Find the next straightforward useful task

Type: types/instruction.md

Find one useful task on the unresolved frontier—the discoverable field of possible further work—whose need, scope, approach and verification are clear enough for an agent to carry out without inventing a solution to a hard problem.

Invocation and authority

An operator or parent agent loads this instruction when choosing the next KB task. Use any supplied objective, search scope, exclusions and search budget. If none are supplied, look for concrete improvements to this KB within a bounded inspection of local files. Choose where to look from current evidence; do not attempt an exhaustive survey or claim to have found the globally best task.

This instruction authorizes read-only discovery and a recommendation. It does not authorize edits, external acquisition, agent launches, running a workflow, executing the selected task or committing changes. Separate permission to execute must specify its scope; selection does not expand that permission. Follow the repository's instructions for any subsequently authorized execution.

Find and check a candidate

  1. Inspect working-tree status and any supplied coordination constraints. Avoid work already owned by another actor or whose relevant files are changing unless coordination is explicit.
  2. Search promising work surfaces: open log observations, Open Questions in notes, pending ledger entries, live proposals, unfinished workshop work, and reviews or other recurring work that may need attention. These are examples, not separate eligibility classes. A due item is a candidate, not an automatic priority. Start with the caller's focus when provided; otherwise choose likely entry points. Use titles, descriptions and headings before full files. For theory questions, this is a discovery aid, not a complete inventory:

bash rg -ni '^#{1,6}[[:space:]]+.*open questions' kb/theory --glob '*.md'

  1. Read a candidate's source and enough current evidence to check that it is still needed. Look for an existing answer, superseding decision or completed implementation. Historical observations and archived proposals are not live tasks merely because they contain unresolved language. Respect explicit deferral triggers; do not turn an unmet trigger into permission to proceed.
  2. Select a candidate only when you can name:
  3. a concrete benefit, not just another document or a closed-entry count;
  4. the exact defect or small unresolved task, with current evidence;
  5. a bounded change using a known approach;
  6. a check that can establish the intended result, including semantic support where a structural validator would not suffice.
  7. Skip candidates that require an unsettled design choice, speculative research, new policy, unavailable evidence, or an unbounded audit to determine the remedy. For an established workflow, check that its current procedure is usable for the bounded task; do not start it just because work is overdue. Do not redesign a system merely to make a candidate executable. A clear local repair may remain useful even when the surrounding problem is hard. Stop searching once a sufficiently useful candidate meets the checks.

If the inspection budget is exhausted or plausible candidates remain unclear, return that no straightforward task was established in the inspected scope. Name the limiting uncertainty; do not infer that the KB contains no useful work. Do not fill a quota by proposing cosmetic churn or continuing a failing approach.

Return

Recommend one task, briefly stating:

  • Task and source: the action and its originating path or section.
  • Benefit: what improves for a reader, operator or executing agent.
  • Current evidence: what establishes that it remains needed.
  • Bounded change: intended files and why the approach is straightforward.
  • Verification: what would show the repair worked.
  • Authority and stop condition: permission needed and what would make the task cease to be straightforward during execution.
  • Search coverage: surfaces inspected and important omissions, without claiming exhaustive prioritization.

Do not execute the recommendation under this instruction alone. If separately authorized execution uncovers a consequential choice or missing support, stop that task, preserve the finding and return for direction or select another task within the caller's remaining authority. Do not repeatedly elaborate a speculative solution merely because this task was selected first.