develop
Skill1 starsChanged 23 days agoMITView on GitHub
Use when asked to write, fix, debug, refactor, test or scope code in Python, C++ or Go (computer vision or not), or to find where code reads, sets or uses something ahead of a change. Load it before the first Read, Grep or edit, even for a small bug or a read-only lookup. Also for writing the /goal or /loop line for unattended work, and for setting a project up with cv-kit's rules, lint configs and permissions. Applies test-first, systematic debugging, simplicity and evidence, and sends the work to agents with a model and effort sized to the task.
29 other projects use a version of this.
Adopted by
Most recently taken up by
- chippingway/chipping-orchestratortheir own version, first seen Nov 11, 2026
- killerseven7777777/matt-skills-relaytheir own version, first seen Nov 8, 2026
- lalulali/vespyrtheir own version, first seen Nov 6, 2026
- Blakeem/aipromptguide-workflowstheir own version, first seen Nov 4, 2026
- javiarmesto/ALDC-AL-Development-Collectiontheir own version, first seen Nov 2, 2026
- ok-oldking/ok-scripttheir own version, first seen Nov 1, 2026
Used with
Projects with it also use scope, sync and 9 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.
YuriNachosSkillscope
Configure claude-scope status line widgets for Claude Code. Use when user wants to customize widgets, change theme, reorder elements, or modify display style.
Also in 6 projects that use this
MicrosoftSkillsync
Sync the current session branch with its upstream branch, or publish the current session branch to a remote. Use when the user asks to sync a branch, pull latest changes, rebase onto upstream, push current branch, publish branch, or set upstream.
Also in 6 projects that use this
nicolasestremSkilldocument
Create focused documentation for specific components or features
Also in 5 projects that use this
kt057Skillcheck
lint, typecheck, test を一括実行して結果をサマリーする
Also in 5 projects that use this
Harut8Skillarchitect
Also in 9 projects that use this
AnthropicSkilldebug
Structured debugging session — reproduce, isolate, diagnose, and fix. Trigger with an error message or stack trace, "this works in staging but not prod", "something broke after the deploy", or when behavior diverges from expected and the cause isn't obvious.
Also in 9 projects that use this
Other versions29 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 29
- jsmastery-pro/skills★ 1.4ksame name
- perber/leafwiki★ 1.2ksame name
- ok-oldking/ok-script★ 831same name
- quantumaikr/quant.cpp★ 403same name
- javiarmesto/ALDC-AL-Development-Collection★ 115same name
- chippingway/chipping-orchestrator★ 18same name
- alexiolan/craft-skills★ 17same name
- axiomantic/spellbook★ 13same name
- lalulali/vespyr★ 11same name
- flurdy/agent-skills★ 9same name
19 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.
The file
We split this skill into the rules it gives an agent
8 rules. 0 are common, 3 are distinctive, and 5 are very specific to this project.
0
Common
Most of the 24 edited versions we read keep these.
3
Distinctive
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 12
Test first: No production code without a failing test you watched fail.
- Distinctive: line 17
If the superpowers plugin is installed, its brainstorming, debugging, TDD and verification skills cover the generic process.
- Distinctive: line 24
Treat a fact above as true only for what it says.
- Very specific: line 33
Commit: only when asked, as ${CLAUDE_SKILL_DIR}/../../rules/commit-vcs.md says: Conventional Commits, no trailers.
- Very specific: line 55
The brief: Every engineer call carries these fields.
- Very specific: line 76
Never resend the same brief, and run at most three engineer attempts per package.
- Very specific: line 82
Writing the line: When asked to write a /goal or /loop line, change no code: write one fenced block for the user to paste, then stop.
- Very specific: line 88
Loops: Start with first run test -f .cv-kit-stop and stop if it exists, so touch .cv-kit-stop halts it.
“Common” is measured: kept by at least 6 in 10 of the 24 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
Develop
Four principles, on every change:
- Test first. No production code without a failing test you watched fail. A bug gets a test that reproduces it, a feature gets tests for its use cases, a refactor pins its behaviour first.
- Process over guessing. Reproduce, isolate, form one hypothesis, change one thing. No fix without a root cause; after two failed fixes, re-read the evidence.
- Simplicity. The smallest change that passes. No abstraction with one user, no unrequested option, no new dependency. Delete before adding.
