Roadmap to Improve Plans

Turn selected strategic-roadmap opportunities or accepted feature slices into grouped batches of numbered improvement plans. This is the wide planning stage: preserve the audit judgment, feature decisions, and architecture diagnosis behind each selected item, then turn it into enough detail for future executors to work safely.

Treat plan writing as meta-work: the output should improve future execution loops by carrying forward impact, risk, confidence, design claims, regression checks, approval boundaries, and STOP conditions.

This owns this pipeline's NNN plan-bank contract directly.

Source Frame

Before writing plans, load what applies:

• The selected .agents/ROADMAP.md or equivalent roadmap/opportunity artifact.
• Accepted feature-planning artifacts when splitting a feature outline or final plan into PR-sized improvement plans.
• improve for the original audit posture: category, evidence, impact, effort, risk, confidence, fix sketch, and “not worth doing” verdicts.
• The coding-standards skill as the canonical source for standards concerns. Load only the reference files matching the selected work; do not copy the standards into this skill or the plan.
• improve-codebase-architecture for architecture/deepening candidates: current friction, shallow-module diagnosis, deletion test, locality/leverage, and recommendation strength.
• DLS-style exemplar plans when present in the target repo, especially effort indexes, numbered plans, memos, rejected/deferred notes, and reconciliation logs.
• Project docs and commands: AGENTS.md, README.md, CONTEXT.md, architecture docs, ADRs, package/test config, CI config, and current VCS state.

Workflow

1. Select plan scope
   • Start from selected roadmap items, accepted feature artifacts, or ask the user which items to plan.
   • Treat one effort directory as one improvement batch: one theme, one outcome, one README index.
   • Split unrelated projects, features, or outcomes into separate batches.
   • Preserve dependency order from the roadmap or feature outline; do not flatten everything into parallel tasks.
2. Reconcile current state
   • Read current project instructions and existing plans.
   • Check current VCS state and active branch/bookmark so the plan does not assume stale code.
   • Re-open source evidence from the roadmap before planning; old roadmap citations are leads, not facts.
   • Identify exact validation commands and any missing verification baseline.
   • Carry forward the roadmap's definition of “better”; if missing, define the plan-bank success criteria before decomposing work.
   • If credentials or secrets appear, reference only the file path, line, and credential type; never copy secret values into plans.
3. Decompose into numbered plans
   • Each NNN-*.md should be independently landable or have an explicit dependency on an earlier plan.
   • Preserve the source audit frame for roadmap/audit opportunities: audit category, impact, effort, risk, confidence, evidence, and fix sketch or direction.
   • Preserve the source feature frame for accepted feature artifacts: source design/outline, accepted decisions, phase/slice, dependencies, verification, and implementation routing.
   • Prefer vertical slices over layer-by-layer rewrites.
   • Include architecture/deepening language when relevant: module, interface, depth, seam, adapter, leverage, locality.
   • For architecture plans, carry forward the diagnosis: current friction, deletion test result, locality/leverage claim, recommendation strength, and ADR conflicts if any.
   • Do not pretend an unresolved seam or interface decision is implementation-ready. Route it to feature-planning-artifacts, a research/spike plan, or a memo instead.
   • Use the coding-standards skill to name the standards concern and design claim each plan protects.
4. Write the effort index and plans
   • Prefer .agents/plans/improvements/<effort-slug>/ unless the repo already has an established planning convention.
   • Use references/index-template.md for README.md.
   • Use references/plan-template.md for each numbered plan.
   • Use references/memo-template.md only when a design fork or handback needs a durable answer.
   • Choose the next available NNN- prefix by inspecting the destination directory.
5. Mark deeper-planning candidates and autonomy boundaries
   • In the effort index, mark plans that deserve deeper feature-planning treatment because the audit confidence is low, the architecture shape is unsettled, the recommendation is speculative, or user decisions are required.
   • For each plan, state what can be done as routine execution, what needs design review, and what requires human approval.

Plan Requirements

Every numbered plan must be self-contained for an executor with no planning-session context. Assume competence at following explicit instructions and weakness at filling gaps. Each plan must include:

• source frame: roadmap/audit category, impact, effort, risk, confidence, and fix sketch or direction when available; feature artifact, accepted decision, and slice dependency when splitting a feature;
• standards concern from the coding-standards skill, when the plan changes code shape;
• purpose and payoff;
• definition of “better” for this plan;
• current-state evidence with paths and line ranges where possible;
• desired end state;
• scope and out of scope;
• implementation sequence at PR-sized granularity;
• exact automated verification commands;
• evals or regression checks that catch the change getting worse;
• manual verification only when it proves something automation cannot;
• autonomy boundary for routine execution, design review, and human approval;
• drift checks to run before implementation;
• STOP conditions for stale assumptions, missing commands, design forks, or failed validation;
• rejected approaches and why they are not being used;
• executor handoff notes for a fresh executor.

Architecture plans must also include current friction, deepening direction, deletion-test result, locality/leverage claim, and recommendation strength. If those are unknown, the plan is not ready for routine implementation.

Guardrails

• Do not implement source changes.
• Do not write giant all-or-nothing plans.
• Do not hide unresolved design questions in TODO prose; either stop for deeper planning or write a memo request.
• Do not turn speculative architecture into implementation instructions; preserve the uncertainty and route it to design review.
• Do not leave recurring decisions as repeated advice; turn them into a policy, memo, ADR suggestion, or plan-bank rule when durable.
• Do not keep old shapes by default. Prefer the intended end state unless migration compatibility is required.
• Do not invent validation commands. If commands are unknown, make establishing the verification baseline an explicit first plan or STOP condition.
• Do not copy secrets into plans, even as evidence.

Done Criteria

The effort is done when the index explains plan order and readiness, each numbered plan can be handed to a fresh executor, and low-confidence or unsettled architecture work is flagged for deeper planning or design review instead of being marked ready prematurely.