Claude Skill

forge

Drive a software or general-work outcome through Forge's composable Spec, Plan, Build, Acceptance, and Ship lifecycle. Use when the user explicitly asks to use Forge, asks Forge to explore, spec, plan, build, review, accept, verify, simplify, finish, ship, reconcile a Spec change

LLM Mart · 0 points · 1 views 53 listing impressions 0 install-command copies
Download brightstack-forge-skills_forge-ba47c05.zip · 2 KB

Install

skills CLI npx skills add https://github.com/brightstack/forge/tree/main/skills/forge
Claude Code claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install brightstack-forge@llmmart
Git git clone https://github.com/brightstack/forge.git

The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole brightstack/forge collection as a plugin from our marketplace. Git is the plain clone.

Skill manifest

Forge

Files (forge)
  • SKILL.md 6.4 KB
    ---
    name: forge
    description: "Route software delivery and related Forge operations when the user explicitly invokes Forge or the caller selects it. Use for Forge this, Forge build/fix/review/QA, or Forge delivery requests. Select the named phase and stop at the requested boundary. Ordinary unselected requests do not activate portable Forge; generic routing never grants Auto or publication."
    ---
    
    # Forge
    
    <skill name="forge">
    <activation>
    Activate only when the user explicitly invokes Forge or the caller identifies
    this skill. Do not infer portable Forge from an ordinary unselected build, review,
    plan, research, or writing request. A target harness may select Forge for those
    requests. Re-evaluate the current request; loading Forge does not start a workflow. Route to
    the smallest requested operation. A ticket, plan, mock, or reference alone does
    not authorize Build. Actual host commands such as `/forge:spec` delegate to their
    named phase; valid skill folder names do not by themselves prove slash support.
    Use the installed host documentation for verified invocation and gaps.
    </activation>
    
    <setup>
    Run `forge context --engine <actual-engine>` from where you started, without `--repo`
    and without changing directory; read every file it lists, follow its printed rule,
    and use the Target it prints as `--repo` for later forge commands.
    If `forge` is missing or doesn't know `context`, follow the fallback in
    [target harness resolution](../../references/target-harness.md#resolve-applicable-authority-progressively).
    Read
    [authority](../../references/authority.md) and
    [target harness resolution](../../references/target-harness.md), then the
    actual applicable repository sources they route to. The Coordinator owns conversation, authority,
    Launch, routing, and owner inspection. Load only the selected phase's professional
    instructions. Identity belongs to those professionals; procedure belongs here.
    </setup>
    
    <launch>
    Show or reuse [Launch](../../references/workflows.md). For a new Launch,
    read and fill the shared [chat template](../../templates/launch.md): the
    task and stopping point, a compact Flow/Team table, and the actual approval ask or
    authorized next action in plain language. Keep routing taxonomy internal. Guided
    waits before dispatch and at new Spec/consequential Plan gates; an existing accepted
    Launch is reused. Auto proceeds only within its cited grant, retaining protected
    Spec and publication restrictions. A clear direct bounded Build request supplies
    Build authority; faithful tactical Plan notes create no second approval gate.
    The shared direct read-only Review exception also applies: a bounded review
    request authorizes inspection and its report after a brief scope/team/boundary
    update, without an additional Launch acceptance ask.
    </launch>
    
    <routing>
    - Spec, scope, or spec design: [Forge Spec](../forge-spec/SKILL.md).
    - Plan or architecture strategy: [Forge Plan](../forge-plan/SKILL.md).
    - Build, implement, or fix: [Forge Build](../forge-build/SKILL.md).
    - Integrated review: [Forge Review](../forge-review/SKILL.md).
    - Verify or browser QA: [Forge Verify](../forge-verify/SKILL.md).
    - Ship: [Forge Ship](../forge-ship/SKILL.md).
    - Explicit standalone code-only review: [Forge Code Review](../forge-code-review/SKILL.md).
    - Explore: [bounded research](../../references/research.md).
    - Explain an artifact/change: [read-only Explain](../../references/reporting.md#explain).
    - Simplify: [Forge Simplify](../forge-simplify/SKILL.md), a bounded post-Review operation.
    - Finish: [actual-completion reconciliation](../../references/finish.md).
    - Spec apply, legacy natural wording Spec merge, or KB add/update/remove/verify/history:
      [guarded memory operations](../../references/memory.md).
    - Evaluate Forge behavior: use the separately supplied `forge-eval` development
      skill. Its evaluation corpus is not part of a runtime-only installation.
    </routing>
    
    <workflow>
    <step n="1" name="Establish scope and authority">
    Identify the requested boundary, accepted sources and revision, actual state, and
    remaining material choices. Follow the [workflow assignments and gates](../../references/workflows.md).
    Use [record mechanics](../../references/protocol.md) only as needed.
    </step>
    <step n="2" name="Run only authorized phases">
    Full delivery follows `Spec -> [Build <-> Review] -> Simplify -> Verify -> Ship`.
    Reuse relevant accepted artifacts/evidence explicitly. Plan remains a job before
    Build, with brief Engineer notes for Quick work and a direct Plan skill when
    requested. Spec owns discovery; Build/Review converges on a working reviewed
    candidate. A separate Engineer then owns Simplify; changes receive affected proof
    and independent delta Review bound to the resulting complete candidate, while an
    evidenced no-op retains prior Review. Verify exercises that result. Ship performs only authorized publication. Direct Build/Review
    stops before Simplify unless the request also authorizes it; direct Simplify stops
    before QA.
    </step>
    <step n="3" name="Inspect handoffs and repair">
    Use one accountable owner and disjoint or serialized writes. Required specialists
    are actual native assignments, not personas loaded into the Coordinator. Read each
    return and new relevant log evidence under [the return contract](../../references/returns.md)
    before advancing. Apply shared severity, remedy authority, and finite repair limits.
    </step>
    <step n="4" name="Close the requested work">
    At actual authorized completion, run Finish for approved unapplied meaning and
    owned-process cleanup. Direct Spec/Plan stays proposal-only. State actual outcome,
    candidate, proof, gaps, and next required action; no extra lifecycle or publication
    claim follows from a local PASS.
    </step>
    </workflow>
    
    <output>
    Use the selected phase's owning artifact template and shared returns. Launch uses
    the shared [chat template](../../templates/launch.md), not a saved report.
    Other human gates use [brief reporting](../../references/reporting.md), with a plain-language
    decision brief followed by links to the exact revision.
    </output>
    <checklist>
    - Explicit current activation and requested stopping boundary.
    - Launch and authority reused or shown; human approval and delegated progression distinct.
    - Actual phase professionals, independent Review, and distinct Verify as required.
    - Returned artifacts and new relevant log evidence inspected before advancement.
    - Findings retain scope; shared severity, finite repair, and proof limits respected.
    - Completion records actual apply/cleanup outcomes within existing authority.
    </checklist>
    </skill>
    

Comments (0)

Sign in to join the conversation.

No comments yet.

Reviews (0)

No reviews yet.