Deliverable intake

SkillDocs & knowledge

Turns raw captures, dictated braindumps in a deliverable’s inbox/, or recent macwhspr dictation-log entries into proposed planning-wiki changes, shown as a unified diff for author acceptance or rejection. Classifies every capture line as a decision, evidence claim, question, task, or dropped item with a reason, and traces it to its source. Uncertain names and citations remain UNRESOLVED. Applies decisions only with explicit assent. Use when the user says "intake", "absorb my notes", "process the inbox", "I dictated something", or after notes from a talk, meeting, or reading session. Never edits source files, only planning/ and open questions.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Deliverable intake skill

What this skill tells your AI

The instructions your AI receives, as published by scdenney/open-science-skills in codex/deliverable-intake/SKILL.md and read by ahel’s review.

The author thinks out loud; the wiki stays authoritative. This skill is the seam between the two. It reads captures verbatim, proposes where each statement belongs, and applies nothing the author has not seen.

Inputs

  • The nearest deliverable.yml upward from cwd (or the path given). Read it, then planning_dir/00-README.md and every file that index names. That is the current state of the wiki; the proposal is a diff against it.
  • inbox/*.md that do not end in .done.md, oldest first. A capture is read verbatim and never edited.
  • --from-log N: show the last N entries of ~/.config/macwhspr/cleanup_log.jsonl ({ts, raw, cleaned} pairs) as a numbered list with the first line of each, and ask which to import. An imported entry is written to inbox/YYYY-MM-DD-HHMM-log.md with the cleaned text (the raw text appended under a <!-- raw --> marker) before anything else happens. Nothing is imported without being named.

Classify, then trace

Walk the capture paragraph by paragraph. Each statement gets exactly one class:

ClassGoes toNeeds assent?
decisionthe wiki file that owns the topic (00-README.md "Decisions recorded here", or the numbered file)yes, always
evidence claimthe wiki file for the section it supports, marked with the capture referenceno, but a claim naming a source that is not in sources_of_truth or the bib is written as UNRESOLVED: <claim>; when the manifest names a knowledge_base, the report lists each such source as a process-source candidate (drop the PDF in sources/unprocessed/) so intake feeds the knowledge base rather than bypassing it
question08-open-questions.md, open listno
task08-open-questions.md under a Tasks heading, or the top HANDOFF entry's next actions if it is due before the next sessionno
droppednowhere; listed in the report with the reason (duplicate, already resolved, off-topic)no

Every proposed line carries a trace: (inbox/2026-09-10-debrief.md ¶3). A capture paragraph that yields no proposal is listed as dropped with its reason, so the author can see the capture was read whole.

Names, dates, numbers, and citations that the capture states but the wiki cannot confirm stay as dictated and are tagged UNRESOLVED. Do not correct a dictated name against a guess; do not invent a citation to match a claim; do not resolve a number from memory.

Propose

Produce one unified diff across the wiki files (new files included) plus the dropped list. Show it. Then put the decisions to the author as a numbered list, each with the sentence that would be written and where; use AskUserQuestion when there are three or fewer, a numbered list otherwise.

Apply

On assent: apply the diff and rename the capture to *.done.md. Do not edit any existing HANDOFF.md entry, today's included; the intake is recorded by the closure step ($finished) as Intake: <capture> -> <n> decisions, <n> claims, <n> questions, <n> tasks, <n> dropped, and if the session ends without closure the renamed capture and the wiki diff are the record. Record which capture paragraphs were promoted (a <!-- promoted: ¶1,¶3 --> line at the end of the .done.md file) so a partially accepted capture is never promoted twice. On refusal of any decision, leave that line out and keep the capture unrenamed until the author says otherwise.

Report

Five lines: captures read; counts by class; UNRESOLVED items (each with what would resolve it); decisions applied; what to run next ($deliverable-lint if the wiki changed materially, otherwise nothing).

Notes

  • This is a single-worker task and a cheap model suffices. Do not spawn subagents.
  • A reaction to a lint report is a capture like any other: dictate it into inbox/, then intake it; findings the author accepts become tasks, findings they reject go under a Rejected findings heading in 08-open-questions.md (the authoritative wiki, not NOTES.md) with the finding id and one line of reason, so the next lint does not raise them again.
  • Register vocabulary as it appears: a proper noun the dictation mangled twice belongs in ~/.config/macwhspr/vocab.md (offer, never append silently).

Codex notes

Single worker, no delegation. Present the complete patch before assent and apply only the approved hunks. Never load dictation-log entries the user did not name. Files and formats are identical to the Claude skill.

Signals

GitHub stars
55
Forks
3
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
deliverable-intake
Source
github.com/scdenney/open-science-skills