work
Skill2 starsChanged 2 days agoMITView on GitHub
Build a feature end to end with an agent pipeline, in any repository, and show it live on a run board. Starts from a GitHub issue number, an issue URL, or a plain-text request. It writes a run-specific workflow from its templates and runs it - research, a plan, an approval agent, the owner's approval, tests first, implementation, the gate and E2E, verification, per-role agent QA recorded on video, UI before/after screenshots when the UI changes, an evidence report, and the owner's ship decision. Every stage shows on the run board (dashboard.html; an Artifact in cloud sessions) and in a stage table. Use whenever the owner says "/work", "work on", "use the work skill", or "implement issue N". Also use it to continue a run - "/work approve", "/work ship", "/work status" - or to set a repo up for it - "/work setup".
- Commits and pushes
47 other projects use a version of this.
Adopted by
Most recently taken up by
- arxdsilva/ai-loopstheir own version, first seen Nov 16, 2026
- BoredHF/albertskillstheir own version, first seen Nov 16, 2026
- Weaverse/.agentstheir own version, first seen Nov 15, 2026
- yagizdo/quivertheir own version, first seen Nov 15, 2026
- automagik-si/genietheir own version, first seen Nov 15, 2026
- Firzus/agent-skillstheir own version, first seen Nov 8, 2026
Used with
Projects with it also use authoring, done and 10 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
automagik-devSkillauthoring
Write or revise a Genie skill so it survives the shipped contract — frontmatter, house size, starter card, and runtime-neutral voice.
Also in 3 projects that use this
robzolkosSkilldone
Reset the working directory for the next task by ensuring no uncommitted changes exist, then switching to the main branch and pulling latest.
Also in 3 projects that use this
Google CloudSkillfix
Plan a remediation strategy after a test failure.
Also in 4 projects that use this
AlexPEClubSkillrefine
Always use when the user wants to discuss an exsiting feature or specification. Open an existing feature spec to improve, extend, or fundamentally challenge it. Pass the feature ID as argument (e.g. /refine PROJ-2).
Also in 3 projects that use this
MicrosoftSkillmerge
Merge changes from the topic branch to the merge base branch. Use when the user wants to merge their session's work back to the base branch.
Also in 3 projects that use this
Microsoft AzureSkilltest
Also in 6 projects that use this
Other versions47 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 47
- cwinvestments/memstack★ 433same name
- automagik-dev/genie★ 353same name
- automagik-si/genie★ 353same name
- namastexlabs/automagik-genie★ 351same name
- ohad6k/emulo★ 292same name
- first-fluke/fullstack-starter★ 226same name
- guardana/guardana★ 225same name
- OutThisLife/brooklyn-skills★ 215same name
- Towow-ai/Flowness★ 107same name
- assaio/assaio★ 63same name
37 more not listed
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 repository9 other files it gives its agents.
- control-roomSkill
- dsa-conscious-codingSkill
- excalidraw-architectureSkill
- improve-codebase-architectureSkill
- issue-to-phasesSkill
- pr-screenshotsSkill
- raise-prSkill
- visual-reviewerSkill
- wayfinderSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
22 rules. 0 are common, 6 are distinctive, and 16 are very specific to this project.
0
Common
Most of the 43 edited versions we read keep these.
6
Distinctive
This author's own choices. Worth a look.
16
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Very specific: line 10
make sure the repo has a profile (.claude/work.json), and write one with the owner if it does not.
- Distinctive: line 11
start the run board, so the owner sees every stage live.
- Very specific: line 12
write the workflow files for this run from the templates in workflows/, and run them with the Workflow tool.
- Distinctive: line 13
stop at the two owner gates.
- Very specific: line 33
Do not post to GitHub issues or add CI workflows unless the owner asks.
- Very specific: line 39
Check .agent/. If it belongs to another unfinished run (state.json with "finished": false and another run), stop and ask.
