extract-artifact-candidates

Mine the current chat/history for forgotten durable-residue candidates and backfill `.plans/<task>/ARTIFACT_CANDIDATES.md` without touching tracked `AGENTS.md`. Use when implementation/debugging ran l

SKILL.md

GitHub

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.

Candidate bar

Only record a candidate when all are true:

  1. not recoverable reliably from code/tests alone
  2. durable enough to matter beyond the current moment
  3. costly to forget
  4. not strong enough yet to become tracked AGENTS.md directly

Edge cases

  • If the current packet status is already done, prefer a light harvest only when the candidate inbox is clearly incomplete; do not reopen finished implementation work just because the chat was interesting.
  • If the chat contradicts code/tests or explicit human decisions in AMENDMENTS.md, trust code/tests and amendments; record at most a mismatch note in task artifacts.

Output shape

Default outcome should say one of:

  • no_new_candidates
  • backfilled_candidates
  • packet_missing_candidates_reported_in_chat
  • packet_repaired_and_candidates_backfilled

Do not

  • write tracked AGENTS.md
  • write AGENTS.override.md
  • summarize the whole implementation
  • treat every chat decision as durable
  • mine history so aggressively that ARTIFACT_CANDIDATES.md becomes a worklog
  • use raw chat archaeology to override code/tests or explicit human decisions in AMENDMENTS.md

Initiative mode

When harvesting from an initiative child packet:

  • keep the candidate capture child-local first
  • make it easy for sync-initiative to fold stable working-memory facts into the parent initiative
  • make it easy for distill-knowledge to decide later whether any candidate is durable enough for tracked memory

Communication

Honor active caveman mode for user-facing replies per ../../references/communication-mode.md. Keep durable artifacts normal unless the human asks otherwise. Drop caveman for safety/clarity when needed, then resume.