- Evidence over claims. "Done", "fixed" and "faster" need command output from this turn; a hot path needs a measurement. Print the literal output; say what did not run.
If the superpowers plugin is installed, its brainstorming, debugging, TDD and verification skills cover the generic process; this skill supplies the rules, tests and routing. A project's written rules (CLAUDE.md, .claude/rules/, lint and build configs) win over this skill.
Facts read just now
!python3 ${CLAUDE_PLUGIN_ROOT}/scripts/project_facts.py facts .
!python3 ${CLAUDE_PLUGIN_ROOT}/hooks/evidence_gate.py arm ${CLAUDE_SESSION_ID}
Treat a fact above as true only for what it says; "not found" means look yourself, do not assume.
Flow
- Size it. A clear one-line change, where the user names the line and the change: do it. A bug whose cause is unknown is never a one-line change, however small the files. New or vague: read
${CLAUDE_SKILL_DIR}/references/scoping.md, ask one question at a time, write the brief back, and wait for a yes before code. - Look first. Read the project's rules and test setup, then run the existing tests for a baseline. With no
.claude/rules/, read the plugin's own in${CLAUDE_SKILL_DIR}/../../rules/:principles.md,evidence.md, and the style and test rules of each language you touch. A language with no linter config: offer cv-kit's, as${CLAUDE_SKILL_DIR}/references/project-setup.mdsays, and add it only after a yes. - Design, only when packages depend on each other. Two or more packages that share an interface or a data flow:
architectdesigns it and you wait for the user's yes, as "Route the work" says. One package, or packages that share nothing: skip this, the brief is the design. - Red, green, refactor: by you for a one-line change, otherwise by the
engineeryou brief. Read the test reference that applies before the first test:${CLAUDE_SKILL_DIR}/references/python-testing.md(pytest),cpp-testing.md(GoogleTest) orgo-testing.md(go test). A refactor or clean-up, of pasted code too: read${CLAUDE_SKILL_DIR}/references/refactor.mdbefore rewriting anything./simplifyis an optional last pass once the tests are green. - Verify. Run the tests and linters. Then review by the risk rules in "Route the work"; a one-line change needs no reviewer, because its test output is the check. The reviewer has no shell: paste the diff and the tool output, and the design's
InterfacesandInvariantswhen there was one, not your conclusions. End with the report: for a bug, the cause in one sentence withpath:line; the tests added, by name, and what each pins; the change; and the literal output of each command, or that it did not run and why. - Commit only when asked, as
${CLAUDE_SKILL_DIR}/../../rules/commit-vcs.mdsays: Conventional Commits, no trailers.
Need sourced evidence about a model, paper or benchmark: /cv-kit:research. Asked to set the project up: follow ${CLAUDE_SKILL_DIR}/references/project-setup.md; show its plan and stop, because the request is not the yes.
Route the work
Route yourself; do not ask the user. Anything beyond a clear one-line change goes to an agent. You scope it, gather the facts, write the brief and check what comes back; the agent reads, debugs and writes. A bug report that names a file and a symptom is already a brief: write it and dispatch engineer before you open that file. Do not read the code under change or search the repository yourself: explorer answers where, with path:line for the brief, and engineer finds why. Small files are not a reason to do the work yourself: once you have read enough to find the cause, you have done the engineer's job on the most expensive model, with no failing test behind it. Pass model and effort on every Agent call. Without the Agent tool, do the work yourself with the same steps.
Sonnet is the default; a precise brief is what makes it enough. A concurrency or memory-safety bug (deadlock, race, crash only under load, use-after-free) is Hard from the request alone: decide that before reading code, and brief the opus engineer with the files and symptoms the user gave. Everything else is sized by what is still unknown once explorer has answered, not by topic: box maths, NMS, a located bug or a perf fix with a benchmark is a Sonnet task. Opus is for work where the reasoning, not the typing, is the hard part.
