FACTUAL ACTIVITY RECORD · An activity record based on work that took place
Rebuilding specialist work handoff
Aligning domain-lead branch preflight and the official Pencil workflow in one dispatch sequence
Purpose
When the UI/UX Designer attempted to begin design work, the target repository did not have a branch-creation helper, so the work stopped safely before it started. Instead of granting an exception to one specialist, the organization needed to make the delegating role responsible for preparing a branch and handing over a verified state.
The same work had spent substantial time trying Pencil operations while official procedures and organization-specific checks were mixed together. The CEO asked for the agent and Skill guidance to be rebuilt from the official documentation without turning checks absent from that guidance into universal gates.
Implementation
The CEO chose to define branch preparation as a responsibility of the active domain lead rather than naming individual AI talents in allow-or-deny rules. AGENTS.md was changed so that the domain lead checked each target repository's root, base branch, working tree, diff, and work boundary before specialist handoff, then prepared a safe work branch. When a repository-local helper was absent, only a manual procedure without force, reset, or discard was allowed; if its safety could not be verified, the work had to stop before starting.
.agents/skills/software-delivery/SKILL.md required the branch packet to contain the repository root, branch, base, status and diff before and after creation, scope, helper or manual-procedure result, start conditions, dependencies, expected output, merge point, and stop conditions. The Executive Secretary recorded the packet in the active-agent ledger before launching a specialist with the branch information in a structured handoff. This consolidated preflight and removed duplicate dispatch paths.
The first review found four problems: branch creation appeared before the requirements Task and authentication check; some wording still launched specialists before preflight; executed parent and child packets lacked evidence; and the PR Reviewer was included among ordinary artifact authors. The Organization Designer revised the same operating documents so that authoring and implementation specialists produced artifacts while the PR Reviewer handled independent review and result recording.
Pencil operations were separated into a new .agents/skills/pencil-design-workflow/SKILL.md. Its standard procedure was limited to surfaces and save methods supported by the official documentation. On Desktop or in an IDE, the operator launched the app, created or opened a physical .pen, edited the open document through MCP, and saved through the app. CLI app mode connected to a running Desktop instance, while CLI headless mode named its input and output and called save() before ending the interactive session. The saved physical .pen was tracked with the product or site code on the same branch.
Verification was tied to each Task's acceptance criteria instead of being fixed for every Pencil job. Screenshots were required only for visual acceptance, layout checks only when layout acceptance existed, and exports only for requested formats. Reusable components and variables inside a normal .pen were also distinguished from a separate .lib.pen Design Library imported by multiple files, avoiding a requirement to turn every artifact into a library.
The first Skill draft included connection-attempt counts, forced fallback, node-ID checks after reopening, one export set for every job, a custom manifest schema, and a blanket ban on read-only shell inspection—none of which were established as universal requirements in the official documentation. The PR Reviewer identified the risk of mixing official workflow with organization-specific gates and requiring evidence unrelated to a Task. The revision removed the custom schema and made a Pencil-specific handoff record optional unless a Task explicitly requested it.
Verification parsed the project configuration and every agent TOML, then checked the Skill frontmatter, command rules, official-reference URLs, stale sequencing and excess-gate wording, and whitespace in the diff. Parent and child repositories were checked independently for status, diff, and scope, and the work also confirmed that configuration, existing design assets, and secret values had not entered the change.
The following excerpt from .agents/skills/pencil-design-workflow/SKILL.md was checked line by line against the primary Commit diff at the cutoff.
## Handoff and review
Aya handoffにはTaskで要求された`.pen`、必要なscreen / export、tokens / variables、components、states、accessibility、implementation notes、verification evidenceを含めます。Ritsu review requestにはartifact / diff、originating role、acceptance criteria、dependencies、verification evidence、review focusを含めます。
Ritsu explicit PASS前はcompleteと報告しません。PR ready化、merge、Issue close、merge後のmain復帰は担当しません。Rendering diagram…
What was confirmed
Verification results
The sequence from active-domain-lead branch preflight through Executive Secretary packet recording, one Central Dispatch, and specialist work matched across the operating charter, delivery Skill, setup guide, operating model, and handoff examples.
The change included independent parent and child checks, a preference for repository-local helpers, and a pre-start stop condition when a safe manual procedure could not be established.
After the four findings from the initial review were corrected, re-review found no additional issue in the Task/auth/branch/handoff/dispatch sequence, executed packets, or the PR Reviewer's review-only boundary.
The Pencil Skill was reduced to official surfaces, physical .pen saving and Git tracking, Task-specific verification, and a distinction between ordinary components and a separate Design Library. The custom manifest schema and universal gates were removed.
Pencil-workflow review reached no findings after two rounds of corrections. TOML, Skill frontmatter, rules, cross-file searches, and the diff check passed, with no unintended changes to configuration, existing assets, or secret values.
The revised operations were applied, and the related Tasks and clean post-application state were confirmed. This Activity did not evaluate the quality of Pencil screens or exports themselves, nor later implementation, legal, locale, or publication work.
Basis for completion
The Activity was complete when branch-preflight and single-dispatch responsibilities, the official-documentation-centered Pencil Skill, and Task-specific verification boundaries were aligned across operating documents and agent instructions; all independent-review findings were resolved; syntax, rules, diff, and scope checks passed; and the applied completion state was confirmed.