work
Skill2 starsChanged yesterdayMITView on GitHub
Token-lean executor that gets projects finished. Turns a /mind-map plan into task cards with a small v1 finish line, splits work 70% AI / 30% the user, coaches the user on their own cards, reviews diffs to teach, parks "this idea is dumb" doubts until v1 runs, and tracks progress and cost per card. Use when the user types /work, says anything like "let's work on this project", "start working", "carry on", "what's next", "I'm done with this bit", "I'm stuck" or "check my code" in a project with a mindmaps/ or .work/ folder, asks to execute their mind map, or says mid-project that the idea is dumb or not worth finishing.
46 other projects use a version of this.
Adopted by
Most recently taken up by
- 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
- isanchez-aguilar/relayrace-aitheir own version, first seen Nov 8, 2026
- Mharbulous/probanda-devtheir 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 versions46 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 46
- 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
36 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 repository17 other files it gives its agents.
- autopilotSkill
- autopsySkill
- confidence-mapSkill
- delegateSkill
- expert-askSkill
- failure-museumSkill
- ghost-usersSkill
- grill-dumpSkill
- intent-diffSkill
- md-to-mapSkill
- mind-mapSkill
- premortemSkill
- red-teamSkill
- shipitSkill
- stuck-detectorSkill
- taste-thiefSkill
- tournamentSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
13 rules. 0 are common, 2 are distinctive, and 11 are very specific to this project.
0
Common
Most of the 41 edited versions we read keep these.
2
Distinctive
This author's own choices. Worth a look.
11
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 13
If ~/.claude/agents/delegate.md is missing, copy it from ${CLAUDE_SKILL_DIR}/agents/delegate.md (create ~/.claude/agents/ if needed).
- Distinctive: line 16
Never write code on a drive card.
- Very specific: line 17
Never mark a card [x] without running its done: check.
- Very specific: line 18
Never read map-data.js, index.html or the export JSON.
- Very specific: line 19
Never start later or polish cards while v1 cards are open without a confirmation.
- Very specific: line 20
Keep state.md at 6 lines or fewer.
- Very specific: line 24
/workalone, or "let's work on this": run where.mjs first. It reads .work/ and finds the mind map for you, and prints one action: line. - Distinctive: line 36
Mid-session, work out the sub from what the user says: Don't ask them to type it:
- Very specific: line 47
All state lives in the project's .work/ folder.
- Very specific: line 86
Finish line: Write why: <one sentence on why the user wants this> and v1: <one sentence on what done looks like> at the top of cards.md.
- Very specific: line 97
drive: no code. Give the goal, the files, the done check, and one design question to think about first.
- Very specific: line 98
Write state.md, recording the start reading in its budget: line.
- Very specific: line 109
Run git diff (or git diff HEAD~1 if nothing is staged) limited to the card's files.
“Common” is measured: kept by at least 6 in 10 of the 41 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
Execute a plan cheaply, keep the user coding and learning, and reach a finished v1.
Requires the mind-map skill installed next to this one (the scripts here stop with a one-line message if it is missing) and Node 18+. Not for one-off tasks with no plan: do them directly, or use the delegate skill.
First run
If ~/.claude/agents/delegate.md is missing, copy it from ${CLAUDE_SKILL_DIR}/agents/delegate.md (create ~/.claude/agents/ if needed). Without it, spawn a general-purpose agent with the same card as the brief.
Hard rules
- Never write code on a drive card.
- Never mark a card
[x]without running itsdone:check. - Never read
map-data.js,index.htmlor the export JSON; use the scripts. - Never start
laterorpolishcards while v1 cards are open without a confirmation. - Keep
state.mdat 6 lines or fewer.
