Local LLMs

Local LLMs usually do not auto-discover skills. Use the compiled prompt bundle instead.

Open:

prompts/local-llm-featurekit.md

Paste it into your local model, then paste the selected context and skill.

For a manual bootstrap, paste the using-featurekit skill first. For a raw feature, also paste grilling so the model can begin the design-tree interview without another prompt. For other work, paste the narrower workflow skill selected by the bootstrap.

Minimum Context

For a local model, keep context tight:

1. Product repo AGENTS.md.
2. Relevant docs/agents/*.md.
3. The relevant glossary (CONTEXT.md by default, or the one docs/agents/domain.md names), accepted ADRs, PRD, or slice.
4. using-featurekit plus grilling for a raw feature, or one focused FeatureKit SKILL.md; paste orchestrate-work or validate-feature when the task needs them, and paste secure-design, secure-code, secure-api, or secure-deploy when the focused skill tells the model to use one.
5. Your exact task.

Do not paste the whole repo unless the local model has enough context and you really need it.

Keep task state in the plan file’s slice Status, Evidence and Notes rather than in the conversation. Never paste an API key, transcript, or unrelated source into the model prompt.

Example: Planning

Use FeatureKit with the grill-with-docs skill.

Project context:
<paste AGENTS.md>
<paste docs/agents/domain.md>
<paste relevant ADRs, if any>
<paste the glossary (CONTEXT.md by default)>

Skill:
<paste skills/engineering/grill-with-docs/SKILL.md>

Task:
I want to add saved dashboards to this product. Build a design tree and interview me until its selected planning depth has no material decision left open.

Example: Implementation

Use FeatureKit with the implement skill.

Project context:
<paste AGENTS.md>
<paste docs/agents/verification.md>
<paste docs/agents/security.md if relevant>
<paste relevant accepted ADRs>

Spec:
<paste PRD>

Slice:
<paste one approved slice only>

Skill:
<paste skills/engineering/implement/SKILL.md>

Task:
Implement only this approved slice. Ask for any file contents you need before editing.

Example: Test Verification

Use FeatureKit with the verify-tests skill.

Project context:
<paste AGENTS.md>
<paste docs/agents/verification.md>
<paste relevant PRD or slice>

Skill:
<paste skills/engineering/verify-tests/SKILL.md>

Task:
Check whether these generated tests actually protect the intended behavior. Do not run mutation testing unless the skill's gate is met.

Example: Security Review

Use FeatureKit with the security-review skill.

Project context:
<paste AGENTS.md>
<paste docs/agents/security.md>
<paste relevant PRD, slice, ADRs, or diff>

Skill:
<paste skills/engineering/security-review/SKILL.md>

Task:
Review this change for exploitable security risk. Lead with findings.

Limitations

With local LLMs:

• expect more manual file pasting,
• use numbered text with at most three simple questions per interview round when the model has no structured question tool,
• keep tasks smaller,
• use implement one approved slice at a time unless the model can hold the full approved PRD, slice queue, relevant files, and verification output for execute-slices,
• ask for plans before code, and approve the written plan or slice before implementation,
• paste or restate the Agent Operating Principles before coding, refactoring, or verification work,
• verify outputs more aggressively,
• use the orchestration guide for capability routing and runtime evidence, and consult the provider reference linked there for CLI mechanics,
• use fresh sessions only for substantial work or when the current session is context-constrained, and reopen the same checkout and branch.

The workflow still works because the skills are plain Markdown.