bladeforge

Stack-agnostic pre-push review framework — eight reviewer agents (five on by default, plus skill, leak and security-scan for repos that publish skills), the /review orchestrator, and the secret-scan + attestation gate, all driven by a per-project .claude/review.config.json.

2 skills · 2 commands · 8 agents

Meta skills — authoring new skills, the OCKHAM minimal-entities discipline, and the WITTGENSTEIN clarity gate for specs & plans.

10 skills

Frontend Vue 3 skills — single-file components with Composition API and TypeScript, component architecture, BEM in scoped styles, Pinia stores, VueUse-first composables, Nuxt conventions, motion, testing.

9 skills

Frontend React skills — component structure, hooks registry, skeleton components, Storybook stories.

8 skills

Personal Salesforce skills — Apex test authoring standard and salesforce-dx MCP usage.

5 skills

Whose Turn, the Claude Code side of a local monitor of your Jira, GitHub, Linear and other sources: whose turn it is, what changed, your own to-dos, a summary at the start of each session, a hand-off from the monitor window into the session, a /whose-turn command that opens the window, and an MCP server for any agent. The monitor itself is the published whose-turn npm package; the plugin installs its latest version into the monitor's data directory on first start and runs it from there; each newer version is then announced in its window and at session start and installs only when the user says yes or clicks Update. Employer-specific values live in a workspace config on the machine, not in the plugin.

1 skill · 1 command · 1 MCP server · hooks

CERBERUS — the two-headed guard at the gate for a PUBLIC skills marketplace. A PostToolUse hook fires on every skill / reference / script / eval / manifest edit and reminds you to run BOTH heads before committing: cerberus:leak-check (outward — no real work identifier or secret escapes into the public marketplace) and cerberus:security-scan (inward — the eight-point consumer-safety checklist: prompt injection, exfiltration, secrets, dangerous commands, obfuscation, external fetches, credential access, privilege escalation, so nothing unsafe reaches whoever installs the skill). Before a push it also guards the commits themselves — their messages, authors and branch name, which no file check reads: a classic PreToolUse hook has a nested claude judge the outgoing commits and blocks a push that leaks or that it cannot judge, and on Claude Code 2.1.292 or newer a push-guard mod judges them in-process first; an older Claude Code gets a red notice at session start to update. The verdict cache is guarded best-effort, so the hook stops a forgotten check, not a deliberate forgery. Both are agent judgment passes, no denylist by design — a list of the real names to catch would itself be the leak; the intelligence lives in the agent, not a regex.

2 skills · hooks

Frontend CSS/SCSS skills — rem units and SCSS-module structure, scoping & naming.

3 skills

Tells you and Claude, at session start, what is slowing the machine: orphaned agent and test processes, sustained memory pressure and swap growth, low disk, a runaway fseventsd. Each finding comes with one suggested step that Claude offers and runs only on your yes; the plugin itself executes nothing. Ships a status-line command your own status-line script can call. macOS first; Linux and Windows builds are silent until measured.

2 commands · hooks

Passive skill that reads the compiled catalog to discover/recommend/install marketplace skills, surfaces declared side effects, and treats catalog text as untrusted data.

2 skills · hooks

Test standard as an installable plugin — the execution-tier split, the test contract, the axes a case space is derived from, factories, matchers and fakes, a JSON failure envelope, and the coverage and mutation gates. Ships a whole-tree audit agent and the config a repository states its own numbers in.

2 skills · 1 agent

A portable documentation standard shipped as installable artifacts — a rules skill, reference templates, a read-only audit agent, and one deterministic pre-push check that blocks.

1 skill · 1 agent

Error skills — the unified error envelope + reason-code vocabulary (format) and the framework-agnostic client-side error-handling architecture (architecture).

2 skills

CICERO — the house voice. The numbered communication rules under one governing readability rule (result first, plain words, recommend one option, push back with reasons, honesty about verification, stay in scope) ship as a force-for-plugin output style applied at the system-prompt level. Four floor rules are invariants; every other rule is a user-overridable default. 2.3.0 inverts the shape default: prose unless a trigger fires, with a checkable situation-to-shape map, a ban on two adjacent blocks sharing a shape, a measured device inventory in which bold is no longer load-bearing, and a bar-and-title heading. The honesty floor now forbids ending a turn on a promise — a turn ends DONE, BLOCKED, or CONTINUING-and-continuing — because asking an agent to narrate its progress is itself a documented cause of premature stopping. 2.5.0 adds an unambiguity section borrowed from ASD-STE100: one term for one thing inside a document, a noun cluster of at most three words, the condition stated before the instruction, and a ban on dropping a word to hit a length target. That standard's approved-word dictionary and its ban on -ing forms are deliberately not adopted. A SessionStart hook shows a banner with a worked example of the finding-tree notation and picks the conversation language; a UserPromptSubmit hook keeps the reply language current.

hooks

Adversarial critique — the house method for red-teaming a design, spec, or plan with diverse independent lenses so a critique detects real defects instead of laundering a green stamp.

1 skill

Architecture / flow diagram authoring — turn a spec or raw code into a readable, clickable diagram (classes with methods, objects, permission sets, relations) rendered as one browsable HTML page under docs/diagrams/<name>/. Builds a <name>.diagram.json model laid out with React Flow + ELK (elkjs), hardened by parallel Sextant reviewer agents over a capped review loop.

1 skill

Feature briefing films in Claude Motion: turn a feature's spec, plan, code or ticket into one honest presentation film that opens on the problem, carries one running example through the feature, shows what is proven and what is still open, and invites the viewer to answer the author. film:feature owns the storyboard (chapters, stakes, showcase, finale, honesty, on-screen copy rules, a read-only Braintrust review), builds and checks the scenes locally, and publishes the finished film to Claude Motion.

1 skill

Personal frontend dev-harness skills — types + tests runner.

1 skill

Personal JavaScript/TypeScript style conventions.

1 skill

Personal Git workflow skills — atomic commit splitting.

1 skill

Personal i18n skills.

1 skill

Personal jira skills.

1 skill

A comment is a pointer, a directive, a three-line doc block on an export, a member of one or the module itself, or one another house standard requires — never an essay. One language-agnostic rule with a per-language syntax table, enforced by the review gate's craft lens.

1 skill

Spec+Plan+Branch gate — a PreToolUse hook that blocks Edit/Write to code unless you are off the base branch (default `main`) AND a plan exists for the work: a branch task-id (e.g. PROJ-1234) must match a plan filename, else any branch-new or uncommitted plan file suffices. docs/ is always writable so the spec/plan itself is never blocked. Opt-in per repo via .claude/plan-gate.config.json; bypass one edit with SKIP_PLAN_GATE=1.

hooks

Workflow script for the pre-push review gate — dispatches the review plugin's lenses in parallel with a forced response schema, reconciles findings, recomputes each lens's score with the attempt-aware formula, and checks the eight gate criteria before the review command may attest. No skill; invoked directly by the review plugin's /review command.