Size the run
Contract
| Field |
Bound contract |
| Trigger |
User asks how hard a run should be ('how hard should I think about this', 'is this worth a deep run', 'size this task') or a subagent dispatch is about to launch. |
| Authority |
Read-only. No file, VCS, credential, paid, published, deployed, or remote mutation. |
| Side effect |
One advisory recommendation in chat; no config, execution, or model change. |
| Done |
The report names exactly one capability tier plus one effort level (the cheapest sufficient pair), the move-up/move-down triggers, and the proof surface. |
Inputs
- Task description (required): the work being sized: a concrete unit of work, not a vague goal.
- Context (optional): domain constraints, risk factors, or prior sizing history that affect the judgment.
Procedure
- Frame the unit. Name the exact work being sized. Done when: the unit of work is named in one sentence.
- Score once. Assess risk and complexity across ownership boundaries, reversibility and blast radius, safety or privacy risk, ambiguity and synthesis load, need for research or adversarial review, and the cost of a wrong answer. Done when: all six risk dimensions are assessed.
- Read off tier. Choose the cheapest tier whose ceiling covers the judgment and risk. Risk beats size: one high-risk file can want
frontier, while a broad mechanical rename can stay fast.
fast: local, mechanical, reversible work with cheap, complete verification.standard: ordinary repo-grounded reasoning, multi-step drafting, normal coding and docs.frontier: architecture, high ambiguity, safety or security risk, release-critical review, or work where one wrong assumption wastes a large run.
Done when: one tier is chosen with its rationale.
- Read off effort. Default to the effort that tracks the tier. Raise it for ambiguity or long multi-step reasoning; lower it for a bounded task under a strong model. Effort buys deliberation, never capability.
glance: minimal deliberation, the direct path.measured: ordinary, everyday deliberation.thorough: deliberate extra: alternatives and assumptions checked.exhaustive: maximal deliberation, the search exhausted and re-checked.
Tier and effort move together by default (fast→glance, standard→measured, frontier→thorough, reserving exhaustive for the hardest stakes), then part where deliberation-hunger and capability-need diverge.
Done when: one effort level is chosen with its rationale.
- Report both coordinates. Give one shared rationale,
move up if and move down if triggers for each dial, and the proof surface the work still needs regardless of tier or effort. Done when: both dials, their triggers, and the proof surface are stated.
- Stop. Do not execute the sized task, change config, or switch models. Done when: no execution action is taken; the recommendation is the sole output.
Failure and recovery
- Ambiguous task boundary: refuse to size; ask the user to name the unit of work before proceeding.
- Insufficient risk signal: default to
standard/measured and state the assumption explicitly; do not guess.
- Partial result: if the task cannot be fully assessed, return the best-effort sizing with the uncertainty named; never silently default.
- Non-convergent: if repeated re-assessment yields different tiers without new evidence, report the last stable pair and the oscillation.
- No rollback applies: this skill is read-only and advisory.
Output
A report naming exactly one capability tier (fast/standard/frontier), one effort level (glance/measured/thorough/exhaustive), a shared rationale, move-up and move-down triggers per dial, and the proof surface — no routing, vendor, or orchestration claim.
Machine-readable block:
recommended_tier: fast|standard|frontier
recommended_effort: glance|measured|thorough|exhaustive
rationale: <one sentence, covering both dials>
move_up_if: <signals that would justify a stronger tier or higher effort>
move_down_if: <signals that would justify a cheaper tier or lower effort>
proof_surface: <verification still required>
1---2name: size-the-run-23description: Use when a run could be over- or under-powered, before dispatching a subagent, or when the user asks how hard to think about a task. Recommends one capability tier and one effort level on a neutral two-dial scale, without pinning a vendor model.4---5
6# Size the run
7
8## Contract
9
10| Field | Bound contract |
11|---|---|
12| Trigger | User asks how hard a run should be ('how hard should I think about this', 'is this worth a deep run', 'size this task') or a subagent dispatch is about to launch. |
13| Authority | Read-only. No file, VCS, credential, paid, published, deployed, or remote mutation. |
14| Side effect | One advisory recommendation in chat; no config, execution, or model change. |
15| Done | The report names exactly one capability tier plus one effort level (the cheapest sufficient pair), the move-up/move-down triggers, and the proof surface. |
16
17## Inputs
18
19- Task description (required): the work being sized: a concrete unit of work, not a vague goal.
20- Context (optional): domain constraints, risk factors, or prior sizing history that affect the judgment.
21
22## Procedure
23
241. **Frame the unit.** Name the exact work being sized. Done when: the unit of work is named in one sentence.
252. **Score once.** Assess risk and complexity across ownership boundaries, reversibility and blast radius, safety or privacy risk, ambiguity and synthesis load, need for research or adversarial review, and the cost of a wrong answer. Done when: all six risk dimensions are assessed.
263. **Read off tier.** Choose the cheapest tier whose ceiling covers the judgment and risk. Risk beats size: one high-risk file can want `frontier`, while a broad mechanical rename can stay `fast`.
27 - `fast`: local, mechanical, reversible work with cheap, complete verification.
28 - `standard`: ordinary repo-grounded reasoning, multi-step drafting, normal coding and docs.
29 - `frontier`: architecture, high ambiguity, safety or security risk, release-critical review, or work where one wrong assumption wastes a large run.
30 Done when: one tier is chosen with its rationale.
314. **Read off effort.** Default to the effort that tracks the tier. Raise it for ambiguity or long multi-step reasoning; lower it for a bounded task under a strong model. Effort buys deliberation, never capability.
32 - `glance`: minimal deliberation, the direct path.
33 - `measured`: ordinary, everyday deliberation.
34 - `thorough`: deliberate extra: alternatives and assumptions checked.
35 - `exhaustive`: maximal deliberation, the search exhausted and re-checked.
36 Tier and effort move together by default (`fast`→`glance`, `standard`→`measured`, `frontier`→`thorough`, reserving `exhaustive` for the hardest stakes), then part where deliberation-hunger and capability-need diverge.
37 Done when: one effort level is chosen with its rationale.
385. **Report both coordinates.** Give one shared rationale, `move up if` and `move down if` triggers for each dial, and the proof surface the work still needs regardless of tier or effort. Done when: both dials, their triggers, and the proof surface are stated.
396. **Stop.** Do not execute the sized task, change config, or switch models. Done when: no execution action is taken; the recommendation is the sole output.
40
41## Failure and recovery
42- Ambiguous task boundary: refuse to size; ask the user to name the unit of work before proceeding.
43- Insufficient risk signal: default to `standard`/`measured` and state the assumption explicitly; do not guess.
44- Partial result: if the task cannot be fully assessed, return the best-effort sizing with the uncertainty named; never silently default.
45- Non-convergent: if repeated re-assessment yields different tiers without new evidence, report the last stable pair and the oscillation.
46- No rollback applies: this skill is read-only and advisory.
47
48## Output
49A report naming exactly one capability tier (`fast`/`standard`/`frontier`), one effort level (`glance`/`measured`/`thorough`/`exhaustive`), a shared rationale, move-up and move-down triggers per dial, and the proof surface — no routing, vendor, or orchestration claim.
50
51Machine-readable block:
52
53```text
54recommended_tier: fast|standard|frontier
55recommended_effort: glance|measured|thorough|exhaustive
56rationale: <one sentence, covering both dials>
57move_up_if: <signals that would justify a stronger tier or higher effort>
58move_down_if: <signals that would justify a cheaper tier or lower effort>
59proof_surface: <verification still required>
60```
majiayu000