co-customize

name

co-customize

description

Configure the project's skill stack and supervision preferences. Reads the curated registry in `skillpacks/skill_dictionary.yaml`, asks a short preset-first set of questions about workflow, dependency tolerance, autonomy style, and resource policy, then writes or updates `.co-researcher/skills.yaml`. Trigger phrases: "customize my stack", "configure skillpacks", "set up my skills", "choose presets", "configure supervision and packs", "personalize this project".

Co-Customize

Use this skill when the user wants to choose a preferred stack for the current project instead of manually copying many external skills.

---

Goals

co-customize should:

1. Read the curated registry in skillpacks/skill_dictionary.yaml.
2. Recommend packs and subsets by capability, not just by repo name.
3. Configure supervision policy alongside skill selection.
4. Write the result to .co-researcher/skills.yaml.
5. Keep the result easy to read and edit manually.

If .co-researcher/skills.yaml is absent, create it. If present, preserve unrelated fields where possible.

Use templates/skills.yaml.template as the starting shape when you need to create the file from scratch.

When expanding a preset into the full file, merge: preset's packs and supervision.mode/risk_profile take precedence; all other fields (preferences, full supervision subtree) are filled from template defaults unless the user explicitly changed them.

---

Inputs

Registry sources

Read:

• skillpacks/skill_dictionary.yaml
• skillpacks/presets/*.yaml

These are the source of truth for available packs, curated subsets, and preset bundles.

Existing project state

If present, also read:

• .co-researcher/skills.yaml
• RESEARCH.md

Use them to infer whether the project is literature-heavy, experiment-heavy, publication-focused, or still exploratory.

---

Interaction model

Use a short preset-first flow. Ask all questions together when possible.

Required questions

1. Workflow type

   • core-only
   • low-dependency
   • literature-heavy
   • experiment-heavy
   • academic-rigor
   • balanced
   • custom

2. Dependency tolerance

   • minimal
   • balanced
   • best-in-class even if heavier

3. Autonomy style

   • manual
   • checkpointed
   • autonomous
   • wild

4. Resource policy

   • local only
   • APIs allowed
   • remote compute allowed

5. Preference knobs

   • minimize overlap?
   • prefer subset installs?
   • allow experimental packs?

Existing config editing

If .co-researcher/skills.yaml already exists, begin by asking:

• keep current selection and only adjust supervision?
• adjust packs only?
• adjust both?

---

Recommendation rules

Capability-first mapping

• If the user wants stronger literature search/synthesis, prioritize aris/research-lit, aris/research-refine, feynman/alpha-research, or academic-research-skills/deep-research.
• If the user wants stronger experiment planning, prioritize aris/experiment-plan.
• If the user wants claim/interpretation checking, prioritize aris/result-to-claim.
• If the user wants academic rigor in writing/review, prioritize academic-paper-reviewer.
• If the user wants compute-heavy execution, surface NanoResearch as a niche/heavy option rather than a default.

Dependency warnings

• feynman/alpha-research requires AlphaXiv authentication (feynman alpha login). Always surface this when recommending literature-heavy or any feynman skill.
• nanoresearch/project-experiment requires local GPU or cluster access. Always surface this when recommending experiment-heavy with NanoResearch.

Default posture

• Always keep core enabled.
• Prefer curated subsets over full-pack installs.
• Avoid recommending heavy overlapping packs unless the user explicitly wants them.

Preset choice

Prefer a preset when the user's answers fit one cleanly. Fall back to custom pack selection only when necessary.

balanced should map to the balanced preset in skillpacks/presets/balanced.yaml.

---

Output file

Write or update:

.co-researcher/skills.yaml

Expected shape:

version: 1
profile: balanced

selection:
  preset: balanced

enabled_packs:
  core:
    enabled: true
  aris:
    enabled: true
    mode: subset
    selected_skills:
      - research-lit
      - research-refine
      - experiment-plan
      - result-to-claim

preferences:
  minimize_overlap: true
  prefer_low_dependency: true
  prefer_subset_installs: true
  allow_experimental_packs: false

supervision:
  mode: checkpointed
  risk_profile: balanced
  approval_policy:
    install_new_skills: ask
    update_registry: ask
    modify_config: ask
    launch_experiments: ask
    use_external_api: ask_first_use
    run_long_review_loops: ask
    create_commits: ask
  resource_policy:
    local_only: false
    cloud_allowed: false
    cluster_allowed: false
  budget_policy:
    enabled: true
    api_budget_level: medium
  execution_limits:
    max_review_rounds: 2
    require_confirmation_for_new_pack_imports: true

---

Required behavior after recommendation

Before writing the file:

1. Show the recommended preset or custom stack.
2. Show selected packs and selected skills.
3. Show supervision mode and any notable approval/resource settings.
4. Explain tradeoffs briefly.
5. Ask for confirmation.

After writing the file:

1. Tell the user where the config was written.
2. Summarize the chosen stack.
3. Mention that the file is intended to be hand-editable.

---

Constraints

• Do not install external skills silently.
• Do not mutate skillpacks/skill_dictionary.yaml; that belongs to co-evolve.
• Do not recommend full-pack imports unless the user clearly asks for them.
• Prefer simple, project-local configuration over hidden session-only state.