Tsurai7
- 17 skills
- 0 followers
- 7 hours ago last updated
- ▌ Create Pr · tsurai7Create or update GitHub pull requests using gh. Use when the user asks for a PR, or to open/update a pull request after branch work is ready.
- ▌ Git Safety · tsurai7Safe git commit workflow. Use when the user asks to create a commit, or before running git commit. Covers staging, message style, HEREDOC format, amend rules, and hook failures.
- ▌ Plan Review · tsurai7Engineering review of an implementation plan before any code is written: scope challenge first, then architecture, code quality, tests, performance. Use when asked to review/critique a plan, design doc, or proposed approach, or before approving a multi-file implementation plan. Skip for reviewing a code diff (staff-review), for deciding whether the idea is worth building at all (office-hours), and for single-step plans with an obvious shape.
- ▌ Office Hours · tsurai7Pressure-test an idea before building it - premise challenge, status quo, evidence of value, narrowest wedge, risk, cost of keeping. Use when asked to vet/scope a proposal, decide whether something is worth adding ("стоит ли внедрять X"), or before writing an implementation plan for a new capability. Skip for bug fixes with an obvious cause and for work already scoped by an accepted plan.
- ▌ Staff Review · tsurai7Staff-level code review method for a diff or branch: review against stated intent, confidence-calibrated findings, verify every finding against the code before reporting, fix-first protocol. Use when asked to review code like a senior/staff engineer, or when review findings must be high-precision. This is the review method — it complements harness-native review commands (e.g. /code-review), not replaces them. Skip for reviewing plans (plan-review) or vetting ideas (office-hours).
- ▌ Answer Shapes · tsurai7Optional output skeletons for common coding deliverables — bug fix, code review, investigation. Apply when the deliverable is one of these shapes to keep the answer complete and parseable (root cause / fix / verification, etc.); especially helps weaker models and machine-consumed handoffs. Skip for freeform questions.
- ▌ Review Ratchet · tsurai7Convert every human review finding or incident into a permanent automatic check so the error class cannot recur. Use after receiving human code-review feedback, after a retrospective, or when the same finding class appears a second time. Skip while findings are still being triaged (findings-ledger comes first).
- ▌ Findings Ledger · tsurai7Track every review finding to an explicit, recorded disposition so nothing is lost in conversational triage. Use when triaging code-review or audit findings (agent or human), before merging a branch that received review feedback, or when deferring or rejecting a finding. Skip when a review produced zero findings.
- ▌ Invariant Audit · tsurai7Adversarially sweep a repo's standing invariants (billing/idempotency, PII, published contracts, deploy order) across every execution path a diff touches — retry, redelivery, timer, error, concurrency, rollout — not just the happy path. Use before merging changes that touch money, external calls, PII, published contracts, or workflow/state-machine definitions. Skip for docs-only diffs and refactors that add no execution paths.
- ▌ Claim Verification · tsurai7Extract every checkable claim from a PR description, code comment, doc, or review reply and verify each against the actual tree before trusting or publishing it. Use when reviewing someone else's PR, when writing your own PR description or review replies, or when acting on assurances found in docs or comments ("a test locks this", "matches production"). Skip for claims whose verification you watched in this session.
- ▌ Model Orchestration · tsurai7 bundleRoute each sub-task to the cheapest model tier that can do it well, and delegate execution to subagents. Apply when planning a multi-step or token-expensive task, choosing between strong/mid/cheap models (Opus/Sonnet/Haiku), or deciding whether to spawn subagents. Plan and decide with a strong model; execute well-specified sub-tasks with a cheap one; escalate on failure. Skip for a single trivial step - orchestration has overhead. Prompt/cache token costs of an LLM system belong to prompt-caching-playbook.
- ▌ Spec Fidelity Review · tsurai7Review a diff against what was asked — ticket, acceptance criteria, PDR/ADR, product requirements, PR description — instead of against the code. Use when asked whether a change matches its spec/ticket/requirements, or as the spec-fidelity lens of a multi-lens review. Skip when no spec source exists (report that instead of inventing one) and for code-quality review (staff-review) or invariant sweeps (invariant-audit).
- ▌ Systematic Debugging · tsurai7Systematic root-cause debugging: investigate before fixing, log hypotheses, trace data flow, and stop to reassess after 3 failed fix attempts. Use when debugging a non-obvious failure, when a fix attempt did not work, or when asked to find a root cause. Skip for trivial errors whose cause is stated in the error message.
- ▌ Test Adequacy Review · tsurai7Judge whether the tests in a diff can actually catch what they claim to guard: can each test fail for its named reason, are all trigger paths exercised, are barriers observable rather than sleeps, do fixtures share contracts with production. Use when asked whether tests are adequate/any good, or as the test lens of a multi-lens review. Skip for writing new tests (that is implementation work) and for reviewing non-test code.
- ▌ Lossless Doc Compress · tsurai7Compress a document "without losing information": remove only provable redundancy (filler, hedging, LLM slop, restatement, verbose phrasing), never a fact, number, decision, or caveat; flag every judgment call instead of cutting it; account for every removed word. Use when asked to compress, tighten, shorten, condense, or de-slop a doc, note, PRD, RFC, or report — including "сожми без потери информации", "compress as much as possible without losing any information". Skip for summaries, abstracts, or TL;DRs where loss is acceptable — this skill's contract is that nothing is lost.
- ▌ Prompt Caching Playbook · tsurai7 bundleMaximize prompt-cache hit rate and cut token cost on LLM/agentic systems (Anthropic, OpenAI, Gemini, vLLM). Use when designing, auditing, or debugging the prompt/message structure of an LLM system, or when asked about cache hit rate, cache_control, prompt_cache_key, context caching, cache TTL, KV-cache, prefix matching, static-prefix/dynamic-suffix, "why is my LLM bill so high", "почему API LLM дорого", "как уменьшить токены", or reusing prompt prefixes across requests.
- ▌ Verification Before Completion · tsurai7Require evidence before claiming work is complete, fixed, or tests pass. Use before commits, PRs, or any "done", "fixed", "works now", "should work", or "tests pass" statement.