agentleFS
YulivuYulivu/debuffer-skills/skills/research-repo-architect/SKILL.md

research-repo-architect

Skill9 starsChanged 3 months agoMITView on GitHub

Organize research code and skill repositories into lightweight reproducible architecture. Use when Codex needs to scaffold or migrate a research repo for different startup stages, separate reusable code from scripts, standardize data/config/results/run directories, add Python src-layout packaging, create AutoDL/HPC-ready entrypoints with smoke/formal gates, keep local work light, or integrate skills/tools/shared-references with mainline and Codex mirrors.

  • Installs packages

The file

Show its rules and text

We split this skill into the rules it gives an agent

60 rules. About 0 are common, 13 are distinctive, and 47 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈13

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 47

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Add AutoDL/HPC hooks when experiments may need GPU time, but keep execution gated and manual by default: runbook, preflight, smoke suite, …

    line 31
  2. Very specific:

    Read ../shared-references/project-guide-protocol.md when creating project memory.

    line 34
  3. Very specific:

    Read ../shared-references/autosci-lite-patterns.md when scaffolding idea discovery, failure memory, pilot gates, or the macro state machine.

    line 47
  4. Very specific:

    When the repo will eventually produce a paper package, prefer an ICDE_YU_Memory-style manuscript layout or better: thin main.tex, modular …

    line 51
  5. Distinctive:

    Start by deciding which path applies:

    line 58
  6. Distinctive:

    New repo scaffold: Create the architecture first, then add code and experiments into the right layer.

    line 60
  7. Very specific:

    research blueprint handoff: When a stable method is about to become a formal experiment protocol or paper plan, preserve …

    line 64
  8. Very specific:

    AutoDL/HPC-ready research project: Add target-machine setup, preflight, smoke, result-transfer, and formal-run approval boundaries without …

    line 68
  9. Distinctive:

    Lightweight project start: When materials are sparse, create only the minimal brief, plan, architecture, and next-action files needed for …

    line 69
  10. Very specific:

    AutoSci-lite lifecycle: Use A-E idea paths, append-only failure memory, pilot gates, and the explicit PROJECT_STATUS.md phase map without …

    line 73
  11. Very specific:

    Keep reusable implementation under src/<package_name>/.

    line 86
  12. Very specific:

    Put runnable entrypoints under scripts/:

    line 87
  13. Very specific:

    Put hand-written configs under experiments/configs/ and suite definitions under experiments/suites/.

    line 92
  14. Very specific:

    Add a short root README.md for installation, smoke checks, data, and reproduction commands.

    line 93
  15. Very specific:

    Create docs/project/, docs/experiments/, docs/evidence/, docs/paper/, docs/theory/, and docs/runbooks/ as needed.

    line 94
  16. Very specific:

    Add a conservative .gitignore so raw data, generated runs, caches, model checkpoints, and large binary artifacts are not tracked.

    line 97
  17. Very specific:

    If the repo uses project-local skills, keep skill runtime state out of Git by default: .debuffer_skills/traces/, .debuffer_skills/cache/, …

    line 98
  18. Very specific:

    Keep compact project memory tracked when it exists: idea-stage/IDEA_MEMORY.md and experiments/NEGATIVE_RESULTS.md are decision memory, not …

    line 99
  19. Distinctive:

    If the reference is a strong manuscript package template, also record what is worth copying at the package-architecture level: section …

    line 108
  20. Distinctive:

    When idea exploration or experiment planning begins, initialize compact memory only as needed:

    line 127
  21. Very specific:

    Do not create a full research-wiki/ or graph unless the user asks for a long-lived knowledge base.

    line 133
  22. Very specific:

    Use rg --files to inventory code, notebooks, scripts, configs, data, and outputs.

    line 140
  23. Distinctive:

    Read package/build files and the main run scripts.

    line 141
  24. Very specific:

    Update imports, CLI commands, README instructions, tests, and .gitignore after each phase.

    line 150
  25. Very specific:

    Never overwrite existing data/raw/, experiments/results/, or user-created run outputs unless the user explicitly asked for that exact …

    line 151
  26. Very specific:

