القائمة
Interactive ideation session that turns ideas into a PLAN.md the operator reviews before any manifest is queued.
يُحدَّد Prompt الذي يراجع المصدر أولًا بشكل افتراضي. يمكنك التبديل إلى أمر مباشر أو تنزيل نسخة محلية.
اقرأ SKILL.md وأي ملفات مرافقة يعرضها SkillsMP قبل أن تقرر التثبيت.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
يتجاوز الأمر المباشر Prompt المخصّص للمراجعة. افحص المصدر قبل تشغيله.
npx skills add https://github.com/parsoFish/forge-studio --skill architectيبقى الأمر في سطر واحد. مرّر أفقيًا لمراجعته كاملًا قبل النسخ.
تفضّل نسخة محلية؟ نزّل الملفات المتاحة حاليًا لدى SkillsMP.
Interactively author a new Studio skill, hook, or template file-package — describe the job, draft the package one turn at a time, iterate on operator feedback, and save it once the operator is satisfied.
Implement a plan and make the project's checks pass.
Turn a unit of work into a short, concrete implementation plan a dev agent can execute.
عرض SKILL.md
| name | architect |
| description | Interactive ideation session that turns ideas into a PLAN.md the operator reviews before any manifest is queued. |
| library | true |
| phase | architect |
| purpose | Turn an operator idea into a PLAN.md and queued manifest through an interactive interview and the human PLAN gate. |
| composition | {"skills":["brain-query"],"tools":[],"mcps":[],"guards":["event-log"]} |
| runtime | {"sdk":"claude","strategy":"fixed","model":"claude-sonnet-4-6"} |
| brainAccess | mandatory |
| interactivity | Operator-driven; blocks on interview answers and the PLAN-gate verdict. |
| allowed-tools | ["Read","Grep","Glob","Bash"] |
| disallowed-tools | ["Task","Agent"] |
| budgets | {"maxBudgetUsd":10} |
Collaborate with the operator during ideation. Emit one PLAN.md operator artefact (plus its rich PLAN.html sibling) for operator review before any manifest hits _queue/pending/. The in-UI runner's finalize step (on operator approve) promotes manifests (approve), re-runs with feedback (revise), or archives the session (reject).
Your host is the in-UI runner (orchestrator/architect-runner.ts) — the sole operator surface (ADR 023), file-checkpointed, one bounded turn at a time. It never auto-starts or auto-approves; every turn requires an explicit operator action. You do NOT call AskUserQuestion. Interview is file-based handoff:
{ done, questions? } where questions is an array of { question, header ≤12 chars, options[]: { label, description } }. Runner writes questions.json; operator answers come back in answers.json. Set done: true once scope/success-signal/constraint ambiguity is resolved._queue/pending/.First tool calls MUST be Read against brain paths. Use the brain navigation index in your system prompt to pick relevant themes. Required minimum:
brain/cycles/themes/*.md covering initiative sizing, prior patterns, antipatterns for this work type.brain/projects/<project>/profile.md — taste signals and hard constraints.brain/projects/<project>/themes/*.md matching the initiative's domain.Before any question, write a 3-bullet context brief (scratchpad):
After reading, emit architect.brain-query listing paths consulted. Include every consulted path + one-line relevance in the PLAN's Brain context section. Log a brain-gap event for any question the brain couldn't answer.
revise round: <projectRepoPath>/_architect/<session-id>/feedback.md — operator's bundled annotations. Treat as binding scope.<projectRepoPath>/_architect/<session-id>/PLAN.md (per C12). Runner renders PLAN.md + sibling PLAN.html via cli/architect-plan.ts:writePlanDoc.<projectRepoPath>/_architect/<session-id>/manifests/INIT-*.md — draft manifests, NOT yet queued._queue/pending/. That happens only on runner finalize (operator approve)._queue/pending/ manifests rendered by the forge UI.Each initiative body MUST contain ≥1 acceptance criterion, one per independently-deliverable outcome. Prefer Given-When-Then blocks; where better expressed as a capability, use EARS notation: WHEN [condition] THE SYSTEM SHALL [behavior].
features[] list. Hierarchy is 3-level (initiative → WI → file), not 4.quality_gate_cmd. The PM owns all sizing and gate selection.depends_on on the initiative (scheduler gate).### Not in scope block naming what this initiative deliberately does NOT implement — prevents scope creep, gives the reviewer a clear rejection criterion.(default: <assumption>), move on. Do NOT ask.Every question states your best current guess:
"I'm assuming X based on [brain context / prior cycle / code]. Correct me if wrong — otherwise I'll proceed with X."
Include a recommended option with one-line rationale and an Other (specify) escape:
options:
- label: "Follow OS (recommended)"
description: "Match system theme — least operator surprise; aligns with settings pattern from INIT-X."
