Planning — the work-item craft

A plan is work items in the repo. Everything else — a published doc, an artifact link, a chat summary, a slide — is a view of those files and links back to them. This skill exists because the opposite happened: a plan was produced as an external doc, in a table shape nobody had defined, and no guard saw it.

The domain decisions inside the plan belong to their owning roles (SYS sequences the architecture, COORD the delivery, FA the stories, PROD the priority). This craft governs the form: where the plan lives, which standard it follows, and how you prove it followed it.

1. Resolve the effective standard before writing a line

The standard is the project's, with the crew template as fallback:

1. The project's own template wins — docs/requirements/README.md and docs/stories/README.md, the fenced block under the template heading. A project that translated or reshaped its template has declared its standard by doing so.
2. Where the project has none, the crew template applies (the one /crew:setup scaffolds).
3. Declared deviations in the crew:standard block of docs/DEVIATIONS.md are applied on top. Each carries its rationale; honor them, never "fix" them.

Print it instead of remembering it:

node "${CLAUDE_PLUGIN_ROOT}/scripts/conformance.js" docs/requirements/<plan>/001-x.md

The output names the source (project or crew), the header fields, the sections, the exact table columns, and any deviation applied or ignored. Use those columns verbatim — the heading ## Estimation and its column names are what the closure gate reads.

2. Repo first, view second

• Write the work items as files first: docs/requirements/<plan>/README.md (context, index, recommended order) plus one NNN-slug.md per requirement, or stories under docs/stories/<feature>/.
• When the request asks for a link, a doc or an artifact, publish it after the files exist, as a summary that names the repo path as its source. Never put estimate tables only in the view: the view may summarize totals; the milestones live in the files.
• A plan that exists only in chat or only in an external doc is a defect, whatever the request's wording.

3. Estimation happens at two levels — respect which one you are at

• Project level (a brief, a proposal awaiting go/no-go): rough magnitude only. No milestone breakdown.
• Planning level (a work item taken for execution): whoever executes adds the ## Estimation table — milestones, both estimates, closed by a Total row — and the ## Verification table at the same moment. Stories are authored without estimation; an analyst never estimates.
• Every estimate is double, always, without waiting to be asked. Each milestone carries Est. hours at person pace (what it takes a person without agents, fixed before deciding who executes) and Est. AI hours (the agents' processing time; 0 when the milestone has no agent work). A plan with only person-pace hours is incomplete, and so is one that hides the AI figure in Notes.
• Human review is its own milestone row, never a line outside the table and never folded into another milestone: the person's time reading, correcting and approving agent work, with Est. AI hours of 0.
• The Total row sums all four columns (Est. hours, Est. AI hours, Actual hours, Actual AI hours). A plan README that summarizes several work items reports both totals.
• During execution, Started and Finished are written in real time, with a timezone offset, when the milestone begins and the moment it closes. Never reconstructed. Actual hours is the person's time; Actual AI hours comes from a capture (factory, the harness's session log) — write — when nothing measured it, since a figure typed from memory loses its value as evidence.
• When the project's own template still carries the legacy crew columns, conformance prints a notice: write the current columns anyway (both shapes conform until the next minor version) and flag in the seal that the template needs updating. When the project's template is its own shape (translated, reshaped) without an AI column, the template wins: still estimate both, write the AI figure as AI est. X h in Notes, and name the gap in the seal.

4. Delegating to subagents

When you ask a role to estimate or draft work items, give it the path of the standard (or the conformance.js output) — never a table format of your own. A role that receives a format contradicting the effective standard applies the standard and reports the requested deviation.

5. When the project has no delivery circuit

In solo mode, or when the project has no docs/requirements/ or docs/stories/, do not scaffold ceremony on your own. Say what is missing and ask where the plan should live before creating structure. The answer is the project's to give.

6. Seal

End with one line: which standard was applied (project template at <path>, crew template, or deviation N of docs/DEVIATIONS.md), what was checked (conformance.js --check on each file), and what stayed unverified.