    Use this path when the repository is a debuffer-style skill pack: it has skills/<name>/SKILL.md, skills/skills-codex/, …

    line 155
  27. Very specific:

    Read references/skill-pack-architecture.md before editing.

    line 157
  28. Very specific:

    Treat skills/<name>/SKILL.md as the mainline source of behavior.

    line 158
  29. Very specific:

    Put reusable cross-skill contracts under skills/shared-references/.

    line 159
  30. Distinctive:

    When adding or renaming a skill in a skill repo, update all required surfaces in the same pass:

    line 161
  31. Distinctive:

    Run the repository's inventory and targeted tests after integration.

    line 167
  32. Very specific:

    Keep skill packages under skills/.

    line 174
  33. Distinctive:

    Use this path when a normal research project is being operated by skill workflows.

    line 178
  34. Very specific:

    Keep reusable implementation in src/<package_name>/ and experiment runners in scripts/ as usual.

    line 180
  35. Very specific:

    Keep skill handoff directories where downstream skills expect them unless the user asks for a breaking migration: idea-stage/, …

    line 181
  36. Very specific:

    Place Markdown handoff artifacts under docs/ when creating or refreshing them: docs/project/RESEARCH_BRIEF.md, …

    line 187
  37. Very specific:

    Treat .debuffer_skills/ as runtime state.

    line 193
  38. Very specific:

    If research-wiki/ exists, preserve it as persistent project memory.

    line 194
  39. Distinctive:

    Use this path when the repo will run on AutoDL or another SSH GPU/HPC machine.

    line 199
  40. Very specific:

    Read references/autodl-hpc.md before adding or changing remote-run scripts.

    line 201
  41. Very specific:

    Keep target-machine setup in scripts/autodl_setup.sh and machine checks/wrappers in scripts/hpc/.

    line 202
  42. Very specific:

    Keep AutoDL smoke suites under experiments/suites/, with raw outputs under experiments/runs/autodl_smoke/ and preflight reports under …

    line 203
  43. Very specific:

    Add a pilot gate before formal suites: bounded pilot spec, raw outputs under experiments/runs/pilot/, and pass/fail/inconclusive decision …

    line 204
  44. Distinctive:

    Keep formal suite definitions disabled or explicitly gated by default.

    line 207
  45. Very specific:

    Keep AutoDL operational docs in docs/runbooks/AUTODL_HPC_RUNBOOK.md or an equivalent runbook.

    line 208
  46. Very specific:

    Treat large uploaded data as an explicit manifest concern: expected paths under data/raw/ or data/processed/, with data/DATA_MANIFEST.md …

    line 209
  47. Very specific:

    Treat src/<package_name>/ as the source of truth for reusable behavior.

    line 214
  48. Very specific:

    Treat scripts/ as orchestration only: parse args, load config, call package APIs, write outputs.

    line 215
  49. Very specific:

    Treat experiments/configs/ as hand-written intent and experiments/suites/ as reproducible task selection.

    line 216
  50. Very specific:

    Treat experiments/runs/ as raw run provenance and experiments/results/ as curated, stable results.

    line 217
  51. Very specific:

    Prefer metrics.json, summary.md, resolved_config.yaml, and run_metadata.json for experiment output bundles.

    line 218
  52. Very specific:

    Use --config, --out-dir, and optional --set key=value CLI conventions for experiment scripts.

    line 219
  53. Very specific:

    Keep data directories present with .gitkeep, but ignore large files under data/raw/, data/interim/, and data/processed/.

    line 220
  54. Very specific:

    Add .gitattributes with LF line endings for shell scripts when HPC/Linux execution matters.

    line 221
  55. Very specific:

    Use assets/repo-template/ for a new repo baseline or as a target shape during migration.

    line 240
  56. Very specific:

    Read references/architecture.md when deciding where files belong.

    line 241
  57. Very specific:

    Read references/skill-pack-architecture.md when the repo contains skills, shared references, installer scripts, Codex mirrors, or skill …

    line 242
  58. Very specific:

