| name | work-estimation |
| description | Estimate the necessary work before converting it into agent tokens, time, or money. Use when comparing implementation options, setting an agent budget, choosing estimate granularity or error tolerance, planning a discovery probe, or asking "how much work is this really?" / "预算怎么估？" / "工作量评估" / "误差容忍度". Do not use to invent a token price, approve spend, or replace strategic choice, execution, or calibration evidence. |
| argument-hint | [estimate | compare | review] |
Work Estimation
Principle expression
Primary: P07
Supporting: P02, P08, P15
Scope
Own one judgment: given a concrete desired state and evidence, what is the
smallest necessary work graph, what resolution can the present decision tolerate,
and what remains discovery rather than committed work?
Work is executor-independent. It is not person-days, tokens, dollars, elapsed
time, model capability, or a universal point scale. Conversion into resources,
human approval of a hard envelope, runtime enforcement, and retrospective
calibration belong to their respective owners.
Principle source
Use a host Sequence and matching interpretations when present. Otherwise use
this package's read-only fallback in references/sequence.md. Read
only P07, P02, P08, and P15 for this skill's stable lineage.
Start
Ground the work before assigning any size:
Current state and source evidence:
Target state / decision to change:
Actor and decision horizon:
Hard constraints and irreversible surface:
Current alternatives, dependencies, and unknowns:
Strongest case for using no estimate or a coarser estimate:
If the decision is one-step, reversible, and no later actor must compare or
approve a resource commitment, use ordinary disciplined development. Do not
manufacture a work graph.
Core method
- Recover the concrete state transition. Read the object, its owners,
existing evidence, and acceptance boundary. Separate observed work from a
hoped-for implementation or an executor's claimed effort.
- Construct the necessary work graph. Add a node only when it changes a
required state, resolves a decision-changing unknown, verifies a hard
constraint, or preserves a later settlement. State its dependency and
acceptance observation. Do not use a fixed investigation/design/build/test
sequence.
- Separate committed work from discovery branches. A discovery node exists
to decide whether later work is needed; it is not hidden contingency. Name
the observation that opens or closes each branch and the smallest allocation
needed to learn it.
- Choose decision-derived resolution. Use
resolution and tolerance. Estimate
only as finely as needed for the current choice. If two alternatives remain
indistinguishable within their tolerated uncertainty, route to a bounded
discovery practice rather than inventing precision.
- Return a Work Estimate, not a price. Use the
record shape when approval, handoff, or a
later comparison requires recovery. Name what an execution projection would
need—executor profile, relevant observations, and confidence—but do not
fabricate P50/P80/P95 or set a hard cap.
- Route the next owner. Route resource conversion and hard enforcement to
Work Cell/runtime policy; multi-alternative direction to strategic-advisory;
post-run calibration to practice-cycle; carrier choice to form-guidance;
durable placement to artifact-organization; and semantic insufficiency to
principle-cultivation.
Boundaries
- Do not map a work node directly to a Work Cell, person-day, token count, or
price. One executor may combine nodes; another may need several attempts.
- Do not hide uncertainty in an undifferentiated percentage buffer. Preserve
discovery branches and their opening observations separately.
- Do not use forecast overlap to pretend alternatives are equal. If uncertainty
can change the decision, identify the next information-gathering practice.
- Do not approve a Budget Envelope, release a child-cell allocation, or allow a
forecast view to accept its own result.
- Do not turn the record into a mandatory plan or universal work-point taxonomy.
Completion standard
The estimate is ready when it names the concrete state transition, necessary
nodes, discovery branches, dependencies, resolution/tolerance, omitted work,
and a disconfirming observation. It is supported only when an action probe
compares real alternatives without using tokens/person-days as work, a boundary
probe declines a one-step request or unsupported point forecast, and a context
probe reads the object and acceptance boundary rather than an executor price
sheet.
