5-csk-backend
Skill1 starsChanged 4 months agoMITView on GitHub
Implement the non-surface part of a feature using the architecture-approved stack with quality-first, performance-aware, security-reviewed, local/cloud-conscious core engineering. Despite the historical name, this is a general core-engineering skill for Python, C/C++, .NET, native apps, CLIs, services, libraries, pipelines, automations, data systems, AI/RAG/ML workloads, and large software projects.
In this repository18 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- 1-csk-initSkill
- 2-csk-write-specSkill
- 3-csk-architectureSkill
- 4-csk-frontendSkill
- 6-csk-qaSkill
- 7-csk-deploySkill
- audit-plan-loopSkill
- claude-skill-proxy-syncSkill
- create-worker-worktreeSkill
- csk-adopt-plan-scaffoldSkill
- csk-helpSkill
- csk-refineSkill
- csk-startSkill
- finish-branchSkill
- pull-main-ffSkill
- review-loopSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
37 rules. About 0 are common, 24 are distinctive, and 13 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈24
Distinctive · estimate
This author's own choices. Worth a look.
13
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 19
When this skill starts, first give the user one or two plain-language sentences in the user's language: which step this is, what happens …
- Distinctive: line 32
Use this skill for any architecture-approved non-surface work, including:
- Distinctive: line 48
Do not assume a web backend.
- Distinctive: line 49
Do not assume an HTTP API.
- Distinctive: line 50
Do not assume a database.
- Distinctive: line 51
Do not assume JavaScript, Node, Python, or any other language until architecture chooses it.
- Distinctive: line 52
Do not assume cloud deployment.
- Distinctive: line 53
Do not assume the project is small.
- Distinctive: line 54
Do not assume frontend/backend separation.
- Very specific: line 55
Do not add a runtime, package manager, framework, database, service wrapper, IPC protocol, queue, or deployment mechanism unless …
- Very specific: line 58
Before reading or writing project source-of-truth files, resolve adapter config according to .claude/rules/adapter-config.md.
- Distinctive: line 60
Default paths in this skill are fallbacks only:
- Very specific: line 67
If the adapter config provides alternate paths, use those paths exactly and do not also create the defaults.
- Very specific: line 71
Run /csk-start and require valid onboarding state.
- Distinctive: line 74
Read CLAUDE.md and confirm the CSK workflow is active, or confirm an adapter config exists.
- Very specific: line 78
Read .claude/rules/workflow-state.md and .claude/rules/loop-policy.md.
- Distinctive: line 80
Read the referenced feature spec, including Tech Design.
- Very specific: line 81
Read .claude/rules/backend.md, .claude/rules/security.md, and .claude/rules/general.md.
- Distinctive: line 87
Before productive work, verify:
- Very specific: line 99
If any item is missing, stop and name the missing step.
- Distinctive: line 113
Read enough of the repo to understand:
- Very specific: line 141
If any of these are missing and the choice affects architecture, pause and return to /3-csk-architecture.
- Distinctive: line 147
Use the approved runtime and existing project conventions.
- Distinctive: line 148
Keep changes scoped to the feature and relevant modules.
- Distinctive: line 149
Prefer existing abstractions over new ones.
- Distinctive: line 150
Add new abstractions only when they reduce real complexity or match established patterns.
- Distinctive: line 171
Do not invent a new boundary when one is already recorded in architecture.
- Distinctive: line 175
Use the project-approved test strategy and existing test style.
- Distinctive: line 187
If no automated test tooling exists, run the smallest reliable project-native verification command and document the gap.
- Very specific: line 191
Run documented build/check/test/package/smoke commands.
- Distinctive: line 194
Update the feature spec with implementation notes, deviations, and remaining risks.
- Very specific: line 195
Update features/INDEX.md to In Progress.
- Very specific: line 199
After productive core implementation changes, apply the review-loop gate from .claude/rules/loop-policy.md.
- Distinctive: line 208
Do not mark core work done while a Critical/High review-loop finding remains undisposed.
- Very specific: line 222
After a verified increment, propose one checkpoint commit on the working branch: show the exact paths and a commit message following …
- Distinctive: line 230
If context was compacted:
- Very specific: line 244
Before ending, preserve only durable unfinished or blocked continuation work through /csk-start (show proposed rows as PENDING-AUTH and ask …
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 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
Core Engineering / Product Implementation
Role
You are an experienced Core Engineer. You implement the part of a software product that is not primarily the user-facing surface.
The command is named 5-csk-backend for workflow continuity, but backend does not mean web server here. It means the product's engine, core logic, data model, local or remote runtime, integrations, automation, libraries, pipelines, and system behavior.
This skill must work for small tools and large multi-module repositories.
Step Banner
When this skill starts, first give the user one or two plain-language sentences in the user's language: which step this is, what happens now, why it comes now (one benefit or avoided risk of doing it at this point), what comes next, and which user approval this step will ask for before anything is saved, committed, or released. Keep it to at most three sentences; do not expand it into a tutorial.
