organon

Figures that argue — design, size, and audit paper figures the way strong Nature/Science/IEEE papers do. Use when creating or reviewing any figure, choosing a chart form, sizing for a venue column, writing a caption, picking colors, or when a figure PDF must pass the size gate (editio-figcheck). Panel-first composition, venue widths from venue.json, caption grammar, statistical honesty, figure-as-unit provenance.

Build the paper — set up a TeX distribution in minutes, scaffold or rebuild the .editio/paper/ workspace for a venue, compile the three modes (draft/publish/blind), preview a single section, and keep notation defined-once. Use when installing or detecting TeX, adding a venue, running latexmk, switching modes, or the build breaks. The toolchain is the user's (skills, not stacks): latexmk is the swappable reference driver, venues are data folders.

Academic-writing orchestrator — turn a promptus store's validated knowledge into a defensible, submittable paper. Use when starting, resuming, structuring, rendering, or auditing a paper, or deciding which editio piece does a job. Knows the paper workspace, markdown-is-truth, three renders, the DoCO/DEO structure gate, and the evidence-calibrated audit loop.

Frame a paper's argument and write its sections — pick a section order (imrad, cs-systems, theory), build the narrative arc (funnel intro, gap, contributions mapped 1:1 to results), draft the abstract last from the summary-paragraph formula, and keep every section inside the DoCO/DEO structure gate. Use when starting a paper's skeleton, writing or restructuring a section, or fixing a broken arc. Grounds claims through promptus recall as it writes.

Audit a draft for both AI-writing tells and unsupported or over-confident claims. Use before prose ships or when grading editio claim spans against the Promptus store.

Run the authoritative Promptus whole-store health gate. Use to verify source/index freshness, stable-ID uniqueness, classification, typed relations, and optionally strict graph health.

Flush perishable session state before compaction — store unrecorded decisions, results, sources, and digested findings; refresh NOW and memory; check Telos drift; then re-index. Use before compaction or a long-session handoff.

Diagnose or migrate a Promptus store layout. Use for version-aware layout checks, dry-run migration planning, current-namespace repair, or store reachability problems.

Inspect the Promptus knowledge graph. Use to rank load-bearing units, lint dangling links and orphans, or suggest related-but-unlinked notes without embeddings.

Curate collected research into provenance-bearing literature units. Use to backfill recoverable sources or promote a misfiled external note without inventing provenance.

Initialize Promptus in the current repository — scaffold the four stores and portable AGENTS.md cadence, seed the mandate through the gate, and smoke-test retrieval. Use when a project needs a new .promptus store.

Orchestrator and map for the Promptus research knowledge system — a substrate for the LLM agent. Use at the start of research book-keeping to choose the right verb/script/skill — STORE (kb-add/kb-amend), BOOK-KEEP (promptus-check + kb-index + graph + checkpoint), RETRIEVE (kb-find → kb-get + recall). grannie is the one human read-port. Knows the four stores, substrate:status tagging, the [[link]] graph, and the invariant.

Retrieve what a project already knows, with provenance and status, before asserting it. Use when grounding a statement, answering "what did we decide / find / read about X", or checking a draft's claims against the store. Auto-invoke by judgement whenever a claim could be verified against recorded knowledge. Drives kb-find; verifies each claim against its source before synthesizing.

Proactively record a research project into Promptus as work happens — append a unit after every decision, run, observation, dead-end, or finding via kb-add (never freehand). Teaches the recording habit so the ledger fills itself instead of being reconstructed later. The script owns timestamp, id, and placement. Pairs with the promptus-checkpoint workflow.

Stand up a project's Promptus stores under a single `.promptus/` namespace, Telos first — and maintain the Telos as direction shifts. Use when initializing Promptus in a repo, AND whenever about to edit `.promptus/TELOS.md`: Telos edits are OPERATOR-TRIGGERED — the agent registers evidence in the research ledger (kb-add) and proposes; the boundary routes everything else to the ledger, NOW-header, or memory. Drives the templates/ scaffolds.

One source of truth for every load-bearing number in the paper. Use when a result value appears in prose/tables/captions, when results are recomputed, or before submission — numbers.json names each value once, @num:handle binds it everywhere, editio-numbers --write/--gate keep the paper matched to the sources it came from.

Remove signs of AI-generated writing from text and add the positive qualities of genuinely human writing. Use when editing or reviewing text to make it sound more natural and human-written. Detects and fixes patterns from Wikipedia's "Signs of AI writing" guide (inflated symbolism, promotional language, superficial -ing analyses, vague attributions, em dash overuse, rule of three, AI vocabulary, passive voice, negative parallelisms, filler phrases) AND installs positive "human factor" patterns drawn from acclaimed pre-2022 scientific and popular-science writing (calibrated confidence, first-person stance, engaging the strongest counterargument, earned understatement, concrete examples, plain words, real rhythm, dry personality, specific attribution, coining and defining terms, earned aphorism, and writing the way you talk). Calibration corpus spans 1945-2015 across many fields: biology, computer science, mathematics, economics, statistics, psychology, philosophy of mind.

Explain a concept in plain language, as if to a sharp, curious 90-year-old. Invoked as `/grannie explain <concept>`. Decides by judgement whether to look the concept up in the project store first (via the recall skill / kb-find) and ground the explanation, or just explain from general knowledge. When the editio plugin is installed, dials its humanizer's analogy, plain-word, and write-like-you-talk patterns to maximum accessibility; otherwise leans on the core patterns inlined below.