- Very specific: line 40
Profile: Read .claude/work.json. If it is missing, run the first-run setup in references/setup.md now and get the owner's OK before you go …
- Very specific: line 45
Add .agent/ to .git/info/exclude (create the line once) so the run state can never be staged.
- Very specific: line 46
Branch: Create <branch_prefix><n>-<slug>, or <branch_prefix>local-<slug> without an issue, from the profile's base (default main) in the …
- Very specific: line 54
Board: Start it now, before any agent runs (references/board.md):
- Very specific: line 57
App: When the profile has app: run app.check.
- Distinctive: line 74
If approval is still REVISE after 2 rounds, report the open issues to the owner.
- Distinctive: line 111
Ask "Ship" / "One more fix round, then ship" / "Stop".
- Very specific: line 114
Run build.js once more with mode: "fixup" and fixup: {themes}.
- Distinctive: line 116
If everything is green, ship with no second ask.
- Distinctive: line 123
Run the ship checks. They must all pass:
- Very specific: line 124
No committed file points into the run state or a home folder: git diff --name-only origin/<base>...HEAD | xargs grep -lE …
- Very specific: line 127
If ui: .agent/ui/review.json exists and is newer than .agent/ui/shots.json.
- Very specific: line 133
Run state.py step ship --status DONE and state.py finish.
- Very specific: line 138
The board and the table stay current: Every stage you start or finish goes through state.py.
- Very specific: line 142
Secrets: Never print or commit anything in the profile's secrets.
- Very specific: line 143
Project rules: Every agent follows the files in the profile's rules.
“Common” is measured: kept by at least 6 in 10 of the 43 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
/work: the feature pipeline
You are the orchestrator. You do not write feature code yourself. You:
- take in the request;
- make sure the repo has a profile (
.claude/work.json), and write one with the owner if it does not; - start the run board, so the owner sees every stage live;
- write the workflow files for this run from the templates in
workflows/, and run them with the Workflow tool; - stop at the two owner gates;
- ship when the owner says so.
S = this skill's base directory (the Skill tool prints it when the skill loads; with the plugin installed it is inside the plugin cache, not the repo). Use the absolute path in every command. Run state lives in <repo>/.agent/; the layout is in references/state.md. The rules behind each stage are in references/pipeline.md. Read it once per session.
Before you write or edit any workflow script, load the workflow-authoring skill (Skill tool) if it is available. That includes the copies of the templates when the plan needs a stage they lack. It holds the script API and its traps: meta must be a pure literal, Date.now() and Math.random() throw, scripts are plain JavaScript, args must be passed as a JSON value. When it is not available, change only the args, never the scripts.
0. Read the invocation
| Invocation | Means |
/work 123, /work #123, /work <issue URL> | New run from GitHub issue 123 |
/work <free text> | New run from that text; no issue |
/work alone | Ask what to build (one question), then start |
/work setup | Write or update .claude/work.json only (references/setup.md) |
/work approve [changes…] | Owner gate H1 passed → build. With text: one re-plan applying it, then build |
/work ship · /work fix then ship: <themes> | Owner gate H2 → ship (§6), or one more fix round, then ship |
/work status | Print the stage table (state.py show), the board link, and what waits on whom |
/work board | Start the board again (or republish the snapshot) and give the link |
Do not post to GitHub issues or add CI workflows unless the owner asks. Reading an issue with gh issue view is fine.
1. Intake (inline, no agents)
- Check the base.
git statusmust show no tracked changes, or only this run's changes.- Check
.agent/. If it belongs to another unfinished run (state.jsonwith"finished": falseand anotherrun), stop and ask. Never delete another run's state.
- Profile. Read
.claude/work.json. If it is missing, run the first-run setup inreferences/setup.mdnow and get the owner's OK before you go on. Never guess the gate, the app URL or the roles: a run without a profile is the run that gets lost. Write the request into
.agent/issue.md.- From an issue:
gh issue view <n> --json number,title,body,labels,comments,author. Quote the title, body and comments verbatim under headings. - From free text: write the text verbatim, plus what you inferred (roles, pages, risk) marked as inferred.
