extract-artifact-candidates

extract-artifact-candidates

Manual salvage pass.

Not a workflow gate. Not a replacement for distill-knowledge.

Run this at any point when the conversation may contain non-obvious rationale, hidden constraints, owner-boundary notes, rejected paths, revisit triggers, or start-here breadcrumbs that should be preserved as task-local candidates before they fade.

Read first

• ../references/task-packet-contract.md
• ../references/persistent-artifacts-contract.md
• ../references/initiative-workflow-contract.md when harvesting from an initiative child packet
• ../../references/communication-mode.md
• relevant tracked AGENTS.md chain
• applicable AGENTS.override.md only if local execution constraints matter
• current task packet if one exists: INDEX.md, PLAN.md, AMENDMENTS.md, ARTIFACT_CANDIDATES.md

Scope

This skill owns:

• manual history mining into ARTIFACT_CANDIDATES.md
• backfilling candidates earlier phases forgot to record
• compact INDEX.md note that a harvest occurred

It does not own:

• writing tracked AGENTS.md
• writing AGENTS.override.md
• replacing distill-knowledge
• changing PLAN.md, AMENDMENTS.md, or review verdicts
• bootstrapping a brand-new task packet from scratch

Workflow

1. Establish the target packet.
   • If one current .plans/<task>/ packet is already in play, use it.
   • If multiple plausible packets exist, stop and ask the human which packet to backfill.
   • If no packet exists, do not invent one silently. Report candidates in chat and recommend packet bootstrap via plan-task.
2. Repair candidate-file drift only if a packet exists.
   • If ARTIFACT_CANDIDATES.md is missing, create the empty placeholder and note the repair in INDEX.md.
3. Read current packet state before mining.
   • Avoid duplicating existing candidates.
   • Use PLAN.md and AMENDMENTS.md to align terminology, owner boundaries, and already-explicit decisions.
4. Mine current history deliberately.
   • Start with the current visible conversation.
   • Use available history/context-log evidence to surface earlier user decisions, rejected approaches, hidden constraints, review dispositions, and carryover summaries.
   • Ignore tool spam, raw logs, and dead-end debugging unless they reveal a durable trap, a tempting rejected path, or a non-obvious constraint worth preserving.
5. Triage with the candidate bar.
   • Keep only task-local residue that may matter later and is not already obvious from code/tests.
   • Prefer exact facts over softened paraphrase.
   • Good candidate classes:
     • non-obvious rationale
     • hidden constraints
     • owner-boundary clarifications
     • rejected tempting approaches / do not casually resurrect
     • revisit triggers
     • start-here navigation / first validation commands
     • repeated slop traps exposed by the chat
   • Reject:
     • implementation summaries
     • transient debugging noise
     • todo/checklist residue
     • behavior prose already obvious from code/tests
6. Write compact candidates.
   • Append or merge into ARTIFACT_CANDIDATES.md.
   • Keep each candidate short, concrete, and attributable.
   • Include:
     • Type:
     • Why keep:
     • Candidate:
     • Source:
     • optional Related code: / Related packet refs:
   • If a prior candidate is duplicated or superseded, merge instead of stacking near-identical entries.
7. Note the harvest in INDEX.md.
   • Add a short note with what was harvested or no_new_candidates.
8. Hand off clearly.
   • Report what was added, what was intentionally skipped, and whether distill-knowledge should later revisit any item.