    Read references/autodl-hpc.md when the repo mentions AutoDL, /root/autodl-tmp, FileZilla/SFTP, scripts/autodl_setup.sh, …

    line 243
  59. Very specific:

    Read references/migration-playbook.md before reorganizing an existing repo.

    line 244
  60. Distinctive:

    Do not copy SpectralStore's graph-compression domain code.

    line 245

≈ “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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Research Repo Architect

Capability Routing

This is a first-layer entry skill. Keep it loaded as the user-facing route; when a request needs a specialized capability below, resolve the debuffer repo root from .debuffer_skills/installed-skills-codex.txt (repo_root) when available, read the referenced library SKILL.md, then follow that skill. Do not copy the whole library skill into this file.

  • /system-profile: read ../library/experiments/system-profile/SKILL.md.
  • /training-check: read ../library/experiments/training-check/SKILL.md.
  • /run-experiment: read ../library/experiments/run-experiment/SKILL.md.
  • /overleaf-package: read ../library/paper/overleaf-package/SKILL.md.
  • /submission-repo-cleanroom: read ../library/review/submission-repo-cleanroom/SKILL.md.

Use this skill to turn research code or a skill bundle into a maintainable, reproducible repository. The base style is inspired by SpectralStore: reusable library code lives under src/, scripts are thin entrypoints, configs and suites describe experiments, raw runs are kept separate from curated results, and large artifacts stay out of Git.

Customized Pack Defaults

For research projects, read ../shared-references/lightweight-research-pack.md when available. Default to a lightweight, AutoDL-first shape:

  • Classify startup mode: venue-only, reference-paper, reference-codebase, idea-doc, existing-repo, or partial-results.
  • Local repo work should prepare structure, code, configs, tests, tiny smoke checks, prompts, and audit scripts. Design the scaffold around lightweight local verification plus remote-ready handoff.
  • Add AutoDL/HPC hooks when experiments may need GPU time, but keep execution gated and manual by default: runbook, preflight, smoke suite, data manifest, and formal-run approval.
  • Read ../shared-references/project-guide-protocol.md when creating project memory. Keep PROJECT_STATUS.md current so future sessions know the macro phase; keep other Markdown project memory under docs/ category directories, such as docs/project/PROJECT_BRIEF.md, docs/project/NEXT_ACTIONS.md, docs/evidence/findings.md, and docs/experiments/EXPERIMENT_LOG.md; create or refresh docs/project/RESEARCH_BLUEPRINT.md, docs/project/BLUEPRINT_GATE.md, or docs/project/PROJECT_GUIDE.md only at stage gates.
  • Enforce root Markdown hygiene for ordinary research repos: root .md files are limited to README.md, PROJECT_STATUS.md, and tool-managed AGENTS.md / CLAUDE.md when present. Put all other Markdown under docs/ with a category directory. Read legacy root files when they already exist, but write new or refreshed artifacts to docs/.
  • Read ../shared-references/autosci-lite-patterns.md when scaffolding idea discovery, failure memory, pilot gates, or the macro state machine.
  • For external review, create review-prompts/ rather than wiring direct API, MCP, or SSH automation into the repository.
  • When the repo will eventually produce a paper package, prefer an ICDE_YU_Memory-style manuscript layout or better: thin main.tex, modular section files, dedicated figure assets, and contribution-bearing benchmark/discussion sections instead of a flat monolithic manuscript.

Workflow

Start by deciding which path applies:

  • New repo scaffold: Create the architecture first, then add code and experiments into the right layer.
  • Existing repo migration: Audit first, produce a migration map, then move code in small validated phases.
  • skill framework or skill repo: Preserve the skills/, skills/skills-codex/, skills/shared-references/, tools/, templates/, docs/, tests/, and mcp-servers/ contracts instead of forcing everything into src/ and scripts/.
  • skill-managed research project: Combine the SpectralStore experiment layout with skill runtime artifacts such as research-wiki/, idea-stage/, refine-logs/, review-stage/, paper/, and .debuffer_skills/.
  • research blueprint handoff: When a stable method is about to become a formal experiment protocol or paper plan, preserve docs/project/RESEARCH_BLUEPRINT.md and docs/project/BLUEPRINT_GATE.md as tracked project memory instead of burying the decisions in terminal logs.
  • AutoDL/HPC-ready research project: Add target-machine setup, preflight, smoke, result-transfer, and formal-run approval boundaries without weakening source/config/run/result separation.
  • Lightweight project start: When materials are sparse, create only the minimal brief, plan, architecture, and next-action files needed for the current startup mode. Avoid generating a full research wiki or large Markdown bundle until the project has evidence worth preserving.
  • AutoSci-lite lifecycle: Use A-E idea paths, append-only failure memory, pilot gates, and the explicit PROJECT_STATUS.md phase map without adopting a full wiki platform.

