Using Skills

Using Forge Skills

If there is even a small chance a skill applies, consider it before answering, asking a clarifying question, editing files, or spawning an agent. Skills are the contract layer for this kit; skipping them is how pipeline drift returns.

This router is injected only into top-level sessions by hooks/session_init.mjs (installed as .claude/skills/forge/hooks/session_init.mjs). Sub-agents receive their own lighter context and should follow the handoff they were given.

Dispatch

| Situation | Consider | | --- | --- | | Goal in an unfamiliar area, "what am I missing", unknown unknowns before brainstorm/plan | blindspot | | Vague idea, product shape, options before planning | brainstorm | | Non-trivial feature or change needs an implementation source of truth | plan | | Approved plan or concrete brief needs implementation | cook | | Bug, failing test, regression, stack trace, or broken behavior | fix | | Need read-only repo, git, plan, diagnosis, research, or review evidence | scout | | Finished work needs readiness, merge, release, rollback, or handoff decision | ship | | Review pending code or a diff before merge | code-review | | Preview, explain, ASCII-map, or HTML-showcase an idea/plan before implementation | preview | | New UI, visual direction, typography, or reshaping an existing interface | frontend-design | | Need to create, update, validate, or benchmark a skill | skill-creator | | Stuck after repeated failed attempts, need a new framing, or complex multi-step reasoning needs revisable branches | problem-solving | | Need isolated workspace for feature work, or branch integration choices once implementation is complete | git-workflow | | Test level, regression guard, proof strategy, auth/secrets/PII/trust-boundary risk, production diagnosability, schema/data/API migration safety, or docs/ADR/architecture rationale | engineering-practices | | Version-sensitive framework, library, API, or external source required | source-grounding | | Long session near context pressure or clean phase boundary | strategic-compact |

Agent Mention: @advisor

advisor is an agent (strongest model tier), not a dispatch row. When the user explicitly writes @advisor ..., skip every escalation gate: package a compact evidence bundle (question, relevant artifacts, prior attempts) and spawn the advisor agent directly. When the mention accompanies a skill call (e.g. /brainstorm @advisor ...), run that skill normally and route its heaviest analysis step through the advisor. Without a mention, only the gated escalations inside fix, problem-solving, plan, brainstorm, and code-review may spawn it — never by default.

Pipeline Order & Strictness

The main line is:

blindspot? -> brainstorm -> plan -> cook -> code-review -> ship
                                     ^
                            fix (side loop for bugs)
scout / preview support any stage read-only

Strictness is a gradient along that line, not a constant. Each stage's mode flags (--fast/--auto/--hard) tune strictness within the stage; this table sets the baseline between stages:

| Stage | Strictness | Behavior | | --- | --- | --- | | blindspot, brainstorm | Low — diverge | Explore freely, generate options, loose format. Hard rules: read-only, claims grounded in evidence. | | plan | Medium — converge | Contract and artifacts required for non-trivial work; ceremony scales with risk, not habit. | | cook, fix | High — verify | Scoped edits only; every phase produces runnable verification and evidence. Creativity lives in the how, never in the scope. | | code-review, ship | Highest — gate | Verdicts derive mechanically from evidence tables; no done/ready claim without fresh proof. |

Loosen within a stage only via its mode flags; never loosen a gate stage by skipping its evidence.

Precedence

When multiple dispatch rows match, choose in this order:

1. Safety/security/data-loss risk before workflow convenience.
2. Concrete bug or failing test (fix) before new feature planning.
3. Existing approved plan (cook) before creating a new plan.
4. Need for evidence (scout, support skills) before implementation or ship.
5. Context-pressure mode (strategic-compact) last unless explicitly triggered.

Inventory

Top-level skills in this kit:

• blindspot
• brainstorm
• cook
• fix
• plan
• preview
• scout
• ship
• code-review
• engineering-practices
• frontend-design
• git-workflow
• problem-solving
• skill-creator
• source-grounding
• strategic-compact
• using-skills

Source Notes

The gating pattern is adapted from the superpowers meta-router; the SessionStart injection shape is adapted from the agent-skills session-start hook (external references, not vendored in this repo). This file is rewritten for this Claude Code kit and its actual skill inventory.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: danielleit241
• Source: danielleit241/forge
• License: ISC

Install and usage instructions live in the source repository linked above.