AGENT SKILL

playtest

Playtest Continuity Loom as a source-and-doc-blind author in the real browser, creating a new story or continuing from a supplied playtest report through one accepted local prose segment. Use when the user asks to playtest the app with a story premise, asks to create a story through the app, or supplies a prior playtest report and asks to continue it. Evaluates the prose prompt and any naturally invoked Ideate, Record Hygiene, or Segment Reconciliation prompts with cold-context subagents without making OpenRouter requests, then writes a cumulative evidence-backed report.

License 许可
GPL-3.0
最近更新
2026-11-10
Stars
★ 0
Forks
0

01 前言 Front Matter

name
playtest
description
Playtest Continuity Loom as a source-and-doc-blind author in the real browser, creating a new story or continuing from a supplied playtest report through one accepted local prose segment. Use when the user asks to playtest the app with a story premise, asks to create a story through the app, or supplies a prior playtest report and asks to continue it. Evaluates the prose prompt and any naturally invoked Ideate, Record Hygiene, or Segment Reconciliation prompts with cold-context subagents without making OpenRouter requests, then writes a cumulative evidence-backed report.

02 资源文件 16

03 SKILL.md 正文

Continuity Loom Author Playtest

One invocation is one sincere author journey through the visible app, ending after exactly one new accepted segment and post-acceptance continuity work, or after the blocker policy is exhausted. Always produce a report. Observe and report; never fix the app during the run.

Non-negotiables

  • Instructed, source-and-doc-blind. Follow this workflow, but learn the product itself through visible UI and field help—not source, tests, app docs, APIs, SQLite, or hidden state. For a new story, seal the intended story, expectations, and initial mental model before opening the app. This method does not establish uninstructed human discoverability. A Cold First-View Witness measures only first-screen comprehension and never human transfer or a full unaided journey.
  • Returning author on continuation. Read only the report the user supplied, reopen its /tmp project through the UI, and verify the latest accepted segment in Accepted Segments. Do not search for or read unrelated earlier reports.
  • Visible UI only. Create and edit project data through controls a human sees. Do not call app APIs directly, mutate the DOM, inject state, edit project files, or use console/runtime evaluation to advance the story.
  • No paid requests. Start an isolated app process with an explicitly blank OPENROUTER_API_KEY and temporary settings directory. Start the guarded browser before the first navigation. Never click Generate, Ideate, Analyze, Review, refresh-model, or any other control that sends to OpenRouter. A provider request attempt is a run defect even when the guard blocks it.
  • Cold prompt proof. Extract only a prompt visibly exposed by the app, pass that exact prompt to a fresh cold-context subagent, and receive its raw response. Give the subagent no story brief, expected answer, suspected defect, prior response, or repository context beyond the prompt file.
  • Human gatekeeping. Judge raw assistance before manually re-authoring useful suggestions in Records or Generation Brief. Judge raw prose before pasting it through Write or paste candidate, visibly editing it, and accepting it. Never treat assistance output or accepted prose as automatic continuity authority.
  • Naturalistic coverage. Fill fields and use Private Notes, Ideate, Record Hygiene, and Segment Reconciliation when a sincere author would. Do not manufacture irrelevant data merely to tick a feature box; record every skipped surface and why it was not naturally needed.
  • Transient story data. New projects, prompts, raw subagent responses, candidate prose, routine screenshots, settings, and the structured scratchpad stay under /tmp. Continuation works only while the project path named by the supplied report still exists.
  • Privacy-safe reports. Never put full prompts, full record payloads, candidate prose, accepted prose, API keys, or raw assistance responses in reports/. Retain fingerprints, counts, short necessary excerpts, assessments, and carefully selected screenshots only.
  • Custody. The run may add its report and evidence directory, update the method register's privacy-safe coverage and method-signal rows after final report validation, advance only the bounded-pilot state table below at that same post-validation point, and create /tmp data. It must not edit any other source, tests, docs, dependencies, configuration, existing reports, or unrelated worktree changes. Never overwrite an earlier report or evidence directory.

Bounded method pilots

These instruments are temporary pilots, not permanent coverage claims:

  • On the next naturally occurring new-story run, capture at most one sealed Cold First-View Witness before the main operator inspects the first screen, then stop for an explicit keep, revise, or retire decision.
  • On at most one naturally qualifying assistance prompt per run, collect two byte-identical fresh draws. The initial eligible kind is Segment Reconciliation. The pilot stops for an explicit keep, revise, or retire decision after two qualifying pairs.
  • Before finalizing each of the first two reports that contain decision-driving claims, challenge at most three such claims independently and resolve them as supported, narrowed, contradicted, or insufficient.
  • The privacy-safe method register stops after its third natural-run row for an explicit keep, revise, or retire decision. Never read it before or during the author journey.

If a pilot does not trigger naturally, record it as pending rather than manufacturing eligibility. At a pilot's sunset, do not claim saturation, reliability, or method validity; route the explicit disposition through skill maintenance.

Authoritative pilot state

This table is the sole pre-journey state source. Do not inspect the method register or unrelated reports to reconstruct counts.