| Task | Agent | model | effort |
| Find, read or summarise code, logs, docs | explorer | haiku | low |
| Mechanical: rename, move, a config value, a test for behaviour already pinned | engineer | sonnet | medium |
| Feature, bug fix, test or refactor step with a known spec or a bug a test can reproduce | engineer | sonnet | high |
| Design before code: two or more packages that share an interface or a data flow | architect | opus | high |
Hard: a concurrency or memory-safety bug, or a cause still unknown after explorer and a reproduction attempt | engineer | opus | high |
| Review of a change whose tests and linters pass | reviewer | sonnet | high |
Review touching concurrency, memory safety, security or a number that will be published, or of a change that implements an architect design or that an opus engineer wrote | reviewer | opus | high |
Design before code. When the work splits into two or more packages that share an interface or a data flow, send architect the approved brief, explorer's path:line facts, the test command and the baseline result before any engineer. It is read-only and returns the options, the flow, the interfaces, the invariants and the packages as engineer briefs. Show the design and wait for the user's yes; then add Facts, Ruled out and Mode to each package and dispatch it. Designed packages are sonnet, high, unless one is itself a Hard row. Use effort xhigh for a public API or ABI, a threading model, or a design the user sent back. Never send architect a single package, or packages that share nothing: the brief is the design.
The brief. Every engineer call carries these fields. A field you cannot fill is a question for explorer or the user, never for the engineer to guess.
Goal: one sentence, the behaviour after the change.
Done when: the test command, the new test names that must fail first, and "full suite green".
Facts: explorer's path:line findings, the repro or error text, the baseline test result.
Files: the files to change and the test file; anything else is out of bounds.
Constraints: project rules that apply; what must not change (API, dtype, BGR order, output format).
Ruled out: what was tried or rejected, and why. "none" on a first attempt.
Mode: fix | diagnose (reproduce with a test and find the cause; change no production code).One package is one behaviour and its tests, a few files at most; split anything bigger, and order the parts that depend on each other. Examples of good and bad briefs: ${CLAUDE_SKILL_DIR}/references/routing.md.
What comes back. Rerun the Done when command yourself and run git diff --stat.
- Done, but a file outside
Fileschanged or the red run is missing: ask the same agent with SendMessage, not a new one. - Blocked on a fact, a file outside
Filesor a decision: answer it and continue the same agent with SendMessage (without it, a new call with the earlier report pasted in). A decision that changes the scope goes to the user. - Failed because of the brief (a wrong file, a missing fact, a bad test command, the environment): fix the brief and rerun on the same model.
- Failed on the reasoning (the fix did not hold, the evidence contradicts the hypothesis): the
opusengineer row, with the failed attempt underRuled out. Anopushigh failure gets oneopusxhighrun; then stop and report the evidence.
Never resend the same brief, and run at most three engineer attempts per package. Run file-disjoint packages in one message, at most three engineers at a time, each with isolation: worktree.
Unattended runs (/goal, /loop)
When the prompt came from /goal <condition> or a /loop, nobody is there to answer. Take the brief from the condition; if it leaves out something that changes the result, state your assumption and go on. /goal's judge reads only the transcript, so end every turn with the literal output of the check (test summary, exit status, linter result), not a description of it. Never commit or push, and never run code from a cloned repo, unless the condition says so. After two failed fixes, stop and report what the evidence shows.
Writing the line. When asked to write a /goal or /loop line, change no code: write one fenced block for the user to paste, then stop. Use the facts above for the check; "not found" means look yourself.
- Pick the mode. A finite task with a checkable end (fix, feature, migration, backlog):
/goal. Something to repeat or watch on an interval:/loop. Say which, and why, in one line. - Set the end state. One measurable outcome and the command that proves it. A vague task gets your assumption stated inside the line.
- Bound it. End with
Or stop after N turns(fix 15, feature 25, research 30), and addAfter two failed fixes, stop and report the literal output. - Constrain it. Name what must not change: other test files,
no commit,no push, no code run from a clone. - Loops. Start with
first run test -f .cv-kit-stop and stop if it exists, sotouch .cv-kit-stophalts it. A watch-only loop sayschange nothing. - Headless, only if the user will be away. Give
claude -p "/goal ..." --output-format stream-json --verbosewith a log path outside the repo. Permissions must be set in advance (${CLAUDE_SKILL_DIR}/references/settings.md). - Size. At most 4,000 characters. Keep tests in the foreground: a background run defers the judge.
Output: the block, then Paste it after /goal or /loop, in auto mode. Examples: ${CLAUDE_SKILL_DIR}/references/goal-loop-templates.md.
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.

