kamehadb: Skill for Claude Code

Estimate Feature

Estimate active engineering hours for a developer using an AI coding agent. Never run this optional skill unless the user requests an estimate.

Process

1. Require a written plan under docs/agent-devkit/plans/, or a change's tasks.md under docs/agent-devkit/changes/<folder>/. If neither exists and the user also requested a plan with a written design.md, call plan-feature as a route handoff: it writes the plan, calls this skill before its approval gate, and presents both together. Otherwise stop and request a plan. Use a repair handoff to plan-feature only to fix missing detail in an existing plan. Never estimate directly from a feature request.
2. Read AGENTS.md, the plan or tasks.md, the linked design, delta.md (when present), relevant wiki, source, and tests. Call the available Skill entry whose local name is read-codebase-context when source exists. If the plan has an ## Impact map, follow read-codebase-context's ## Refresh an impact map; if its baseline is missing, unavailable, or no longer an ancestor, trace existing source from scratch. Do not try to trace planned entry points, callers, or tests that do not exist yet; use planned files, interfaces, behavior, and verification only from the written design and plan. Pending approval does not block estimation: label planned assumptions pending approval and report the gate's current status. It does not authorize implementation. Mark unspecified behavior as unknown and block an estimate when it prevents a defensible range. Re-trace planned paths only after source exists. Read current source for every existing file the plan says implementation will edit. For a planned file that does not exist yet, use the written design and plan; do not treat that file as traced source. If tasks omit files, behavior, or verification, call plan-feature as a repair handoff (see using-devkit) to fix the plan, then resume this estimate; plan-feature must not start implementation from this call.
3. State the AI support profile. Assume the coding agent can inspect and edit source, write and run tests, and update documentation, while a developer reviews output and resolves product decisions. Include context gathering, prompting, coding, tests, expected debugging, review, and planned docs. Exclude waiting, stakeholder response, deployment, and manual QA unless planned.
4. Estimate every plan task with a defensible hours range, confidence, and evidence-based rationale. Consider local patterns, novelty, touched files, migrations, integrations, test cost, and unresolved dependencies. Never apply a generic "AI is N% faster" discount. When an unknown prevents a defensible range, mark the task Blocked: spike required and keep it in the task table with Blocked in Hours, N/A confidence, and the unknown in its rationale. Exclude blocked tasks from the total and report their count. While any task is blocked, total confidence cannot be High. If every task is blocked, report the total as Not estimable — spike required and total confidence as N/A.
5. For a change, save estimate.md beside its tasks.md and link to that file relatively. For a legacy plan, save a new estimate at docs/agent-devkit/estimates/YYYY-MM-DD-<slug>-estimate.md and link to the exact plan filename. Follow the shared artifact naming rule in using-devkit (read it if it is not loaded). Create docs/agent-devkit/estimates/ only now if it does not exist. Use: