You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
1. Large plans cause AI to lose context during /build. Implementation drifts, conventions slip, and work degrades as the context window fills. plan-splitting-agent exists but only runs in /plan-technical-review and only splits across separate PRs (part-N files). It does not chunk work within a single plan for AI-sized execution.
2. /plan and /plan-technical-review fight each other. Review is an optional follow-up that gets skipped, so plan-splitting-agent rarely fires when it matters most. Plan quality checks belong inside plan creation, not as a separate skill the user has to remember.
Proposal: collapse the two skills into one and split work into phases by default during planning so /build can execute one phase per context window.
Requirements
1. Merge /plan-technical-review into /plan
Move review agents (code-simplicity-review-agent, vgv-review-agent, plan-splitting-agent) into /plan as a mandatory step between writing the draft and presenting post-gen options.
Run agents in parallel.
Apply findings inline before the user sees the plan.
Skip review for the minimal template (overkill for small plans).
Delete /plan-technical-review skill, or keep a thin wrapper for reviewing externally-authored plans.
2. Integrate scope assessment into /plan
plan-splitting-agent runs as part of the merged review step.
If a split is recommended, present options: split into part-N files (existing behavior) or keep as a single plan with phased structure.
Skip if the agent reports no split needed — small plans stay single-phase.
3. Add "Implementation Phases" structure to the standard template
Currently only extensive.md has phases. Standard should support phases too, with explicit shape per phase:
Scope (1-2 sentences)
Files touched
Acceptance criteria
Validation steps
Each phase should fit comfortably in one AI context window.
4. Update /build to loop phases, not just tasks
Current flow loops tasks linearly. Change to: load phase → implement → validate → commit → checkpoint → offer fresh context via /clear → next phase.
Reuse the existing "Clear context and [next step]" handoff pattern between phases.
Additional Context
Plan skill: skills/plan/SKILL.md
Plan technical review skill: skills/plan-technical-review/SKILL.md
Plan splitting agent: agents/analysis/plan-splitting-agent.md
Plan templates: skills/shared/references/plan-templates/
Description
Two related pain points in the planning workflow:
1. Large plans cause AI to lose context during
/build. Implementation drifts, conventions slip, and work degrades as the context window fills.plan-splitting-agentexists but only runs in/plan-technical-reviewand only splits across separate PRs (part-N files). It does not chunk work within a single plan for AI-sized execution.2.
/planand/plan-technical-reviewfight each other. Review is an optional follow-up that gets skipped, soplan-splitting-agentrarely fires when it matters most. Plan quality checks belong inside plan creation, not as a separate skill the user has to remember.Proposal: collapse the two skills into one and split work into phases by default during planning so
/buildcan execute one phase per context window.Requirements
1. Merge
/plan-technical-reviewinto/plancode-simplicity-review-agent,vgv-review-agent,plan-splitting-agent) into/planas a mandatory step between writing the draft and presenting post-gen options.minimaltemplate (overkill for small plans)./plan-technical-reviewskill, or keep a thin wrapper for reviewing externally-authored plans.2. Integrate scope assessment into
/planplan-splitting-agentruns as part of the merged review step.3. Add "Implementation Phases" structure to the standard template
Currently only
extensive.mdhas phases. Standard should support phases too, with explicit shape per phase:Each phase should fit comfortably in one AI context window.
4. Update
/buildto loop phases, not just tasksCurrent flow loops tasks linearly. Change to: load phase → implement → validate → commit → checkpoint → offer fresh context via
/clear→ next phase.Reuse the existing "Clear context and [next step]" handoff pattern between phases.
Additional Context
skills/plan/SKILL.mdskills/plan-technical-review/SKILL.mdagents/analysis/plan-splitting-agent.mdskills/shared/references/plan-templates/skills/build/SKILL.md