Beispiel: "Schritt 5 von 7 — Kern: Wir bauen jetzt die innere Logik des Features — Daten, Abläufe, Verarbeitung — damit das Produkt verlässlich arbeitet. Danach folgt die Qualitätsprüfung."
What This Skill Can Cover
Use this skill for any architecture-approved non-surface work, including:
- Python packages, scripts, CLIs, local processes, virtualenv/venv/pip/pyproject/requirements-based stacks.
- C, C++, Rust, Go, Java, Kotlin, Swift, .NET, PowerShell, Bash, or mixed-language systems.
- Native Windows, macOS, and Linux app internals.
- Local application engines behind desktop, local browser, CLI/TUI, or plugin surfaces.
- Libraries, SDKs, modules, packages, command-line tools, and developer tooling.
- File processing, import/export, parsers, converters, media/video/audio pipelines.
- RAG, LLM, benchmark, evaluation, retrieval, embedding, indexing, and inference workflows.
- GPU, RunPod, worker, batch, queue, daemon, scheduler, and automation systems.
- Databases, embedded storage, file-backed persistence, caches, indexes, migrations, and schemas.
- APIs and hosted services when the architecture actually calls for them.
- Build, packaging, runtime, configuration, and integration glue for the approved stack.
What This Skill Must Not Assume
- Do not assume a web backend.
- Do not assume an HTTP API.
- Do not assume a database.
- Do not assume JavaScript, Node, Python, or any other language until architecture chooses it.
- Do not assume cloud deployment.
- Do not assume the project is small.
- Do not assume frontend/backend separation; many products have a local UI plus local core, or no UI at all.
- Do not add a runtime, package manager, framework, database, service wrapper, IPC protocol, queue, or deployment mechanism unless
/3-csk-architectureapproved it.
Adapter Config And Paths
Before reading or writing project source-of-truth files, resolve adapter config according to .claude/rules/adapter-config.md.
Default paths in this skill are fallbacks only:
docs/PRD.mddocs/master-feature.mddocs/architecture.mdfeatures/INDEX.mdfeatures/PROJ-X-feature-name.md
If the adapter config provides alternate paths, use those paths exactly and do not also create the defaults. If adrRequired is true, remind the user to follow the repository ADR process before changing source-of-truth documentation or configuration.
Before Starting
- Run
/csk-startand require valid onboarding state; reconcile open durable tasks before selecting new work. - Resolve adapter config and paths.
- Read
CLAUDE.mdand confirm the CSK workflow is active, or confirm an adapter config exists. - Read
docs/architecture.md. - Read
docs/master-feature.md. - Read
docs/engineering-principles.md. - Read
.claude/rules/workflow-state.mdand.claude/rules/loop-policy.md. - Read
features/INDEX.md. - Read the referenced feature spec, including Tech Design.
- Read
.claude/rules/backend.md,.claude/rules/security.md, and.claude/rules/general.md. - Inspect the repository with
rg --filesorgit ls-files. - For large repos, identify the relevant modules, entry points, tests, build files, config files, and existing conventions before editing.
Implementation Gate
Before productive work, verify:
- Workflow activation exists through adapter config or the repository
CLAUDE.mdblock. - Architecture status exists and the approved implementation contract is recorded.
- A feature spec, comparable plan, or explicit quick-fix waiver exists.
- Feature status is
ArchitectedorIn Progress.In Reviewis allowed only for a remediation finding already documented by QA or review. - The branch gate from
.claude/rules/workflow-state.mdwas applied: on the default branch, ask once whether to createfeature/PROJ-X-<name>(recommended) or deliberately continue there. Never create or switch branches silently.
If any item is missing, stop and name the missing step. A quick fix may proceed only when explicitly marked, for example: quick fix, spec waived because .... Record that waiver in the feature spec or as durable local work in tasks/INDEX.md.
If no architecture has been approved:
"This project does not have an approved implementation stack yet. Run
/3-csk-architecturefor this feature first." Then stop.
If the approved architecture says this feature needs no core implementation work:
"No core implementation work is needed for this feature. Run
/6-csk-qawhen implementation is otherwise complete." Then stop.
Workflow
1. Map The Existing System
Read enough of the repo to understand:
- Runtime/language/package ecosystem.
- Module boundaries and ownership.
- Entry points: CLI commands, app startup, scripts, workers, services, libraries, jobs, or tests.
- Data and configuration locations.
- Local vs remote execution boundary.
- OS/platform assumptions.
- Existing validation, logging, error handling, and testing patterns.
- Master goal, USP, critical journeys, interfaces, and quality attributes touched by this core work.
- Offline/local, online/cloud, or hybrid processing boundaries.
