Twelve sentences. Copy whichever one sounds like your week.

After-action review. Deep, 2-3 pages, after a project phase or a completed agentic session that left a recorded plan and execution log. The printed record is a hard condition, not decoration: written reflection without an objective record to react to produces no improvement at all, and the record is what qualifies this sheet for its evidence base. Page 1: the plan as recorded ("what was supposed to happen") and, where the machine executed steps, a bare factual timeline with no interpretation, in the machine's voice, with a two-column react beside it under the line "The log is not the whole truth." Page 2: the gap zone — "Why the difference?" — human-only; the machine withholds all diagnosis. Then twin sustain/change lists covering successes as well as failures; after a success, reviewing what went wrong anyway is the working part. The four canonical questions are the fixed structure — what was supposed to happen, what happened, why the difference, what to sustain and change — and high structure outperforms low. Nothing gets ranked here: prioritization, if wanted, is a later Fast pick, and next actions get drafted on screen after scan-back. Sketch field: "Draw the timeline where it bent," paired with Second reading. Never offer it as emotional processing after a distressing event — that literature runs null-to-harm, and introspection belongs to the Distanced pair. No record, no AAR; a trivial session with no gap gets no sheet. Sources: evidence brief, after-action review cluster.

Premortem. Deep, 2 pages, only when a concrete committed plan exists. Page 1: the live plan restated in one tight paragraph, machine's voice — genuine work worth reacting to. Then the certainty frame, asking voice: "It is a year from now. The plan failed — completely, not marginally. Write the history." Certainty is the active ingredient, not pastness; the wording must make the failure total and certain. Below it, six to eight numbered lines — a deliberate, sourced exception to the ten-to-fourteen overbuild rule. Overbuilding serves novelty; this sheet serves confidence calibration, and forcing many failure thoughts feels so difficult that the felt difficulty restores the very confidence the exercise exists to reduce. One input constraint: "At least three must be failures that are our own doing, not the world's" — the documented skew runs toward blaming the world. Page 2: sketch field, "Draw how the failure unfolded," with Second reading; then a separate Fast pick — circle the two most likely to be the real cause — with the counterweight adapted to "One circle must be the reason nobody would say in the meeting," because this sheet's bias is self-censorship, not feasibility. Countermeasures are deliberately absent from the sheet; the AI drafts them after scan-back, which closes the documented gap between naming failure reasons and revising the plan. No timer, no mitigation wording anywhere. Not for Light, not for an already-anxious planner — every supporting study targets overconfidence, and deflating the deflated has no support — and not for repeat runs on an unchanged plan. Sources: evidence brief, premortem cluster.

Outside view. Deep, 2 pages, for a timeline, budget, or adoption number the human is about to commit to. Page 1: the reference class in the machine's voice — comparable cases and base rates, with provenance stated on the page in one line; prefer the human's own past cases when they exist. Below it, a distribution strip on dot-grid with the base-rate anchor pre-ticked; the human marks where this case honestly sits. One pen position, and the spatial judgment is the point: prose lets you dodge it. The strip is the kit's sketch zone. Then the forced linkage, non-optional because it is the working part: "Write the short history of this project going the way the reference class typically goes." Displaying past cases beside an estimate does nothing; the linkage writing is the intervention. Page 2: the differences list — "What would have to be true for you to beat the base rate. Each line must be evidence you hold today, not an intention" — then the committed number, written once, that the human will defend. Promise calibration, never accuracy: the tested effect removes systematic optimism without sharpening any individual estimate. The pre-ticked anchor is licensed for this estimation format only and never leaks onto a generative sheet, where the same move is design fixation. Decline the format when no defensible reference class exists or the base rates' provenance cannot be stated: a fabricated distribution printed as an anchor is worse than no sheet. Sources: evidence brief, outside view cluster.

Teach-back sheet. Deep, 2 pages, closed-book by design. Offer it only after the human has genuinely engaged a body of conceptual explanation — an hour of AI-explained material is the canonical trigger — and never after a skim: closed-book prompts after passive processing measurably hurt, because nothing is loaded to retrieve. Page 1: one Serif line — "Explain X, from memory, to someone who has never met it" — over a full ruled page, with a narrow gap column: "Mark the lines where you reached for a word you couldn't cash out." Ask for mechanism, how and why it works, never a list of reasons; the deflation of inflated understanding requires mechanistic explanation and does not happen otherwise. Page 2: dot-grid — "Draw the mechanism from memory" — with Second reading; at most three retrieval probes in the machine's voice, no answers. The machine withholds every scrap of its own explanation, including material it obviously has: printing the content converts a calibration instrument into a recognition test, and paper is the enforcement the screen cannot provide. Closed-book is the calibration instrument, not the only route to retention. Gap marks are annotation, never scoring. The return trip is the feedback event, under the one scoped exception to scan-back's authority rule: the handwritten explanation is authoritative as a record of what the human believes, not as fact, so scan-back checks it against the source, corrects factual errors on return, and re-explains the gap-marked pieces differently. One sheet is one retrieval event, not a retention program. Wrong tool for procedural or worked-example material, and never a substitute for a generative sheet. Sources: evidence brief, teach-back cluster.

