prompts.mikee.pro

Prompt 03 — Worker Spawn Spec

The orchestrator must supply all four parts on every spawn. Missing any one produces overlapping work and coverage gaps — this is the single most common cause of wasted parallelism.

The spec

OBJECTIVE
  <One sentence, artifact-shaped, carrying the success predicate for
   THIS subtask. Not "explore X" — "produce <artifact> such that
   <checkable property>." Include the narrowing assumptions this
   subtask may NOT make.>

OUTPUT FORMAT
  <Exact shape of what to return: file path, schema, structured
   fields, or inline artifact. State how each part should be
   written so it can be verified in isolation — premises and
   conclusion stated locally.
   Concrete artifacts only. Status reports are rejected.>

TOOLS & SOURCES
  <Which tools to use, which to avoid. What external search is
   permitted, and what it must never be used for.
   <If result independence matters:> Do not search for a solution
   to this exact subproblem; background and standard named results
   only. Log external queries.>

BOUNDARIES
  <What is out of scope. What you must NOT do — e.g. do not modify
   tests, do not touch files outside <path>, do not expand into
   adjacent work, do not coordinate with other workers.
   When to stop and return rather than expand scope.>

DOES NOT COUNT
  <The specific near misses THIS subtask invites. Exclude each by
   name — the narrowed-scope version, the reduction, the survey,
   the confident sketch.>

RETURN
  <The artifact>, modular and independently checkable.
  If blocked, return the exact missing step and what you tried —
  not a re-statement of the goal, and not a partial substitute.

Example (domain: codebase investigation)

OBJECTIVE
  Identify the defect that, when corrected, closes the v4-versus-v3
  production gap on the frozen evaluation slice. Return a candidate
  defect plus a reproduction script that demonstrates it. The
  candidate must explain a measurable share of the gap, not merely
  correlate with it.

OUTPUT FORMAT
  A single file at `findings/<id>.md` containing:
  1. DEFECT: one sentence naming the mechanism
  2. LOCATIONS: file:line for each contributing site
  3. REPRO: a script at `repro/<id>.sh` that, when run, produces a
     differential result between v3 and v4
  4. EXPECTED GAP CLOSED: estimated fraction, with the command and
     output that produced the estimate
  Every claim must cite the tool result that supports it.

TOOLS & SOURCES
  Use the repo, the frozen eval slice, and the reproduction harness.
  Do not modify anything outside `findings/` and `repro/`.
  External search is permitted for library behaviour and documented
  APIs only. Do not search for published analyses of this model's
  regression.

BOUNDARIES
  Do not propose fixes — diagnosis only. Do not widen the eval
  slice. Do not test hypotheses that require unavailable data; if
  the data is unavailable, that is a blocking condition to report.
  Do not coordinate with other investigators; your independence is
  the point.

DOES NOT COUNT
  - a correlational narrative with no intervention
  - a defect explaining under 10% of the gap
  - "data drift" without an identified slice and mechanism
  - a list of hypotheses ranked by plausibility
  - results that require modifying the evaluation harness

RETURN
  `findings/<id>.md` + `repro/<id>.sh`, both runnable.
  If blocked: the exact missing input and what you tried.

Orchestrator checklist before spawning