human-handoff
Skill0 starsChanged 14 days agoMITView on GitHub
Read before presenting commands or steps for the user to execute, including commands, launch instructions, and requests relayed to another agent. Also use for requests for user input, physical actions, or approval. Make the remaining human action reliable to carry out.
1 other project uses a version of this.
Adopted by
Most recently taken up by
- hams-ollo/zen-agent-skillstheir own version, first seen Oct 20, 2026
Other versions1 project uses a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 1
- hams-ollo/zen-agent-skills★ 2same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository6 other files it gives its agents.
- coordinate-repository-versionsSkill
- domain-modelingSkill
- grillingSkill
- grill-with-docsSkill
- optimize-agent-docsSkill
- optimize-human-docsSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
13 rules. About 0 are common, 8 are distinctive, and 5 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈8
Distinctive · estimate
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 14
Check a plausible CLI or API when it could remove a manual step.
- Distinctive: line 18
Ask only for missing input that materially affects the result, required approval, or an action the available tools cannot perform.
- Very specific: line 48
Use a plain-text code block without quote prefixes or list decorations that the person must remove.
- Very specific: line 50
Keep shared-task content and state under the handoff rule when available.
- Distinctive: line 58
When a command is long or the user reports wrapping, split it at syntactically valid boundaries.
- Very specific: line 61
Keep each option, path, and quoted value intact.
- Distinctive: line 67
When the remaining procedure amounts to a script (for example, checks sharing variables, branching, or coordinated failure handling), reuse …
- Distinctive: line 72
Create a script only when needed and allowed.
- Distinctive: line 74
Put the file where the intended user can access it within existing permissions.
- Distinctive: line 75
If transfer is needed, provide a way to retrieve just that file, using a fixed revision when retrieving from a repository, followed by a …
- Distinctive: line 82
Before adding a script-like handoff, check whether the existing ordinary entry point can perform the work.
- Very specific: line 102
If a result is needed, ask for the specific non-sensitive output or pass/fail.
- Very specific: line 103
Never request tokens, passwords, or private data in chat.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Handing work to a human
Apply this skill even when the user asks only for a command and no "please run" or handoff heading is needed. A request for instructions is not authorization to perform the requested action instead of showing it.
Finish the authorized machine work before handing over the part that requires the person. Use available tools to resolve facts and routine choices yourself. Check a plausible CLI or API when it could remove a manual step; do not broaden the task into installing tools, changing permissions, or handling credentials just to avoid asking. Respect sandbox and authorization boundaries.
Ask only for missing input that materially affects the result, required approval, or an action the available tools cannot perform. Existing authorization remains valid; do not ask for it again. Continue independent work while waiting.
Give the smallest remaining action, its execution context, and what success looks like. Start from the person's current host and screen, using available evidence and existing entry-point documentation. Do not assume a target input field or terminal is already open: include the steps to open the application, connect to the target environment, and reach the input, as needed, in execution order with the visible success state for each step. Reuse an already open, connected session instead of restarting it. If the starting state is unknown, make those setup steps conditional; ask only when the uncertainty prevents a usable handoff. Distinguish checking the written procedure from observing the actual screen or successfully completing the action.
Relaying a request to another environment's agent is a human action when the
person must carry it. Name the receiving environment and provide the exact text
to paste, separately from the instructions to the person. Include the receiving
environment in the payload itself, so it survives forwarding without the
surrounding explanation. Include necessary URLs
as literal text inside the copyable payload, not only hyperlink attributes.
For an environment-routed request, the payload must also name the concrete
environment and logical runtime principal, its existing entry point, requested
approval/sandbox mode, acceptance risk, material exclusions or forbidden
operations, capability evidence, and any unknown condition. Do not infer those
facts from a shared distro, OS label, or an environment's active state, and do
not describe a contract-changing environment modification as preparation. Keep
the requested approval/sandbox mode distinct from the recipient's observed
effective mode. Risk accepted inside the named environment does not extend to
its host or another environment.
Use a plain-text code block without quote prefixes or list decorations that the
person must remove. Do not ask them to reconstruct references or edit the request.
Keep shared-task content and state under the handoff rule when available; this
skill governs the human relay, not an additional task ledger.
Commands, relayed requests and their references must survive the actual display
and copy-paste path, not merely parse as the original text. Console or harness wrapping can insert real
whitespace, even inside a quoted argument: workspace-write can become
workspace- write. A fenced one-line command alone does not prevent this.
When a command is long or the user reports wrapping, split it at syntactically valid boundaries. Prefer existing short entrypoints, entering the target shell and directory first, or native argument arrays with separate short elements. Keep each option, path, and quoted value intact. Include the full WSL/SSH/container prefix when assessing length. Do not rely on backslash/backtick continuations or ask the user to repair wrapped text. Check the longest displayed line against the known display constraint; if the width is unknown, shorten the invocation without inventing a universal width guarantee.
When the remaining procedure amounts to a script (for example, checks sharing variables, branching, or coordinated failure handling), reuse an existing entry point or save and validate a script file before handing it over. Do not present the script body as a copy-paste procedure, including a heredoc or encoded payload that makes the user recreate the file. Keep ordinary short commands as commands. Create a script only when needed and allowed; respect a no-new-scripts request.
Put the file where the intended user can access it within existing permissions. If transfer is needed, provide a way to retrieve just that file, using a fixed revision when retrieving from a repository, followed by a separate short execution command. Do not require a repository clone just to obtain the script or pipe a download directly into a shell. Verify the saved file, the available retrieval path, and the target shell; state the working directory, success output and failure behavior. Distinguish a validated handoff from execution by its recipient.
Before adding a script-like handoff, check whether the existing ordinary entry point can perform the work. Do not retain an unnecessary workaround as code. Record an unresolved cause only when it threatens a required outcome, data protection, or recovery with substantial human monitoring, approval waits or repeated work. Reuse the designated task or established issue tracker; do not create an issue or handoff update solely because a manual step may recur or an easily recoverable problem occurred. Distinguish an observed failure from a suspected cause, and do not expand the authorized repair scope. Keep any required but unavailable recording pending in the existing handoff, without copying the task's contract or results into a second ledger.
Validate the exact presented syntax in the stated shell and use a safe dry run or version query through the same argument path where available. Syntax checks do not prove that copying from the UI preserves bytes. For relayed text, check that plain-text copying retains the request and literal URLs without manual removal of decoration; a code block's appearance alone is not proof. Report format inspection, observed copy/paste, recipient retrieval and execution as separate evidence. If the actual UI path is unavailable, leave that check pending. Never execute a destructive or externally visible action merely to test the user's command.
If a result is needed, ask for the specific non-sensitive output or pass/fail.
Never request tokens, passwords, or private data in chat. Follow the applicable
session-end-user-work rule for final-response formatting.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