Add
.agent/to.git/info/exclude(create the line once) so the run state can never be staged.- From an issue:
- Branch. Create
<branch_prefix><n>-<slug>, or<branch_prefix>local-<slug>without an issue, from the profile'sbase(defaultmain) in the current checkout. Do not use a worktree when the local app serves this checkout. Manifest. Run:
python3 $S/scripts/state.py init --run <branch> --name "<title>" [--issue <n>] [--session <this session id, if known>] [--no-app]It writes
.agent/state.jsonwith all 18 stages as TODO (--no-appmarks the app stages SKIP when the profile has noapp). MarkintakeDONE.- Board. Start it now, before any agent runs (
references/board.md):- Local session:
python3 $S/scripts/board/wfboard.py --repo .withrun_in_background: true. Read its first line and give the owner the URL. - Cloud session (
CLAUDE_CODE_REMOTEis set, or the owner cannot open localhost): write the snapshot (--snapshot .agent/board.html) and publish it with the Artifact tool; give the owner the link.
- Local session:
- App. When the profile has
app: runapp.check. If the result is notapp.expect, runapp.start(tell the owner it takes aboutapp.start_minutes), then check again. Then runapp.loginso each role's saved login exists. Record the result as thestackstep.
2. Plan workflow → H1
- Load
workflow-authoring(see above). Copy$S/workflows/plan.jsto.agent/workflow/plan.js. Write
.agent/workflow/plan.args.json:{"run": "<branch>", "issue": ".agent/issue.md", "title": "...", "today": "YYYY-MM-DD", "skill": "<absolute $S>", "profile": <the parsed .claude/work.json>}Pass the date from your context; scripts cannot read the clock.
- Mark
researchRUN, then callWorkflow({scriptPath: ".agent/workflow/plan.js", args: <that object>}). Passargsas a JSON value, not a string. The workflow shows on the board on its own while it runs.
The workflow researches in parallel, has the planner write .agent/plan.md, tickets.json and test-plan.json (plus ui/pages.json when the UI changes), and runs the approval agent with up to 2 revisions. It returns {scope, plan, approval}.
- Set
research,planandapprovefrom the result, with one-line summaries and the files as--artifact(they then open on the board). - If approval is still REVISE after 2 rounds, report the open issues to the owner. Do not hide them.
- H1. Set
h1to WAIT. Refresh the board (republish the snapshot in a cloud session). In chat, print the stage table (state.py show) and the plan summary: tickets, test count, risk tier,ui, open questions with their defaults, and the approval agent's verdict. Ask with AskUserQuestion: "Approve" / "Approve with changes" / "Stop". Seth1DONE when approved. If the owner leaves, the run resumes later with/work approve.
3. Build workflow (after H1)
- Read the run's inputs:
tickets.json,test-plan.json, and the plan'srisk,ui,rolesandspecs. - Prepare the media layers (
references/media.md): whenuiis true, put the ui-snap tool in place and build theuiSnapcommand. Checkcommand -v agent-browser ffmpeg; when either is missing, or the profile says"video": false, video is off and the evidence says why. Write the files. Load
workflow-authoring. Copy$S/workflows/build.jsto.agent/workflow/build.js, and writebuild.args.jsonwith:run,today,skill(absolute$S),profile(parsed);tickets(verbatim from tickets.json),roles,risk,ui;e2e: the profile'se2e.cmdwith{specs}replaced by the plan's specs;uiSnap(ui runs),video;budgets: {build: 2, qa: 1}.
The args make the generic template fit this run. Edit the copied script only when the plan needs a stage the template lacks, and note the edit in the evidence.
