lucidpaths/claude.md-workflow/scalpel
Claude Code and agent skill by lucidpaths/claude.md-workflow
- community
Checking that the SKILL.md file is still in the repository.
Deep recon first, minimal cut second - determine the real mechanism, frequency, and blast radius of a problem from the live system, then ship the smallest change at the choke point, with proof proportional to the risk and machinery proportional to the evidence. Use when the user runs /scalpel, or hands over a fix/incident/change where the right SIZE of the solution is part of the question. The skill sizes itself - a question or a one-liner gets a pass of that size.
Install
npx vskill@latest install lucidpaths/claude.md-workflow/scalpelSKILL.md
The same SKILL.md was published earlier as lucidpaths/claude.md-workflow/scalpel.
Scalpel - Full Imaging, Smallest Incision
Two phases: recon until the problem is determined, then the minimal cut. The core law: machinery scales with evidence of need, never with the gravity of the invocation. Being invoked ceremonially does not entitle the work to be ceremonial.
Phase 0 - Size the invocation itself (do this in your head, in seconds)
Before anything: is this a question (it asks for an answer, not a change), a one-liner (the change fits in a diff of at most DEFAULT one_liner_max_lines = 5 lines), or a system change (anything larger)? Answer questions as questions. If the honest response is three sentences or a diff within one_liner_max_lines, say so and do exactly that - invoking this skill on a question or a one-liner and responding beyond that size is the first failure it exists to prevent. State your sizing in one line and move on.
Phase 1 - Recon to determinism (never optional, never wasted)
Understand before designing. Read the LIVE system, not your memory of it:
- Mechanism: what actually happens, step by step, with evidence (file:line, config,
log, census). A wrong mechanism model produces a correct-looking fix for a different problem.
- Frequency and blast radius: how often has this actually fired, and what did it
actually cost? An observed once-a-month self-evident zero-loss failure and a silent daily corrupter deserve different machinery. Count real occurrences; don't imagine ones.
- Existing machinery: what already runs, schedules, alerts, enforces? Inventory the
choke points - the places everything already passes through - and look for the fix INSIDE something that already exists before designing anything new.
- Standing laws: postmortems, forbidden operations, conventions of the repo/host you
are about to touch. A fix that violates a paid-for law is a regression.
- Stop condition: before reading, write down the questions whose answers can change
the fix; the recon ends when each has its answer. Reading past that is spend, not recon.
Why: recon spend is recovered when it answers one of those questions - it is what makes the small fix safe. Artifact spend is not. If a big design gets killed later, the recon it triggered was still the value.
Phase 2 - The cut
- Choke point first: change the one place everything already flows through, not N
callers; edit the existing file, not a sibling system; delete rather than add. Take the second option of a pair only when the first cannot interrupt the Phase 1 mechanism.
- Smallest change that removes the failure mode: not the smallest that looks like a
fix - the mechanism from Phase 1 must be provably interrupted.
- Proof proportional to risk, discipline over artifact: prove the actual change -
replay the real incident against the edited code, feed the checker a known-bad case, show the invariant. Ship a permanent test artifact only where the repo's own conventions expect one; otherwise the reproducible proof run, documented where the change lives (PR body, commit), IS the deliverable. Proving is non-negotiable; shipping the proving machinery is a separate decision.
- Name what you deliberately did not build, with the reason, where the reviewer will
see it. Restraint that isn't recorded looks like ignorance and gets "fixed" later.
The overengineering tells (check the design against each before building)
- Self-justifying components: a part whose necessity is created by another part,
not by the problem (the lock that only protects against the trigger you added).
- Dead-justification shapes: structure kept after its reason died (the shared
primitive nobody calls anymore). If a design decision changed, re-derive what it was carrying.
- Evaporating packaging: count what disappears if you drop ONE choice. If more than
half of the artifact's lines are packaging for a single optional decision, the decision is the design - question it directly.
- Latency/elegance vanity: optimizing a rare state's speed or a working path's
beauty at real artifact cost. Price improvements against observed frequency.
- Ceremony-shaped output: PRDs, harnesses, roles, and fan-outs produced because the
process was invoked, not because the risk demands them.
Delegation calculus (subagents are machinery too)
Fan out only when it buys something real: genuinely parallel INDEPENDENT implementation lanes, isolation of a read that would take your context use past DEFAULT isolate_read_at = 60 % of your context window, or an adversarial pass on something you wrote yourself. One file's fix does not need an implementer; a question does not need a researcher. When you do fan out: locked designs, disjoint files, you judge everything personally. When a locked design changes, stop every agent still working from the old one BEFORE anything ships, while stopping is cheap - keep the recon, discard the artifacts without sentimentality.
Done means
The mechanism is interrupted and proven; every hunk of the diff either interrupts the mechanism or is required by an existing law; existing laws are honored; what was deliberately not built is recorded with reasons; and one honest sentence states frequency/blast-radius of what remains. If any part of the fix serves no mechanism, occurrence or law that Phase 1 found, stop and re-derive - start with tell 2, a design carrying a dead justification.
Closing block (required) — The close - four lines, printed
NOT BUILT is never empty. If nothing was considered and rejected, Phase 1 did not happen - go back and do it rather than writing "n/a".
Security
- Security scanPassed, 90 of 100Pattern checks on the skill's files.
- Model reviewPassed
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingclaude-starter-kit/.claude/skills/scalpel/SKILL.md
Popularity
- 7-day trend score
- 0
Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/lucidpaths/claude.md-workflow/scalpel)