Note completeness verification and task README. Load during Step 5 (Collection).
Verify that all per-step subdirectories contain their expected note files and write the task-level README.md. Collection is Step 5 in the 7-step workflow (Ideation, Plan, Research, Execution, Collection, Memorization, Review). Notes are written during their respective steps — Collection verifies completeness, it does not create the notes themselves.
Navigate deeper from here:
| Document | Covers |
|---|---|
| gotchas.md | Known mistakes and corrections for _collection |
Collection has three responsibilities: note verification, README.md writing, and gotcha recording.
Note verification — Verify that all expected files exist in their per-step subdirectories. Notes are written during their respective workflow steps (Ideation writes ideation/, Plan writes plan/, Research writes , Execution writes ). Collection does not create these files — it confirms they are present and complete. Required files by subdirectory:
research/execution/ideation/ideation.md (orchestrator synthesis) — must existideation/innovative.md and ideation/best.md (PI agent notes) — must existplan/plan.md — must existresearch/research.md (orchestrator synthesis) — must exist for non-trivial tasksresearch/innovative.md and research/best.md — must exist for non-trivial tasksresearch/subtasks/ — verify subtask JSON files exist for non-trivial tasksexecution/execution.md — must existexecution/subtasks/ — verify subtask JSON files existideation/evaluation/, plan/evaluation/, research/evaluation/, execution/evaluation/ — verify evaluation files exist where evaluation was performedIf any required file is missing, report the gap and investigate — do not silently proceed.
README.md writing — After verification passes, write the top-level README.md for the task directory. See the README.md section below for format.
Gotcha recording — Any corrections, surprises, or mistakes discovered during the workflow must be recorded via _gotcha before the cycle closes.
Task note directories follow the structure defined in _note. Each task directory contains per-step subdirectories:
{YYYYMMDD-HHMM}-{slug}-{session_id}/
README.md — task-level index (written by Collection)
ideation/
ideation.md — orchestrator synthesis
innovative.md — Innovative PI stance
best.md — Best-practice PI stance
evaluation/ — evaluation files (if evaluation performed)
plan/
plan.md — plan details
evaluation/ — evaluation files (if evaluation performed)
research/
research.md — orchestrator synthesis
innovative.md — Innovative researcher
best.md — Best-practice researcher
subtasks/
01-{slug}.json — research subtask records
evaluation/ — evaluation files (if evaluation performed)
execution/
execution.md — execution outcomes
subtasks/
01-{slug}.json — execution subtask records
evaluation/ — evaluation files (if evaluation performed)
feedback.md — feedback rounds (if FEEDBACK)
review/
innovative.md — Innovative PI review + verdict
best.md — Best-practice PI review + verdict
review.md — orchestrator synthesis
The task directory and its subdirectories are initialized by gobbi note init and populated during their respective workflow steps. Collection only writes README.md at the task root level.
Collection writes the task-level README.md after verifying all per-step subdirectories. This is the primary artifact Collection creates — everything else was written during earlier steps.
The task README.md must include:
ideation/, plan/, research/, execution/) with its key filesAlso update the parent README.md (the index file that lists all task note directories) to include this task's entry.
Collection runs at different points depending on the workflow phase. The verification checklist adapts to what the phase produced.
Verify all four subdirectories: ideation/, plan/, research/, execution/. Check that subtask JSON files exist in research/subtasks/ and execution/subtasks/. Check that evaluation/ subdirectories exist for any step where evaluation was performed.
Subtask JSON files are written during their respective steps (Step 3 and Step 4) via gobbi note collect with a --phase argument. Collection verifies they exist and conform to the subtask JSON format defined in _note — each file must contain the required fields: agentId, agentType, description, model, effort, timestamp, delegationPrompt, finalResult. Collection does not create these files.
Write feedback.md to the task root directory. Append each feedback round — do not overwrite previous rounds. Re-verify any subdirectories whose contents were modified during the feedback cycle.
The review/ subdirectory is updated with new PI review files. Verify that review/ contains the expected review artifacts. Update the task README.md to reflect the review results.
Always collect when the workflow involved delegation or produced results worth preserving.
Skip collecting when the task was trivial — a single small edit, a quick question, or a direct handle without delegation.
After writing notes and recording gotchas, use AskUserQuestion to ask: "What non-obvious knowledge was discovered this session?"
This complements per-agent Memorize steps by capturing orchestrator-level insights that no single agent saw. The user may say "nothing" — the question ensures knowledge doesn't get lost silently.
Capture categories:
Gotchas — mistakes or wrong assumptions corrected during the session. Record via _gotcha.
CLAUDE.md additions — conventions or patterns discovered that should persist across sessions. Add as one-line entries to CLAUDE.md.
Skill updates — behavioral patterns identified that a skill should teach. When a learning is categorized as a skill update, the orchestrator MUST propose a concrete change to the skill rather than merely flagging it for later. This is opt-in: only prompt if skill-update-type learnings were identified during the session. To propose a change:
gobbi validate lint) to catch anti-patterns before presentingFormat: one-line entries, concise, actionable. "Discovery X because Y" — not narrative.