Autopilot: the user should never need a sub-command
/work alone, or "let's work on this": run where.mjs first. It reads .work/ and finds the mind map for you, and prints one action: line. Do exactly that action, without asking which command to run:
| action | Do |
no-map | Say no mind map was found, and offer to start one with /mind-map |
export | The map has no export yet. Run the run: command, then run where.mjs again |
plan | Run plan on the given map. If there's a note: about other maps, mention it in one line |
continue | Resume the card from state.md's next: line. Drive or pair cards: ask how it's going |
next | Run next |
ship | Run ship |
done / shelved | Ask the question in the line |
Mid-session, work out the sub from what the user says. Don't ask them to type it:
- "done", "finished it", "it works" →
review, thenwrap, then offer the next card - "stuck", "no idea", "help" →
hint - "check this", "look at my code" →
review - "this is dumb", "why bother", "someone already made this" →
doubt - "I'm out", "stopping here", "gotta go" →
wrap(no question), sostate.mdis saved - "I'm done with this project", "give up" → offer
shelve, with a single 5-minute doubt check first - "how far am I" →
status
The subs below still work if typed, as shortcuts.
All state lives in the project's .work/ folder. Scripts are in this skill's scripts/ folder; run them with node ${CLAUDE_SKILL_DIR}/scripts/<name>.mjs.
| File | Holds | Limit |
.work/cards.md | why: and v1: lines, then one line per card | One line each |
.work/state.md | Handoff between sessions | 6 lines |
.work/doubts.md | Parked doubts and their answers | One line each |
.work/learn.md | Explain-back answers, ## Gaps, ## Shipped | Append only |
For doubt, ship, shelve, or when the user voices doubt mid-task, read ref/finish.md first. Don't read it otherwise.
Rule zero: budget
Optional: if your setup reports a budget tier, check it before each sub. AMBER: delegate cards only go to haiku or sonnet, no Opus subagents. RED: finish the current step, run wrap without the question, stop. OUT: write state.md and stop. With no tier, assume GREEN. Never start a card you can't finish in the budget you have left.
Card format
- [ ] T03 | pair | M | v1 | Parse config file | files: src/Config.kt, test/ConfigTest.kt | done: ConfigTest passes | map: loader/configMarks: [ ] todo, [~] doing, [x] done (wrap adds | hints:N | used:N%). Size: S under 30 min, M about 1 h, L about 2 h. Split anything bigger.
Cut:
- v1: the smallest thing that runs end to end and does the core idea. Aim for 5–8 cards, a weekend or two.
- later: real features that come after v1.
- polish: UX, looks, edge cases, refactors. Only after v1 ships.
Modes
- delegate (AI codes): boilerplate, wiring, config, build files, tests for logic that already exists. S cards go to
model: haiku, M and L go tosonnet. - pair: the AI writes a compiling skeleton and leaves
// YOU:markers where the core logic goes. The user fills those in. - drive (user codes): core domain logic, algorithms, concurrency, memory ownership, anything listed under Gaps in
learn.md. The AI writes no code on these cards.
Target: you 30%, AI 70%, weighted by size (pair counts half). ratio.mjs measures it.
Subs
plan [slug|path]
- Run
map2tasks.mjs <path from where.mjs>. It readsmindmaps/<slug>/export/<slug>.json. Never read that file,map-data.jsorindex.htmlyourself. Its# finish/…and# doubt/…lines are the map's Finish and Doubts branches. If it says the map is NOT agreed complete, say so once and ask whether to plan anyway or go back to /mind-map. Leaves withopen:questions get those questions as the card's first step. - Finish line. Write
why: <one sentence on why the user wants this>andv1: <one sentence on what done looks like>at the top ofcards.md. Take them from the# finish/lines. If they're missing, ask the user, in 2 questions at most. - Pre-mortem. Copy each
# doubt/line intodoubts.mdwith its answer (format inref/finish.md). If the map has none, ask: "What's the one thing in the back of your mind that could make you drop this?" Log the answer as a parked doubt. - Group the leaves into cards. Use a leaf's
cut:if it has one, otherwise propose it. Be ruthless: anything that isn't needed for the core loop to run islaterorpolish. Find file paths with Glob/Grep on names only. Don't open files. - Assign modes, write the cards, and run
ratio.mjs --plan. Re-balance until it reports on target. - Show
why,v1, and a compact table (id, mode, size, cut, title). Ask the user to flip any modes or cuts. Createstate.mdandlearn.md(with a## Gapsheading) if they're missing.