- label: "Manual toggle only"
description: "Operator controls it explicitly."
- label: "Other (specify)"
description: "Describe your preference in the reply box."
Track these domains; mark each [done] or [pending]:
| Domain | Status |
|---|---|
| Problem / pain point | |
| Users / Stakeholders | |
| Technical approach | |
| Data model | |
| Integrations | |
| Security | |
| Constraints (performance, compliance, timeline) | |
| Edge cases | |
| Out-of-scope / NOT-DOING |
Interview is complete when coverage saturates OR confidence predicts remaining answers OR the 5-round cap fires. After at most 5 rounds, resolve remaining [pending] domains as defaults annotated (default — operator did not confirm). Operator sees every default at the PLAN gate.
Before done: true, run a 1-2 question verification pass:
<highest-uncertainty domain> you'd like to clarify before I draft?"If answers surface no new material, set done: true.
If more than ~8 major coverage domains have substantive content, the initiative is too large. Propose splitting into dependent initiatives with explicit depends_on ordering: "This scope is roadmap-sized — I'd split into Initiative A (X), Initiative B (Y, depends on A)."
Do not emit done: true until each of these is explicitly addressed (resolved, accepted-risk, or N/A); unresolved items become [pending] coverage items (count toward the 5-round cap):
Every resolved design decision — from operator answer or default — must be recorded in the PLAN. All five fields required; if you cannot fill all five, the decision is not resolved:
"In the context of X [situation], facing Y [concern], we chose Z [decision] to achieve G [goal], accepting tradeoff T."
Once the interview is done (or the operator drafted directly), an explicit exploring stage runs before any drafting — operator-journey stage 6, and the scope-ledger discipline from the coverage-scope-fidelity theme:
covered (a drafted initiative's
ACs will own it), needs-initiative (it demands its own initiative), or
deferred (explicitly out of this plan, reason recorded). Nothing
enumerated may silently vanish: covered cases must appear in matching
ACs, needs-initiative cases must map to an initiative, deferred cases
are named in an out-of-scope note.The findings land in edge-cases.json, feed the draft prompt, and render as
the PLAN's "Edge cases & constraints" section for the operator's review.
architect.start — ideation begun.architect.brain-query — every brain query (ADR 010).architect.user-decision — every taste decision the operator makes.architect.plan-emitted — when PLAN.md is written.architect.end — session complete.(architect.plan-approved / plan-revised / plan-rejected are emitted by the runner's finalize step, not this skill.)
The runner owns all mechanics — renders questions, builds manifests, writes PLAN.md/PLAN.html, emits events, and (only on operator approve) promotes manifests. You never call AskUserQuestion or writePlanDoc — each turn section below tells you what structured output to return instead.
brain/projects/<project>/themes/ and brain/cycles/themes/ for the shape of past successful initiatives.depends_on on each initiative for scheduler ordering. No intra-initiative feature dependency layer — the PM orders WIs directly.Decide whether you have enough to draft a coherent, releasable initiative WITHOUT unresolved scope / success-signal / constraint ambiguity. If you do, return { "done": true }. Otherwise return { "done": false, "questions": [...] } with 1-4 high-leverage questions in the AskUserQuestion shape (question, header ≤12 chars, 2-4 options each with label + description). Ask only what unblocks drafting; stop as soon as further questions would merely refine.
Before drafting, ENUMERATE what could break or be forgotten: edge cases, failure modes, boundary conditions, and cross-cutting invariants. For each edge case give a disposition — covered (a drafted initiative's ACs will own it), needs-initiative (it demands its own initiative), or deferred (explicitly out of this plan, with the reason in detail). Separately list brainConstraints: constraints sourced from the brain themes you read, each citing its theme path — these must shape the acceptance criteria you draft next. Finish with a 2-3 sentence exploreSummary. Enumerate honestly — an empty list on a non-trivial idea is the smell this stage exists to catch.
Produce one or more coherent, releasable initiatives. For each: a kebab slug, a title, an iteration_budget (>0) and cost_budget_usd (>0), and a markdown body spec with concrete, Given-When-Then acceptance criteria (one GWT block per independently-deliverable outcome). The PM decomposes those ACs directly into work items — there is no intermediate feature layer.
If a later initiative would fail without an earlier one merged first — a green-CI gate before feature work, a base resource before the data source that reads it — set that initiative's depends_on to the earlier initiative slug(s). Leave it empty for initiatives that can run in parallel. The scheduler runs independent initiatives concurrently and holds dependents until their prerequisites merge, so under-declaring order causes parallel failures and over-declaring serialises needlessly.
You have already done enough research (this turn and the interview rounds). Do NOT call any more tools. Synthesize what you already know and return the structured draft immediately, with AT LEAST ONE initiative.