overclock

Stop losing working state. Warm resume with verified full-SHA drift anchors, durable workflow lessons, and a ledger of verified fixes share a confined, atomic memory helper. The startup hook reports parked state without injecting repository-controlled prose; writes never auto-commit.

3 skills · hooks

Get clear-eyed analysis instead of agreeable answers. Tests framing and settled decisions against evidence, alternatives, and risk; delegates material local-fact questions to a host-isolated read-only context with exact roots, hard source budgets, provenance, and prompt-injection defenses, or fails closed when isolation is unavailable.

2 skills

Pre-action engineering gates backed by real evidence: leave red or characterization artifacts uncommitted before edits, mutation-check selected behavior through an integrity-checked transactional restore, and recover defensive-code intent from untrusted history plus current-state proof before weakening it.

2 skills

Self-host a local, governed evaluation stack for agent skills: ironside trace store, rubrist judged evaluation with human adjudication, Claude Code and Codex session importers, a pi tracer that tags skill usage, casefile scanning, and a findings-driven skill-maintenance loop with golden-gate validation and human merges.

2 skills

A self-improvement loop for your agent. Corrections to agent behavior and standing workflow rules become durable, evidence-counted lessons in .ai/memory/LESSONS.md through a confined atomic helper. Its startup hook reports ledger availability without injecting file-controlled text. Standalone alternative to session-memory; do not install both.

1 skill · hooks

Work through GitHub PR feedback in two explicit stages. The resolver fetches and centrally judges comments, applies authorized local fixes, and drafts replies but cannot mutate GitHub. A manual-only publisher can post an exact digest-locked plan after complete host, PR, head, source, and action preflight. Never commits, pushes, merges, rebases, or approves.

2 skills

Experimental pull-request review kit: a repository-agnostic adversarial reviewer plus an explicit initializer that creates a source-grounded repository profile.

2 skills

Use another installed coding harness as a bounded leaf collaborator. Consult read-only or delegate implementation into an isolated clean clone, then inspect and digest-validate the scoped patch before applying it to an unchanged active checkout. Never auto-commits, pushes, or silently substitutes providers.

1 skill

Experimental clean-room pre-review analysis that turns change intent and a pinned pre-change repository snapshot into a bounded edge-case risk brief without inspecting implementation code.

1 skill

Dry-run how an application would call the Claude Messages API or an OpenAI-compatible endpoint before writing the app. One JSON spec carries the system prompt, seed messages, tool schemas, stubbed tool results, and hard request, token, and dollar caps; the helper runs the tool-use loop, records the exact transcript, and compares variants. Credentials stay in the environment, mock runs need no network, and nothing is committed or edited.

1 skill

Systematic diagnosis for bugs that resist the ordinary red-test path. Establishes the safest useful observation loop, fails soft when reproduction is unavailable, minimizes the failure, and tests 1-5 credible hypotheses with predictions and proportionate noisy-evidence rules. Diagnosis is read-only unless implementation is also authorized.

1 skill

Requirements elicitation as a discipline. Asks one material decision at a time, distinguishes facts from uncertain evidence, labels decision provenance, and refuses to hide high-impact choices inside delegated defaults. Produces a confirmed decision brief and stops; implementation and planning remain separate.

1 skill

Generate non-obvious engineering approaches by exposing assumptions, applying lateral moves, and prosecuting conventional candidates before ranking grounded alternatives with explicit costs.

1 skill

Draft and revise multi-paragraph prose in the author's established voice. With no supplied voice, uses a plainspoken fallback house style while preserving factual claims, uncertainty, caveats, quotations, and deliberate punctuation. Optional visual revision reports are reconstruction-verified and explicit opt-in.

1 skill

Choose Overclock plugins without guesswork. A manual, report-only advisor that inventories installed packages and project instruction metadata, detects standalone overlap, blocks incompatible combinations, and returns exact proposed commands and minimal instruction diffs without executing or writing anything. User-level instruction inspection is explicit opt-in.

1 skill

Keep one ubiquitous domain language per project. Applies CONCEPTS.md in conversation, proposes exact changes when routing fires implicitly, and writes only on an explicit add/update request or approval. A confined atomic helper protects the root glossary; workflow corrections stay with lessons-learned.

1 skill

Experimental target-only consumer-contract audit skill.

1 skill

Experimental broad semantic contract-gap audit skill.

1 skill

Experimental semantic contract-gap audit with deterministic evidence assembly.

1 skill

Experimental contract-gap audit with deterministic multi-line evidence assembly.

1 skill

Experimental contract-gap audit with strict intended-line evidence assembly.

1 skill

Experimental contract-gap audit with length-bounded intended-line assembly.

1 skill

Internal, unshipped conversation-eval authoring prototype.

1 skill

Turn a sprawling exploratory repository back into one project. A read-only survey finds the spine, sorts the rest into threads, records the user's keep/park/delete decisions, and writes one plan file with an ordered checklist, a bounded foundations check, and a suggested next move. An explicit apply mode works that checklist one verified item at a time. Never commits.

1 skill