Run the check. Paste the output.
- One dispatch per worktree. Parallel dispatches in one checkout collide.
- Derive counts and lists at run time. Write "every skill in the folder," not "the 42 skills." Hardcoded numbers go stale between writing and running.
Workflow
- Confirm the plan is ready. The spec exists, the tickets are ready, and the verification commands are known.
- Copy the template. Start from
references/dispatch-template.md. - Fill the header. Double-check owner, repo, and visibility against the actual remote.
- Write Gate 0. Use
references/gate-zero-checklist.md. Every check has a stop action. - Write scope, then specification. In then out. Then the work, with examples to match.
- Write verification. Exact commands, expected results, files to read back.
- Write delivery and stop conditions. Draft PR only. Name the PR body format from
references/pr-body-template.md. - Read it as the agent. Find every place you would have to ask a question. Answer it in the dispatch or turn it into a stop condition.
- Send it, then log it. Record the dispatch date and outcome so
sprint-operationscan measure the lane.
Failure patterns
- No Gate 0. The agent writes to the wrong repo, the wrong branch, or the canonical checkout.
- "Make it work" as verification. The agent will report success. Name the commands and the expected output.
- Permission to merge. Removes the only human check. Never grant it in a dispatch.
- Hardcoded counts. "Update all 37 entries" is wrong the day someone adds the 38th.
- No stop conditions. The agent improvises through a missing script or an ambiguous spec, and the PR looks finished.
- Spec pasted and paraphrased. Two versions of the truth. Point at the source.
- Scope with only "in." Agents tidy up nearby code unless told not to.
- Two dispatches in one checkout. Edits collide and the PRs contain each other's work.
- Trusting the agent's summary. Read the diff and the pasted verification output, not the closing paragraph.
- Private details in a public repo dispatch. Internal paths, client names, and private repo names leak into commits and PR bodies.
Output format
One markdown file per dispatch, named dispatch-[short-name]-[YYYY-MM-DD].md, using the seven parts in order. Keep it in the planner's own records, not in the target repo. The agent's PR body follows references/pr-body-template.md.
Reference files
references/dispatch-template.md- Fillable seven-part dispatch. Start every dispatch here.references/gate-zero-checklist.md- The pre-write checks and the stop action for each. Use when writing Gate 0 and when reviewing a dispatch.references/pr-body-template.md- The draft PR body the agent returns, with verification output and open questions. Paste into the dispatch's delivery section.references/example-dispatch-new-skill.md- A filled dispatch adding one skill to a fictional skills catalog. Read to calibrate detail and tone.
Files in this skill
6dispatch-authoring skill questions
How do I install the dispatch-authoring skill in Claude Code?
Run npx degit rampstackco/claude-skills/skills/dispatch-authoring ~/.claude/skills/dispatch-authoring. Claude Code loads skills from ~/.claude/skills (all projects) or .claude/skills (one project) and uses a skill when its description matches the task.
Which clients can use the dispatch-authoring skill?
Any client that reads the Agent Skills format: Claude Code, Codex, Gemini CLI and claude.ai (upload as a ZIP). Scripts in the skill run only where the client allows code execution.
Can I use this skill to write a product requirement document (PRD)?
No, you should use the pm-spec-writing skill for writing specs or PRDs; this skill is specifically for creating the handoff for an agent to implement them.