# Plan Auditable Exploratory Code Changes

Deep learning experiments can blur candidate changes with trusted baseline work. This skill guides isolated adaptations and creates rollback-aware implementation records.

## Install

```bash
npx skillstore add lllllllama/explore-code
```

## Metadata

- Status: approved
- Slug: lllllllama-explore-code
- Skillstore revision: r2
- Version status: missing
- Tree hash: cb6e417f2b81b58e905698ba78b80c469470db51644374367bed583376b4452a
- Author: lllllllama
- GitHub username: lllllllama
- License: MIT
- Repository: https://github.com/lllllllama/rigorpilot-skills/tree/main/skills/explore-code/
- Ref: 656be3040aef5c047555a908cd5c695d22a4a548
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: confirmation\_required
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: external\_commands, filesystem
- Quality score: 38
- Quality tier: warning
- Public page: https://skillstore.pages.dev/skills/lllllllama-explore-code
- Manifest: https://skillstore.pages.dev/api/skills/lllllllama-explore-code/manifest

## Capabilities

- Guides authorized exploratory changes on an isolated branch or worktree.
- Ranks up to eight candidate Python and configuration files from repository paths.
- Combines variant specifications, idea cards, and analysis hints into a minimal patch plan.
- Generates rollback guidance and planned syntax, import, configuration, and interface checks.
- Separates candidate findings from trusted baseline conclusions.
- Defines expected changeset, changelog, comparability, run summary, and status outputs.

## Use Cases

- Adapt a Research Backbone: Plan a minimal backbone adaptation while preserving configuration wiring and rollback steps.
- Review a Module Transplant: Map a source-anchored module into likely code and configuration targets before implementation.
- Prepare an Adapter Experiment: Structure a reversible LoRA or adapter change with smoke checks and candidate-level documentation.

## Prompt Templates

### Find Candidate Targets

```
On my isolated experiment branch, identify likely files for this authorized adaptation: [change]. Keep the trusted baseline unchanged.
```

### Plan a Reversible Adaptation

```
Create a minimal patch plan for [module or backbone change]. Include target files, configuration wiring, rollback steps, and planned smoke checks.
```

### Plan from Research Inputs

```
Use [variant specification], [idea card], and [analysis hints] to plan an isolated adaptation for [task family]. Preserve [research anchor].
```

### Audit a Candidate Implementation

```
Review this exploratory implementation against [research anchor]. Summarize changed surfaces, comparability risks, rollback actions, and evidence still needed before trusted claims.
```

## Limitations

- The planner identifies targets from file paths and supplied metadata, not full source semantics.
- It plans validation but does not run training, tests, or performance comparisons.
- It requires explicit authorization and an isolated branch or worktree.
- The output wrapper depends on a shared Python module outside this skill package.

## Best Practices

- Use an isolated branch or worktree and record the trusted research anchor.
- Provide structured variant, idea, and analysis inputs when available.
- Keep adaptations small, reversible, and clearly labeled as exploratory.

## Anti Patterns

- Do not modify the trusted baseline by default.
- Do not use candidate results to claim verified reproduction or contribution.
- Do not use this skill for broad refactors or unsupported implementations from scratch.

## Security Audit

- Audited at: 2026-08-08T08:48:23.149\+00:00
- Summary: All 25 static alerts are false positives caused by Markdown formatting, ordinary Python syntax, fixed documentation paths, and conservative policy wording. However, a separate high-severity issue exists because scripts/write\_outputs.py dynamically executes an unbundled Python module from outside the skill directory.

## Stats

- Views: 0
- Downloads: 6
- Favorites: 0
- Popularity score: 0