claudelicious

Takes any external artifact you link (a repo, a paper, a tool, a blog post, a thread) and decides what to lift from it into your own stack. Resolves the artifact, maps it against what you already have (your notes, skills, rules, memory), extracts only the genuine deltas, red-teams each delta against your real setup, and returns a per-delta decision: borrow, build, install, or skip. ALWAYS invoke when the user shares a link and asks "should we adopt this", "what do we take from this", "what do we lift from X", or "is this worth stealing". NOT a review of the artifact on its own merits (that is a different, grade-it-in-isolation skill); this one grades the artifact against YOUR systems and outputs a lift plan.

Create a well-formatted git commit using conventional-commit format. ALWAYS invoke when the user says "commit", "commit this", "make a commit", or asks to save work to git. NOT for pushing (that has its own gitleaks gate) and NOT for staging review.

Convenes a council of orthogonal thinker-lenses (Munger, Taleb, Thiel, plus idea-lenses base-rates and second-order-effects, plus a functional skeptic seat) to pressure-test a decision from multiple perspectives, force disagreement, then synthesize. Stateless three-phase engine: blind parallel reads, a dissent round against the consensus, then a synthesis (where they agree, sharpest live dissent, the lens you ignored, decision-shaped close). Invoke when the user says 'convene the council', 'council this', 'pressure-test this decision', or wants multiple mental-model perspectives on a decision before committing. Seats fan out in parallel; the main session synthesizes.

Build or improve any frontend UI (component, page, app, dashboard, landing, marketing site) so it does not look like generic AI output. Enforces a brief-read, three dials, and a countable pre-flight gate instead of vague 'be distinctive' advice. Invoke when the user asks to build/create/redesign/improve a web UI, a React/Vue/Svelte component, a page, a dashboard, or an app, or says it 'looks generic / boring / AI-slop / templated'. Pulls a NAMED design style only when the user names one; otherwise reads the brief and picks a fitting direction. Does NOT auto-clone any template and does NOT default to any one font.

Consolidates memory across all layers (the durable memory files, the learnings log, project continuity) by pruning stale entries, merging duplicates, and resolving contradictions. Invoke when the user says 'dream', 'consolidate memory', 'clean up memory', 'merge duplicates', or after a heavy session that produced many similar entries. Different from per-correction capture (memory and the learning loop write new entries; dream cleans the corpus). Run periodically when memory feels noisy or after a multi-day push.

Interrogate the user about a plan or design one question at a time, walking the decision tree branch by branch and recommending an answer for each, until every decision is resolved and understanding is shared. User-invoked only via explicit phrases like 'grill me', 'interrogate this plan', 'stress-test this design', or '/grill-me'. Never auto-fires. Distinct from plan mode (explores and proposes) and a premortem (runs after a rough decision); grill-me runs before a plan solidifies, to extract and resolve the open decisions.

Dispatch an already-approved status update to a channel (chat, social, a webhook) via a CLI. ALWAYS invoke when the user says "post the update", "send the update", "broadcast this", after they have approved the text. This is a DISPATCH point, not an author — it never writes the content itself.

Run a multi-source research sweep on a topic and synthesize a cited summary. ALWAYS invoke when the user says "research X", "sweep X", "what's being said about X", or asks for a grounded summary across sources before drafting something. NOT for a single-source lookup (just fetch that source).

Logs corrections, errors, bugs, feature requests, and learnings into a learnings log, then queues a promotion target (a skill's SKILL.md, an agent's SOUL/SCOPE, a rules file) so a future session applies the fix at the source. Invoke when the user (1) corrects behavior or output ('no', 'wrong', 'actually', 'don't do that', 'next time'), (2) reports a bug or unexpected output, (3) requests a capability that does not exist yet, (4) flags an outdated assumption, (5) confirms a non-obvious approach worked, or (6) asks to promote a learning. Runs in PARALLEL with the memory system: memory writes the next-session sticky note, this skill writes the structured entry that edits the broken skill/agent/rule. Both fire on the same trigger.