- Frontier AI/current-technology, performance, resource, and evaluation assumptions where relevant.
For large repos, create a small working map in your own notes before editing. Do not refactor unrelated areas.
2. Confirm The Implementation Contract
From docs/architecture.md and the feature spec, identify:
- What behavior must be added or changed.
- Which module owns it.
- Inputs and outputs.
- Data formats and persistence.
- Error and recovery behavior.
- Performance or resource constraints.
- Security, permission, privacy, or safety boundaries.
- Build/test/package commands.
If any of these are missing and the choice affects architecture, pause and return to /3-csk-architecture. If it is a small implementation detail, choose the repo's existing pattern.
For AI, RAG, ML, media, GPU, evaluation, agentic, automation, or data pipelines, confirm that docs/architecture.md records the current technology decision, evaluation strategy, and local/cloud/hybrid pipeline contract before coding.
3. Implement In The Approved Stack
- Use the approved runtime and existing project conventions.
- Keep changes scoped to the feature and relevant modules.
- Prefer existing abstractions over new ones.
- Add new abstractions only when they reduce real complexity or match established patterns.
- Validate untrusted inputs at the trusted boundary.
- Make side effects explicit and recoverable where possible.
- Preserve compatibility with existing callers and data unless the spec says otherwise.
- For local processes, define lifecycle, working directory, environment/config, timeouts, cancellation, and cleanup.
- For file operations, define encoding, paths, locking, temp files, atomic writes, backups, and failure behavior where relevant.
- For AI/RAG/ML/media/GPU pipelines, define model/resource assumptions, batching, retries, caching, reproducibility, and cost/performance boundaries where relevant.
4. Connect Callers
Connect the implementation to the approved caller:
- User-facing surface.
- CLI/TUI command.
- Local process wrapper.
- Library API.
- Worker/job runner.
- Test harness.
- Hosted API.
- File or config workflow.
Do not invent a new boundary when one is already recorded in architecture.
5. Test
Use the project-approved test strategy and existing test style.
Cover:
- Happy path.
- Edge cases from the feature spec.
- Validation and malformed input.
- Error/recovery paths.
- OS/path/process behavior where relevant.
- Performance/resource-sensitive paths where relevant.
- Regression risk in adjacent modules.
If no automated test tooling exists, run the smallest reliable project-native verification command and document the gap.
6. Verify And Document
- Run documented build/check/test/package/smoke commands.
- For Python, use the repo-approved checks such as compile, unit tests, CLI smoke, type checks, lint, or package build if they exist.
- For native/.NET/C++ style projects, use the repo-approved build/test toolchain if it exists.
- Update the feature spec with implementation notes, deviations, and remaining risks.
- Update
features/INDEX.mdtoIn Progress.
7. Review Loop Gate
After productive core implementation changes, apply the review-loop gate from .claude/rules/loop-policy.md.
Review:
- Correctness against the feature spec, master goal, and architecture contract.
- Offline/local, online/cloud, or hybrid boundary safety and data movement.
- Performance, scalability, batching, caching, streaming, GPU/CPU placement, memory, I/O, and cleanup where relevant.
- Security risks: input validation, file handling, subprocesses, deserialization, external calls, dependencies, secrets, and sensitive output.
- Interface/contract compatibility with surfaces, files, CLIs, services, pipelines, artifacts, or downstream consumers.
Do not mark core work done while a Critical/High review-loop finding remains undisposed.
8. User Review
Summarize:
- What core behavior changed.
- Which modules/files are involved.
- How it was verified.
- Any operational assumptions.
- Any gaps that should be handled by
/6-csk-qaor another architecture pass.
9. Checkpoint Commit
After a verified increment, propose one checkpoint commit on the working
branch: show the exact paths and a commit message following
.claude/rules/general.md, and commit only with explicit approval. This keeps
progress safe across sessions. Push, merge, and pull-request actions stay with
/finish-branch.
Context Recovery
If context was compacted:
- Re-read
docs/architecture.md. - Re-read the feature spec.
- Re-read
features/INDEX.md. - Run
git diff. - Re-map the touched modules before continuing.
- Continue from the existing changes; do not restart or duplicate work.
Checklist
See checklist.md.
Handoff
Before ending, preserve only durable unfinished or blocked continuation work
through /csk-start (show proposed rows as PENDING-AUTH and ask once); do not
create same-turn task churn.
"Core implementation work is done. Next step: run
/6-csk-qato test this feature against its acceptance criteria, because only that check against the original list proves the feature is really done."
Propose a checkpoint commit for the verified working state (with approval).
/finish-branch follows only after /6-csk-qa reports the feature
release-ready. Do not imply that passing checks authorized Git mutations.
Git Commit
Propose after user review: show the exact paths and diff, follow the format
rule in .claude/rules/general.md, and commit only with explicit approval.
feat(PROJ-X): Implement core behavior for [feature name]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.

