walk
Measures a product from its users' purposes instead of its own feature list — a blind setter (a separate, tool-less model call that sees only what the project says it is for) writes personas and missions, an independent walker subagent tries each mission and its variations on the running product as that person would, and every finding is routed to an issue draft, the roadmap, the stage parking lot, or the owner. Changes no product code.
- License 许可
- MIT
- 最近更新
- 2026-11-14
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- walk
- description
- Measures a product from its users' purposes instead of its own feature list — a blind setter (a separate, tool-less model call that sees only what the project says it is for) writes personas and missions, an independent walker subagent tries each mission and its variations on the running product as that person would, and every finding is routed to an issue draft, the roadmap, the stage parking lot, or the owner. Changes no product code.
- allowed-tools
- Read, Glob, Grep, Write, Edit, Bash, Agent
- argument-hint
- [--reset] [--all] [mission ids]
- disable-model-invocation
- true
02 资源文件 3
- 参考文档references/setter.md3.1 KB
- 参考文档references/walker.md4.9 KB
- 脚本scripts/set-missions.sh3.4 KB
03 SKILL.md 正文
Walk — try the product the way its users would
The point is order: fix what people come to do before looking at what the product does, then find out whether they can do it. Missions, the blind setter, field ownership and where findings go: missions.md — read it first.
Process
1. Resolve and decide what to walk
Resolve <root> (continuity-docs.md §1).
- No
<root>/missions/missions.json, or--reset→ set missions first (step 2). - Otherwise → walk: the ids in
$ARGUMENTS; with--all, everyactivemission; with neither, theactivemissions whosewalked.buildis not the current build, plus those passing through what changed since their last walk (missions.md §10).
2. Set missions — blind
- Gather the purpose material and nothing else (missions.md §2): quote it verbatim with its
source path into
<root>/missions/purpose.md. With a stage contract, itsgoalcomes first. If the project states no purpose anywhere — no contract goal, no README or scope document that says what it is for and for whom — stop: say what is missing and that/iyu:charteror a README opening would supply it. Do not write a purpose yourself. - Run the setter:
sh "${CLAUDE_SKILL_DIR}/scripts/set-missions.sh" "${CLAUDE_SKILL_DIR}/references/setter.md" <root>/missions/purpose.md <a temp file>. Exit2means no JSON came back — run it once more, then stop and report. Do not write missions yourself or ask a subagent to: either would see the code. - Record verbatim into
missions.json(schema in missions.md §4): personas and missions as returned, adding onlyid,set,state: active, variation objects withverdict: null, and asetentry with the current commit. With--reset, mark every earlier missionretiredfirst. Show the user the personas and mission goals in a short list, and the setter'sunclearlines.
3. Walk — independent
Pick NN, the next unused walk number. Dispatch one Agent call (subagent_type: general-purpose) whose prompt is the full text of
references/walker.md followed by absolute paths to
missions.json and <root>/missions/walks/, the walk number, the missions to walk, the surface, and
the command that builds and starts the product (the stage contract's run, else the project's
documented one, else "none"). Pass nothing about how the product is built inside.
4. Route — every finding has somewhere to go
Read the report. Only now compare it with the project's own records: the stage contract and its
non-goals, parking.md, ROADMAP.md, the issue drafts. Route each finding by missions.md §7 —
write the issue drafts, the ROADMAP.md phases with their self-decision record, the parking.md
lines — and fill the report's ## Routing: each finding's route, and the mission × stage map
(missions.md §8).
5. Report
A few lines: verdict counts, the findings by type with where each went, what was unverified and
why, and the mission concerns. Name what the owner should look at — only the routes that need them.
Standing rules
- No product code changes. Defects go to issue drafts; building is
drive's orrun-cycle's. - The setter's text and the walker's verdicts are theirs (missions.md §5). This skill copies, numbers, routes — it does not reword a mission or reconsider a verdict.
- A walk is a heavy verification. Walk only what step 1 selected;
--allis for a stage boundary, a release, a changed purpose, or when asked. - Output language: continuity-docs.md §5.
04 同仓库更多 Skills
drive
Pushes the open stage of a product to its agreed satisfaction bar — implement one failing scenario at a time, have an independent evaluator judge the running build, repeat until the stage passes or needs something only the owner can supply, then show the owner a demo. Use after `/iyu:charter` has a signed stage contract.
charter
Interviews the owner to set a product's goal and writes the stage contract — an ordered ladder of stages that each work on their own, with scenarios an observer can watch, a calibrated satisfaction bar, the promises each stage must keep, and every owner decision that could block it. Use before `/iyu:drive`, and again at each stage boundary.
accept
Judges the open stage by having an independent evaluator subagent drive the real, running build like a user — scenario by scenario, against the rubric — and record an evidence report and the status fields. Use to get a verdict without building anything; `/iyu:drive` dispatches the same evaluator each round.
telemetry-az
Analyzes Azure Application Insights telemetry since the last run to surface defects, performance regressions, and feature-drop signals, tracked run-over-run as a quality trend (error-rate/p95/issue-count direction, with closed-issue recurrence detection), plus a report-only user-analytics reading (active-user growth, feature/page preference shifts, engagement). Triages findings against project philosophy, files issues only for threshold-crossing ones, and leads every report with a both-purpose Trend section backed by a 12-run history.
ship
Takes finished work out — bumps the version, commits, pushes, and watches the CI run to completion, reporting what failed if it fails. Checks first whether now is the right moment — if remaining backlog work would touch the same consumer-facing surface, it says so and asks rather than publishing a version the next few items obsolete.
run-cycle
Executes adaptive iterative development cycles — each cycle is a self-contained plan/execute/verify/reflect loop that reshapes the roadmap and derives emergent follow-on scope from its own output