Be ambitious
First strip a leading depth argument: it applies only Depth and the outside
check, whichever case follows; say Altitude and Range were skipped because I
asked for depth only. Then read the remaining text after the command and pick
one case. Work means files, commands or a deliverable you produced; a plan or
proposal you only offered is not work.
- A request not yet started. Raise it, show the raise, and stop. Write no file, run no command and invoke no skill until I answer. Sections 1 to 5.
- Feedback on work you did in this conversation ("too shallow", "not good enough"), or no text when the last thing here is work you did: deepen that work. Section 6.
- No text, otherwise. The request is my previous message. When that message is an aside or a side question while an earlier request of mine still waits on my answer, raise the waiting request. With neither, ask me for one.
1. Size it
If the request is already the right size, say so in one line, say it will be done as asked once I reply, and stop. Right size means raising would not change what gets built: a mechanical edit whose purpose is already complete (a rename, a version bump, a format change). A fix for one reported failure is not right size: the goal behind it is the whole class of that failure.
If you cannot tell what the result is for, ask your questions in one batch, including one about the goal, say nothing starts until I answer, and raise after I do.
2. Raise it
Apply all three axes, or Depth alone under depth, and say which you applied.
Skip one only when it plainly does not fit, and then say which and why.
Altitude. State the goal behind the goal, then one purpose line: "This is for ...". Screen the fit: creative or taste work is a weaker fit (say so, do not refuse), and an unbounded goal needs a bound. Ask whether a precedent shows the bigger goal is reachable.
Depth. Run the deeper-version pass: what a top expert in this domain would add that the first version skips; the rawest material to work from (primary data, the source, real users); the methods and tools, each named concretely; and the outside check below.
Outside check, always. Whatever axes apply, name one check outside the
model that shows the result is right: something executable first (a test,
benchmark, validator, measurement, browser run), then a fresh-context
reviewer, which /discipline:scrutinize-dont-coast runs.
Range. Show a nouns-to-verbs step: turn the thing asked for into what it must do ("a website" becomes "lets an old friend recognize me and write back"). Bring in at least one field beyond the request's own. Rewrite any "act as X" into "do what X does: ". Add this line verbatim: "What tool from another field would change what's possible here?"
3. Proposals, never actions
- A crew of specialists: propose it as a hand-off to
/planning:plan. - A red team: propose it only for irreversible, shared-infrastructure or security work, for me to decide, and scope it to gaps that affect correctness or the stated requirements.
- Missing tools or data: one environment line naming what is missing and a reduced version that works without it. Never refuse for lack of them.
- When the raise is doing the same work faster (the session's or team's
throughput, not the product's runtime): one line pointing at
/performance:go-fasteror/session-flow:orchestrate, and no speed method of your own; raise the rest as usual. - When the raised request leaves several open decisions: one line pointing
at
/planning:interview.
Show every skill pointer as bare text; invoke none.
4. Hard targets
Add a hard or unreasonable target only when I say I opt in. Whenever the request or the raise carries a hard target, always add: "The specification outranks the target. If the target is infeasible, stop and report the shortfall; never edit tests, criteria or the target."
5. Output
Original: <the request as given>
Raised: <the raised request>
Axes applied: <Altitude | Depth | Range>
Check: <the outside check>
Deliver what was accepted, at that scope. The project's own instructions still govern.
Reply "accept", edit the raised version, or keep the original.
My reply is the go-ahead; ask for no second confirmation. Once I answer, this skill ends and the session works the accepted request normally.
6. Deepen work in progress
The goal is already set, so Altitude and Range do not apply; a bigger goal is a reason to pause under step 3 below. Sections 3 and 4 still apply, and the project's own instructions still govern.
- Answer, in a few lines each, for that work: what a top practitioner would do that I did not ask for; the rawest material they would start from instead of what the first pass used; the methods and tools, each named; and the outside check.
- Rewrite the brief: my original request plus my feedback, deepened by those answers. Show it.
- Pause only when the new brief changes the goal, widens the scope, needs a decision, money or access that only I can give, or would do something destructive or hard to undo. Say which, and wait; my reply is the go-ahead.
- Otherwise carry on without waiting. Redo the work to the new brief, starting from the raw material in step 1 and using its methods. Before you report, run the outside check and say what it showed. A fresh-context review stays a proposal under section 3.
A declared species in this plugin: NOT a corrector
The correctors re-anchor a standing discipline. This skill is a one-shot,
user-fired raise or deepen: only the user knows they want more than they
asked for or got, so the model never fires it, and it carries no
discipline-batch tier.
Nearest siblings, for routing: /planning:brainstorm when the user wants
options rather than one raised request, and /discipline:scrutinize-dont-coast
when the question is whether finished work is sound.
Next
/planning:interview, when the raised request leaves several open decisions; it locks them into a brief.
Gotchas
- Keep the outside check even when the request already says "don't cheat". We rely on a structural check, not on wording. Topic: reward hacking under anti-cheating prompts. Pointers: https://arxiv.org/abs/2510.20270 (impossible-task results by prompt variant) and https://code.claude.com/docs/en/best-practices#give-claude-a-check-to-run. As of 2026-10-30. Recheck when that docs section or the paper is revised.
- Scope any red team to correctness and the stated requirements. Topic: padded adversarial findings. Pointer: https://code.claude.com/docs/en/best-practices#review-against-your-spec. As of 2026-10-30. Recheck when that section drops or changes its scoping callout.
- Ask for a specialist's work, never a persona. Topic: expert-persona prompting and accuracy. Pointers: https://arxiv.org/abs/2512.05858 and https://arxiv.org/abs/2508.19764. As of 2026-10-16. Recheck when either paper is revised or a study of persona prompts in coding agents appears.
- Every raised request carries the scope fence. Topic: vendor guidance on overeager scope. Pointer: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-2#say-who-owns-the-code. As of 2026-11-10. Recheck when that section changes how a scope-limiting rule interacts with broader instructions, or when the model it is written for is no longer current.
- A hard target never outranks the specification. The abort line always travels with it. Topic: impossible tasks and specification-first prompts. Pointers: https://arxiv.org/abs/2510.20270 and https://arxiv.org/abs/2610.04793. As of 2026-10-16. Recheck when either paper is revised or replicated.