For detailed directory rules, read references/architecture.md. For skill-pack-specific work, read references/skill-pack-architecture.md. For AutoDL/HPC work, read references/autodl-hpc.md. For migration work, also read references/migration-playbook.md.

New Repo Scaffold

  1. Identify the project slug, package/module name, research objective, first experiment ids, expected data sources, and whether the repo is primarily Python. If the user did not specify a package name, derive a lowercase underscore Python package name from the repo slug.
  2. Copy or recreate assets/repo-template/ into the target repo. Replace placeholders:
    • {{project_slug}}: repository/project slug, usually kebab-case.
    • {{package_name}}: importable module name, usually snake_case.
    • {{project_title}}: human-facing project title.
  3. Keep reusable implementation under src/<package_name>/. Do not put reusable algorithms, models, loaders, metrics, or query logic directly in scripts/.
  4. Put runnable entrypoints under scripts/:
    • scripts/expN/ for paper or study experiments.
    • scripts/data/ for download and preprocessing.
    • scripts/analysis/ for figures, tables, and summaries.
    • scripts/hpc/ for preflight, cluster wrappers, and machine-specific helpers.
  5. Put hand-written configs under experiments/configs/ and suite definitions under experiments/suites/. New run outputs must go under experiments/runs/, not experiments/results/.
  6. Add a short root README.md for installation, smoke checks, data, and reproduction commands. Use docs/ for internal runbooks and maintenance notes.
  7. Create docs/project/, docs/experiments/, docs/evidence/, docs/paper/, docs/theory/, and docs/runbooks/ as needed. Do not scatter Markdown files in the repo root beyond the allowlist.
  8. Add a conservative .gitignore so raw data, generated runs, caches, model checkpoints, and large binary artifacts are not tracked.
  9. If the repo uses project-local skills, keep skill runtime state out of Git by default: .debuffer_skills/traces/, .debuffer_skills/cache/, .debuffer_skills/runs/, .debuffer_skills/meta/events.jsonl, and project-local .agents/skills/ symlinks.
  10. Keep compact project memory tracked when it exists: idea-stage/IDEA_MEMORY.md and experiments/NEGATIVE_RESULTS.md are decision memory, not runtime traces.

Startup Mode Scaffolds

  • venue-only: create docs/project/PROJECT_BRIEF.md, docs/project/NEXT_ACTIONS.md, a concise venue-risk checklist, and optional review-prompts/venue_direction_review_prompt.md.
  • reference-paper: add references/ or docs/literature/ notes, a claim map, and a reproduction/extension plan before writing experiment code. If the reference is a strong manuscript package template, also record what is worth copying at the package-architecture level: section topology, figure strategy, benchmark-construction handling, and discussion placement.
  • reference-codebase: first audit the upstream repo, license, entrypoints, environment, and tests; add wrapper scripts instead of modifying upstream code blindly.
  • idea-doc: preserve the original idea doc under docs/project/ or docs/ideas/, extract assumptions into docs/project/PROJECT_BRIEF.md, and write the smallest falsifiable validation plan.
  • existing-repo: run the migration workflow; avoid overwriting existing data, results, notebooks, or user notes.
  • partial-results: inventory logs/figures/tables first, write docs/evidence/findings.md and docs/experiments/EXPERIMENT_LOG.md, then backfill missing configs or audit scripts.

For every startup mode, create or update PROJECT_STATUS.md with the current macro phase, target venue, last accepted artifact, next gate, blockers, and the phase map from project-guide-protocol.md.