Instrument State Completed Sunset
Cold First-View Witness awaiting-disposition 1 1
Paired-Draw Check awaiting-disposition 2 2
Independent Claim Challenge report awaiting-disposition 2 2
Method-register natural-run row awaiting-disposition 3 3

Only after the final report validation passes:

  1. increment Cold First-View when cold_first_view_witnesses is 1;
  2. increment Paired-Draw when paired_draw_checks is 1;
  3. increment Independent Claim Challenge report once when independent_claim_challenges is greater than 0;
  4. increment the method-register row only after that row is appended; and
  5. change any instrument reaching its sunset from active to awaiting-disposition.

No historical run increments the three prospective witness pilots. An awaiting-disposition, kept, revised, or retired instrument does not run again unless its explicit owner disposition changes this table and, when needed, its sunset. Apart from those cells, a playtest invocation must not edit this skill. Reaching a sunset does not interrupt report validation or cleanup: mark the state awaiting-disposition, finish custody, and request the owner's keep, revise, or retire decision in the final response.

Process

1. Prepare the author journey

Read Run setup and Observation scratchpad. Run scripts/prepare-run.mjs, capture the worktree baseline, and fill the scratchpad's story intent, author expectations, sealed mental model, and run plan before launching the app. For a continuation, confirm the supplied report resolves to an existing /tmp project. Done when all run paths exist, the baseline is recorded, and the pre-use expectations are complete.

2. Launch the isolated app and guarded browser

Read Browser driver. Build the app, start safe-app-session.mjs, then start browser-session.mjs before the first navigation. Do not reuse an existing server or browser. Done when the app holder reports its loopback URL, the 1440x900 browser session reports its safety state, and provider-request capture began before navigation. For a pending new-story first-view witness, defer all parent-context screen inspection until Step 3.

3. Enter as a new or returning author

Read Author journey. On a new run, capture the sealed first-view witness when its authoritative state is active, then independently assess the unopened project screen before creating the planned /tmp project. On a continuation, open the supplied project, visit Accepted Segments, and establish the latest accepted sequence without copying accepted prose into prompt-facing fields. Done when the scratchpad records what the UI confirmed, corrected, or left unclear about the initial mental model and continuation state.

4. Author the next local segment

Follow the authoring loop in Author journey: configure the story, create and curate records, use Private Notes only as inert scratch, curate the active working set and cast functions, fill a natural Generation Brief, and resolve readiness through the UI. Record the intended influence of every deliberately populated Generation Brief field before compiling. Done when readiness permits prompt inspection or the blocker policy terminates the run.

5. Evaluate prompts without OpenRouter

Read Cold prompt evaluation. Extract the visible prose prompt to /tmp, dispatch one fresh cold subagent, assess its untouched response, and allow at most one fresh retry. When a naturally qualifying assistance prompt reaches the active paired-draw pilot, predeclare the useful output classes and collect exactly two fresh byte-identical draws regardless of Draw A's result; assess them separately and do not take a third draw. Use one targeted counterfactual probe only when it can clarify a high-impact ignored field; never use that diagnostic response in the app. Evaluate Ideate, Record Hygiene, and Segment Reconciliation the same way when naturally invoked. Done when every invoked prompt has a structured usefulness verdict and every populated Generation Brief field has an influence verdict or not assessable.

6. Accept one segment and reconcile continuity

Paste the chosen raw prose through Write or paste candidate, make ordinary author edits in the visible editor, and record the intervention burden before accepting exactly one segment. Inspect the durable-change reminder. Use Segment Reconciliation and Record Hygiene when the accepted segment or record set makes them useful; re-author adopted suggestions manually. Done when the accepted sequence is visibly confirmed and records, Generation Brief, and working set represent the author's chosen continuity for a later run, or a post-acceptance blocker is logged.

7. Report and close

The source-and-doc-blind product journey ends before closeout. Before drafting any repository artifact, read root AGENTS.md and docs/ACTIVE-DOCS.md, then follow only the post-journey authorities they select for those writes. Do not use those reads to revise the sealed mental model, contemporaneous observations, product findings, or evidence tags.

Read Blockers and diagnostics when any probable blocker or visible defect appears, then read Report format. Consolidate the scratchpad into a new schema-v2 cumulative report. Before publication, independently challenge up to three eligible decision-driving claims when that pilot state is active, and let the main operator retain resolution authority. Close browser and app sessions, remove session plumbing and uncited evidence, then validate with scripts/validate-report.mjs. Keep the scratchpad and temporary exchange files until the report passes; then delete them, rerun the validator, and compare final worktree status with the baseline. Only after that final pass, open and update reports/playtest-method-register.md when its pilot is still active, then advance the report-derived authoritative pilot-state cells above. Advance the method-register counter only when its row was successfully appended. Compare final worktree status with the baseline again, allowing only the declared custody deltas. On successful close, remove the exact run root created under /tmp/continuity-loom-playtest/ but preserve the separate /tmp story project. Done only when the report validator passes, no OpenRouter request was attempted, run-owned processes are stopped, and the final response names the report and custody result.

04 同仓库更多 Skills