next [id]
Read only state.md and cards.md. Pick the user's id, otherwise the first [~] card, otherwise the first [ ] v1 card. Don't pick later or polish while v1 cards are open, unless the user insists. Then say once: "That's after v1, are you sure?" Mark the card [~]. Then:
- delegate: spawn the
delegateagent with the card line as the whole brief, at the model from Modes. Then run the card'sdone:check yourself andgit diff --stat, so the worker's "pass" isn't taken on trust. Don't re-read whole files; skim only the hunks if the check fails or the diff touches files outside the card. One retry through SendMessage with the failure output, then step up a model. A card bigger than L, or one that splits into independent parts, goes through thedelegateskill's waves. - pair: spawn
delegatewith "skeleton only" plus the card, athaikufor S andsonnetfor M/L. List the// YOU:markers asfile:line. - Small side jobs during any card (finding usages, reading a long log, writing a test for code that already exists) go through the
delegateskill's quick lane, not the main thread. - drive: no code. Give the goal, the files, the done check, and one design question to think about first. Stop.
Write
state.md, recording the start reading in itsbudget:line.
hint
For drive and pair cards. Each call climbs one level, and the level is stored in state.md as hint:N:
- A nudge: where to look, or what's off.
- The concept or pattern involved, named and explained in 2 to 3 lines.
- Pseudocode.
- Code, at most 15 lines. Add the topic to
## Gapsinlearn.md. Answer in 6 lines or fewer. Read only the file:line range the user is working on.
review
Run git diff (or git diff HEAD~1 if nothing is staged) limited to the card's files. Read nothing else unless the diff references it.
- Code the user wrote: teach. Give up to 5 findings, ranked:
file:line, what's wrong, and why it matters. For bugs, give the fix. For design or style, ask a question that leads them to it. Name one thing they did well. - Code the AI wrote: be terse and fix it directly.
- Checklist by language:
- Java: nulls and Optional, try-with-resources, equals/hashCode, swallowed exceptions, mutability, thread safety.
- Kotlin:
!!,lateinit, misused scope functions, coroutine scope and cancellation, mutable state in data classes. - C++: RAII and ownership, raw new/delete, undefined behaviour (bounds, uninitialised values, dangling references), copies vs moves, const-correctness, iterator invalidation.
- End with
verdict: shiporverdict: fix first.
wrap
- Mark the card
[x]and append| hints:N. If your setup reports a usage percentage, also append| used:N%: the reading now minus the one instate.md'sbudget:line. - For drive and pair cards (skip this in RED): ask ONE explain-back question about the user's own code, such as why it works or what breaks if X changes. Wait for the answer, judge it in one line, then append
date | card | Q | verdicttolearn.md. A weak answer adds the topic to## Gaps. - Rewrite
state.mdand runratio.mjs. If it says v1 is ready, point the user to/work ship. Suggest a fresh session, sincestate.mdcarries everything over.
doubt <text> · ship · shelve
See ref/finish.md.
status
Print the ratio.mjs output (v1 bar, split, cost per card, doubts), the why: line, state.md, and the budget tier. Read nothing else.
state.md (exactly this, 6 lines max)
tldr: <one sentence: where we are>
card: T03 pair M v1 | hint:0
next: <the very next action, with file:line if there is one>
blocked: -
done: T01-T02 · v1 2/7
budget: GREEN 82% · pc 18% @ 05:35 (tier and usage reading at card start; write "-" if unknown)Token rules
- Send any output over 80 lines (test runs, build logs, stack traces, unfamiliar files) to a haiku agent for a summary (the
tldragent if you have one). - Grep, then read line ranges. Never re-read a file after editing it.
- One card per session is ideal.
state.mdreplaces conversation history. - Replies stay short and end with a line
TL;DR: ….
Anti-patterns
- Asking the user which sub-command to run.
- Reading whole files to "get context" instead of grep and line ranges.
- Pep talks in reply to a doubt; classify, park, show the v1 bar.
- Padding v1 with polish cards.
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.