When idea exploration or experiment planning begins, initialize compact memory only as needed:

  • idea-stage/IDEA_MEMORY.md for accepted/rejected/pilot-status idea entries.
  • experiments/NEGATIVE_RESULTS.md for failed or inconclusive run evidence.

Do not create a full research-wiki/ or graph unless the user asks for a long-lived knowledge base.

Existing Repo Migration

  1. Inspect before editing:
    • Run git status --short.
    • Use rg --files to inventory code, notebooks, scripts, configs, data, and outputs.
    • Read package/build files and the main run scripts.
  2. Classify every important file into one of these destinations: src/, scripts/, experiments/configs/, experiments/suites/, experiments/runs/, experiments/results/, experiments/visualizations/, data/, docs/, or archive.
  3. Present or record a migration map before moving files when the repo is large or risky. Include root Markdown cleanup: keep only the allowlist in root, move other Markdown into docs/project/, docs/experiments/, docs/evidence/, docs/paper/, docs/theory/, docs/runbooks/, or docs/archive/. Preserve user changes and do not delete old locations until imports, commands, and tests pass.
  4. Move in phases:
    • Package reusable code into src/<package_name>/.
    • Convert ad hoc scripts into thin CLI entrypoints.
    • Move hard-coded parameters into YAML configs.
    • Route new outputs to experiments/runs/.
    • Move only reviewed, canonical outputs to experiments/results/ or experiments/visualizations/.
  5. Update imports, CLI commands, README instructions, tests, and .gitignore after each phase.
  6. Never overwrite existing data/raw/, experiments/results/, or user-created run outputs unless the user explicitly asked for that exact replacement.

Skill Repo Integration

Use this path when the repository is a debuffer-style skill pack: it has skills/<name>/SKILL.md, skills/skills-codex/, skills/shared-references/, tools/, AGENT_GUIDE.md, or installer scripts.

  1. Read references/skill-pack-architecture.md before editing.
  2. Treat skills/<name>/SKILL.md as the mainline source of behavior. Treat skills/skills-codex/<name>/SKILL.md as the Codex mirror that must preserve semantics while changing reviewer/tool routing only.
  3. Put reusable cross-skill contracts under skills/shared-references/. Put shared executable helpers under tools/. Put single-owner helpers under skills/<owner>/scripts/ and keep compatibility tools/ shims when existing skills or installs depend on them.
  4. Any helper invocation in a skill must use the canonical resolver chain from integration-contract.md: owner layer 0 when applicable, then .debuffer_skills/tools/<helper>, tools/<helper>, and the repo root recorded in the install manifest.
  5. When adding or renaming a skill in a skill repo, update all required surfaces in the same pass:
    • skills/<name>/SKILL.md
    • skills/skills-codex/<name>/SKILL.md
    • overlays only if reviewer routing differs
    • docs/SKILLS_CATALOG.md
    • count-bearing docs/tests such as AGENT_GUIDE.md, the single root README.md, and inventory tests
  6. Run the repository's inventory and targeted tests after integration. For this pack, prefer:
python tools/check_skills_inventory.py
python -m pytest tests/test_codex_skill_mirror.py tests/test_codex_install_update.py tests/test_install_*_tools_symlink.py
  1. Keep skill packages under skills/. src/ is for reusable project/library code, not for declarative SKILL.md bundles.

Skill-Managed Research Projects

Use this path when a normal research project is being operated by skill workflows.

  • Keep reusable implementation in src/<package_name>/ and experiment runners in scripts/ as usual.
  • Keep skill handoff directories where downstream skills expect them unless the user asks for a breaking migration: idea-stage/, refine-logs/, review-stage/, paper/, and research-wiki/.
  • Inside paper/, prefer a thin main.tex, sections/ for prose, figures/ for final assets, tables/ for generated fragments when useful, and appendix/ for appendix-only assets.
  • Place Markdown handoff artifacts under docs/ when creating or refreshing them: docs/project/RESEARCH_BRIEF.md, docs/experiments/EXPERIMENT_PLAN.md, docs/paper/NARRATIVE_REPORT.md, docs/project/RESEARCH_BLUEPRINT.md, and docs/project/BLUEPRINT_GATE.md. Read legacy root-level versions as fallback only.
  • Treat .debuffer_skills/ as runtime state. Keep traces, caches, wakeup state, and local skill symlinks out of Git unless the project explicitly wants reproducible agent provenance.
  • If research-wiki/ exists, preserve it as persistent project memory. Do not move it under docs/ or experiments/.
  • Experiment outputs still go under experiments/runs/; curated tables and figures still go under experiments/results/ and experiments/visualizations/.