- Run it. Mark
testsRUN, thenWorkflow({scriptPath: ".agent/workflow/build.js", args}). When it returns, set each stage from the result (tests,red,ui-before,implement,gate,verify,qa,ui-after,qa-fix,evidence), with artifacts (.agent/qa/<role>.json,.agent/evidence.md). Read the.agent/files the stages wrote; never trust an agent's summary over the file it was meant to write. - When the Workflow tool is missing (a headless run), run the same stages with the Agent tool, in the order and with the prompts in
references/stages.md, and record each stage as it finishes so the board stays current.
4. Media layers
references/media.md covers both.
- UI before/after, only when
uiis true: before shots on the base code after the tests are written, after shots on the final code, then the UI review agent writes.agent/ui/review.jsonwith expected / unexpected / broken per changed shot. - Video, every run with an app: each QA agent records
demo-<role>.webm(the happy path, for the owner) andqa-<role>.webm(its exploration), each with a contact sheet.
The board shows both under "◐ UI before / after" and "▶ Videos".
5. H2: evidence to the owner
Set h2 to WAIT and refresh the board. Show the owner, in chat:
- the stage table;
- the pass table per role × browser, with NOT RUN where a browser did not run;
- gate results and the verify verdict;
- QA counts and open findings;
- UI shots that changed, with their verdicts, and the videos (on the board);
- time per stage, and cost, labelled as estimated unless billed per token;
- the board link.
Ask "Ship" / "One more fix round, then ship" / "Stop".
"Fix then ship" works as follows:
- Run build.js once more with
mode: "fixup"andfixup: {themes}. The owner has chosen to go over the budget. - It re-runs the gate (and the UI after capture and review if
ui). Then run the full suite (e2e.full). - If everything is green, ship with no second ask. Otherwise stop and report again.
6. Ship
Ship only on the owner's /work ship, or a green "fix then ship".
- Rebase the branch onto
origin/<base>, then re-run the gate. A conflict or a red gate sends the run back to a fix round. - Run the ship checks. They must all pass:
- No committed file points into the run state or a home folder:
git diff --name-only origin/<base>...HEAD | xargs grep -lE '\.agent/|/Users/|/home/[a-z]'is empty (files of this skill itself excepted). - Nothing from the profile's
secrets,.agent/, the ui-snap copy,test-results/ornode_modules/is in the diff. - If
ui:.agent/ui/review.jsonexists and is newer than.agent/ui/shots.json.
- No committed file points into the run state or a home folder:
Land it by the profile's
ship.mode:pr(default): commit on the branch with a feature-level message, then ask once beforegit push -u origin <branch>andgh pr create(title from the request; body from the evidence summary, without.agent/paths).squash:git checkout <base> && git merge --squash <branch>, commit with a feature-level message, then ask once beforegit push origin <base>.
End the commit message with the attribution lines the session gives you.
- Run
state.py step ship --status DONEandstate.py finish. Refresh the board one last time. If the run came from an issue, offer to comment the link on it. Do not comment unasked.
Rules that always apply
- Owner gates are never skipped. Nothing is pushed without the owner's yes.
- The board and the table stay current. Every stage you start or finish goes through
state.py; in a cloud session, republish the snapshot at H1, after QA and at H2. - Budgets: build/verify fix rounds stop at 2; QA fix rounds stop at 1. When a budget runs out, stop and ask. Never loop silently.
- Browsers. A browser that did not run is reported as NOT RUN, never dropped. A test that fails once and then passes is a flake: repeat it 20 times and name it in the evidence.
- QA scope. QA agents change only their own role's state. They seed fixtures before they start and remove them afterwards.
- Secrets. Never print or commit anything in the profile's
secrets. Agents read passwords from those files and never echo them. Videos and screenshots show only seeded test accounts. - Project rules. Every agent follows the files in the profile's
rules. Review findings citereferences/review-rules.mdor the project'sreview_rules. - Cost. Record cost per stage where the workflow result has it. Otherwise mark it "estimated".
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.