Weekly review. Recurring ritual, Deep, 2 pages — the weekly self-meeting the skill description already solicits. Page 1: the inventory of open loops the machine can actually see — chat threads, task lists, calendar — laid out two-column react; the human triages with the pen protocol as designed: strike kills, green go, blue do, red review. These are reaction marks on machine work-product, rank-and-circle territory, not selection over creative options, so no selection counterweight prints; feasibility bias has no originality to corrupt here. Never quadrants, never a priority matrix, never scoring of any kind. Page 2: the mind-sweep, a human-only long list — "What has your attention that no system knows about?" — with the machine-visible loops marked as Spent Ground categories so the sweep goes where the machine cannot see; then a structure sketch: "Draw next week as a structure. Mark what carries the weight." No next-action or if-then wording prints anywhere on the sheet. Capturing undone tasks without then planning them leaves the intrusions intact, so the continuation after scan-back is load-bearing, not cleanup: next actions get drafted there, kept deliberately lightweight — elaborately batch-planning many goals in one sitting undermines the planning benefit — and the machine transcribes the human's plans, never audits them. The mind-sweep is never pre-filled: the machine's inventory is precisely not the human's mind, and printing guesses there is the design-fixation failure mode. Weekly, not daily. Never fused with introspection; the Distanced pair keeps its own sheet. Skip the format when the machine holds no genuine visible inventory — never invent loops. Sources: evidence brief, weekly review cluster.

Brainwriting rounds. Deep, group-only: three to six co-located people, one sheet each, passed by hand. Never solo — the solo interleaved variant was evaluated and cut, since re-anchoring a person on their own first concepts may reinstate the fixation the break relieved. The Deep page cap counts per participant, not per table: it protects one human's completion load, and each participant completes one page, so a six-person session is one page pattern printed six times, not a six-page kit. The sheet: the shared problem in one Serif line; a Spent Ground category header aimed at teammates' obvious ideas as much as the machine's, since copying concentrates on common upstream ideas whoever wrote them; then rows of three idea cells per round, each cell lightly dot-gridded so a sketch is as welcome as a sentence. One protocol line, event-based: "When your three are down, pass left." Each recipient reads the column above and adds three that extend, combine, or fork it. One HAND: box per sheet in the infrastructure voice — furniture, not a heading — routes a follow-up to whoever wrote a line. Filling it is optional and leaving it blank is a fully supported choice; nothing printed or said asks for a name, and the nominal-group evidence may favor anonymity outright. Ink keeps meaning intent, never identity, and ideas stay unattributed to their authors in synthesis. Nothing else prints: no timer (the five-minute round is the brand's convention, never isolated as active), no round count as a promise, no idea-count arithmetic, and no attend-and-memorize instruction — reading the column above does the attention work, and a printed memory instruction cut fluency in the one study that tested it. Selection is exiled to a separate Fast pick sheet after the rounds, with the named criterion, distance frame, and counterweight, never on a passing sheet. A slim open-territory band closes each sheet. No group, no format; a task that is not idea generation belongs to a different kit. Sources: evidence brief, brainwriting cluster.

Grinnell field kit. Deep series — the library's first multi-day format, and the multi-day sibling of the Observation log. One kit per sitting, up to 4 pages per kit: the Deep page cap counts per kit, not per project, and more entities means more sessions, not more pages. For projects with recurring entities worth tracking across days — stakeholders, subsystems, failure modes. Page 1, the journal: a printed header of labeled blanks the human fills at the time of writing — date, location, conditions — then a ruled narrative zone in the fact-forward register: what happened, who, where, with human-written times as data. Times are quantification, not schedule; the Observation log's timestamp column is the precedent. A dot-grid band runs beside the narrative for sketch maps and diagrams — the lineage's own instruction, and it satisfies the sketch requirement natively. Account pages, at most two per kit: one per recurring entity, the entity's name printed as the page header in the machine's voice — gathered inventory at category level, categories-not-instances territory — and every field below empty, human-only. A compact catalog closes the working pages: a numbered one-line register of the concrete artifacts captured this session — documents, links, quotes, samples. The journal is not a diary; the lineage's own doctrine excludes the reflective voice, so this kit never merges with the Distanced pair. The same-day writing discipline is real doctrine, and it lives in the chat framing and the human's practice; nothing on a sheet ever schedules, deadlines, or hurries. Kits return days apart and account pages accumulate: scan-back appends to an entity's running account across sessions rather than treating each scan as standalone, and reorders by the human-written dates, which are data. No generative zone, so no defixation pattern is required. Not for Light, not for ideation; a project with no recurring entities is already served by the Observation log. Sources: evidence brief, Grinnell field-note cluster.