AutoDL/HPC-Ready Research Projects

Use this path when the repo will run on AutoDL or another SSH GPU/HPC machine.

  • Read references/autodl-hpc.md before adding or changing remote-run scripts.
  • Keep target-machine setup in scripts/autodl_setup.sh and machine checks/wrappers in scripts/hpc/.
  • Keep AutoDL smoke suites under experiments/suites/, with raw outputs under experiments/runs/autodl_smoke/ and preflight reports under experiments/runs/preflight/.
  • Add a pilot gate before formal suites: bounded pilot spec, raw outputs under experiments/runs/pilot/, and pass/fail/inconclusive decision recorded before formal approval.
  • Keep formal suite definitions disabled or explicitly gated by default. Require a dry-run and user approval before formal execution.
  • Keep AutoDL operational docs in docs/runbooks/AUTODL_HPC_RUNBOOK.md or an equivalent runbook.
  • Treat large uploaded data as an explicit manifest concern: expected paths under data/raw/ or data/processed/, with data/DATA_MANIFEST.md updated for anything not tracked in Git.
  • Preserve the evidence boundary: smoke output validates engineering readiness only; curated claims and paper-facing tables are generated locally from audited raw runs.

SpectralStore-Style Rules

  • Treat src/<package_name>/ as the source of truth for reusable behavior.
  • Treat scripts/ as orchestration only: parse args, load config, call package APIs, write outputs.
  • Treat experiments/configs/ as hand-written intent and experiments/suites/ as reproducible task selection.
  • Treat experiments/runs/ as raw run provenance and experiments/results/ as curated, stable results.
  • Prefer metrics.json, summary.md, resolved_config.yaml, and run_metadata.json for experiment output bundles.
  • Use --config, --out-dir, and optional --set key=value CLI conventions for experiment scripts.
  • Keep data directories present with .gitkeep, but ignore large files under data/raw/, data/interim/, and data/processed/.
  • Add .gitattributes with LF line endings for shell scripts when HPC/Linux execution matters.
  • In skill repos, keep generated runtime traces under .debuffer_skills/traces/ and keep reviewer trace references in audit artifacts instead of pasting reviewer summaries into main artifacts.
  • For AutoDL/HPC repos, keep deploy-key bootstrap, FileZilla/SFTP data exceptions, smoke pass criteria, and formal-run approval gates in a runbook instead of burying them in terminal history.

Validation

For Python repos, validate with the smallest reliable sequence:

python -m pip install -e ".[dev]"
python -c "import {{package_name}}; print('{{package_name}} import ok')"
python -m pytest
python -m ruff check src scripts

For experiment structure, validate one dry run or tiny smoke task that writes only to experiments/runs/local_checks/.... Confirm no command writes directly to experiments/results/ unless it is an explicit curation or visualization step.

Resource Usage

  • Use assets/repo-template/ for a new repo baseline or as a target shape during migration.
  • Read references/architecture.md when deciding where files belong.
  • Read references/skill-pack-architecture.md when the repo contains skills, shared references, installer scripts, Codex mirrors, or skill workflow artifacts.
  • Read references/autodl-hpc.md when the repo mentions AutoDL, /root/autodl-tmp, FileZilla/SFTP, scripts/autodl_setup.sh, scripts/hpc/preflight_autodl.py, run_autodl_smoke.sh, or formal-suite gating.
  • Read references/migration-playbook.md before reorganizing an existing repo.
  • Do not copy SpectralStore's graph-compression domain code; copy only the repository architecture and reproducibility conventions.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.