From odkrywaj-organizator
Scaffold 8 project organizator markdown files (STARTWORK, CONTEXT, FILEMAP, TECHSTACK, WORKFLOW, PROGRESS, DECISIONS, ENDWORK) plus optional Claude Code hooks so Claude keeps full project memory across sessions and stops re-asking resolved questions. Trigger on magic words "Organizuj!", "ustaw organizatora", "Accio organizator", "bootstrap organizator", "set up organizator files". Also trigger WITHOUT magic words when the user describes the underlying problem — "Claude forgets between sessions", "persistent memory for my codebase", "stop Claude re-proposing rejected ideas", or any equivalent complaint about losing project context across Claude Code sessions. Run before serious multi-session work on a growing repo.
How this skill is triggered — by the user, by Claude, or both
Slash command
/odkrywaj-organizator:odkrywaj-organizatorThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Claude Code sessions forget. On any project bigger than a weekend hack, users waste the first 10 minutes of every session re-explaining what the code does, what was tried, and what was rejected. This skill drops 8 small markdown files in the project root that Claude reads on demand.
assets/hooks/on_pre_compact.shassets/hooks/on_session_start.shassets/hooks/on_user_prompt.shassets/hooks/settings.json.templateassets/organizator/CONTEXT.md.templateassets/organizator/DECISIONS.md.templateassets/organizator/ENDWORK.md.templateassets/organizator/FILEMAP.md.templateassets/organizator/PROGRESS.md.templateassets/organizator/STARTWORK.md.templateassets/organizator/TECHSTACK.md.templateassets/organizator/WORKFLOW.md.templatereferences/examples.mdscripts/scan_project.shClaude Code sessions forget. On any project bigger than a weekend hack, users waste the first 10 minutes of every session re-explaining what the code does, what was tried, and what was rejected. This skill drops 8 small markdown files in the project root that Claude reads on demand.
Key insight: STARTWORK.md is a router, not a document. It points Claude at the right organizator file for the current task (CONTEXT for big picture, FILEMAP for navigation, DECISIONS for rejected approaches). Loading STARTWORK first, then pulling one relevant file, beats stuffing one giant CLAUDE.md.
Before doing anything, check the project root for any of: STARTWORK.md, CONTEXT.md, FILEMAP.md, TECHSTACK.md, WORKFLOW.md, PROGRESS.md, DECISIONS.md, ENDWORK.md.
Run scripts/scan_project.sh from this skill's directory against the project root. It returns: top-level tree (depth 2), detected stack from manifests, last 10 commits, README/CLAUDE.md excerpts, folder file counts. Treat this as ground truth for TECHSTACK and FILEMAP — most of what the organizator needs is already here. If the scan returns nothing (fresh repo, missing tools), fall back to manual Q&A; never abort on scan failure.
At most four questions, one at a time, and only for things the scan couldn't determine:
Asking less is better than asking thoroughly. If the scan answered everything, ask nothing.
Read templates from assets/organizator/. Replace {{PLACEHOLDERS}} with scan data and answers. Write all 8 files to the project root — not .claude/, not a subfolder. Organizator files are living documents the user edits daily. Use relative paths (./CONTEXT.md) in cross-references.
Ask: "Install Claude Code hooks so STARTWORK.md auto-loads on every session start and ENDWORK triggers on 'kończymy' / 'endwork'? (Y/n)"
assets/hooks/ contents into the project's .claude/. If .claude/settings.json already exists, merge new hooks into the existing config, dedupe by command, preserve every other key. If merge is ambiguous (malformed JSON, schema conflict), stop and show the user both files.Print: files created, hooks status, next step ("Start a fresh session and say 'pracujemy nad projektem' or 'let's work on the project' — STARTWORK loads, Claude knows where to look"). Do not edit further in the same turn. Fresh session, clean context, organizator-driven is the whole point.
| File | Loaded when | Purpose |
|---|---|---|
STARTWORK.md | Every session start (router) | One-page map pointing to the other 7. ≤50 lines. |
CONTEXT.md | Task touches product goals or audience | Why, who for, what success looks like, what we're NOT building. |
FILEMAP.md | Claude needs to find code | Top-level folder layout + one-line purpose per folder. |
TECHSTACK.md | Task is technical | Languages, frameworks, key deps, runtime, deploy target, versions that matter. |
WORKFLOW.md | Task touches dev process | Run / test / deploy commands. Git conventions. |
PROGRESS.md | Planning next steps | Done, in-progress, next, blocked. Updated at session end. |
DECISIONS.md | Evaluating approaches | Decisions and why, append-only. Stops re-proposing rejected paths. |
ENDWORK.md | Session close ("kończymy" / "endwork") | Wrap-up checklist: update PROGRESS, log new DECISIONS, commit, push. |
STARTWORK is the entry, ENDWORK is the exit, the middle six load on demand. Middle organizator files stay under ~150 lines each. Past that they become noise — move overflow into docs/ or a real spec.
.claude/settings.json merges, never clobbers. Parse, merge, dedupe by command, preserve other keys. On failure, stop and hand the user both files.package.json, believe the user.TBD, never a guessed framework. A TBD invites the user to fill it in; a hallucination misleads every future session.For worked examples on three contrasting projects, see references/examples.md.
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.
npx claudepluginhub odkrywaj-ai/odkrywaj-organizator --plugin odkrywaj-organizator