Create analytic subagent in Claude code based on SKILL.md pattern

[sedlukha]

Summary

Develop an analytic subagent in Claude code following the methodologies and organization outlined in the skills/planning-with-files/SKILL.md file. The new subagent should implement a similar persistent file-based planning pattern but be specialized for analytics-focused workflows.

Details

• Reference the principles in SKILL.md, including use of persistent markdown files for working memory, error tracking, and task planning.
• Ensure the subagent supports creating and maintaining task_plan.md, findings.md, and progress.md, tailored to analytic tasks (e.g. data exploration, model tracking, finding patterns).
• Inherit workflow rituals from the Manus-style approach (read before decide, update after act, strict error/progress logging).
• Consider special requirements for analytic pipelines: logging hypotheses, experimental results, visualizations, data sources and queries, etc.
• Provide clear instructions/templates for analytic users to follow (adapt planning file templates as needed).
• Document the subagent's functions and integration points.

Acceptance Criteria

• A new Claude code subagent for analytics is implemented.
• Adheres closely to the core planning-with-files pattern and error tracking rules from SKILL.md.
• Demonstrates, via example or documentation, how the subagent improves organization and persistence for analytic work.

---

Reference: skills/planning-with-files/SKILL.md

[OthmanAdi]

Hey Arthur, thanks for opening this! Really cool to see someone thinking about extending the planning pattern into analytics workflows.

You're right that the core loop (plan, track, log errors, read before deciding) maps naturally to data work. Logging hypotheses, tracking experimental results, and keeping a record of what queries you ran and what you found are exactly the kind of things that get lost when context resets.

I've been thinking about this too. The tricky part is figuring out what the right templates look like for analytics. A data exploration session has different checkpoints than a code implementation task. Things like "what data sources did I connect to," "what did the distribution look like," "which features mattered" need their own structure.

I'm going to keep this open as a feature track. If you have ideas about what your ideal findings.md or task_plan.md would look like for an analytics workflow, I'd love to see them. Even rough sketches would help shape this.

[mvanhorn]

Implemented analytics templates in PR #115. Added a --template analytics flag to init-session.sh with 4 phases (Data Discovery, Exploratory Analysis, Hypothesis Testing, Synthesis) and domain-specific findings sections (Data Sources, Hypothesis Log, Query Results, Statistical Findings).