← Back to the essay

One of Ida's skills

Deep Reckoning

A skill for a question Ida has held open for months: too tangled and too frightening to settle in one sitting. Over many sessions, it turns that question into a page she can weigh and act from. She launches it by hand, rarely, and only when a question qualifies — unlike her other skills, it never fires on its own.

Most of this page, and the skill it documents, was written by AI, specifically Claude Opus 4.8 and 5. Deep Reckoning is a skill Ida's AI wrote and revises from her feedback; the same AI compiled this page from the skill's files, and Ida edited it. The essay it belongs to, Obsessing Over AI Agents Is Slowing You Down, is entirely her writing.

The Skill Files


The largest skill of the four: one instruction file and seven reference files. Unlike the other three skills, none of these files are shown or downloadable: all of them are woven through with material from the two real reckonings the skill was built from, which stay private. The descriptions below say what each one holds.

deep-reckoning/ ├── SKILL.md — the pipeline, pausing, and resuming └── references/ ├── cycle-procedures.md — the procedure for each pass, with checks ├── gates.md — twelve reasoning disciplines, in full ├── scope-probe-method.md — how the real question gets found ├── staging-and-recon.md — how the evidence gets gathered ├── render-phase.md — how the final page gets designed ├── writing-standard.md — the prose standard every deliverable inherits └── reckoning-log-template.md — the schema for the state file
deep-reckoning/
The eight files, each described.
SKILL.mddescription only

The loaded instructions: the pipeline, the scoping and staging phases, the cycle each pass runs, and how a run pauses and resumes across sessions.

This file is woven through with material from the two real reckonings, so it stays described rather than shown.

cycle-procedures.mddescription only

The full procedure for each cycle step, with acceptance checks and the per-phase execute requirements.

This file is woven through with material from the two real reckonings, so it stays described rather than shown.

gates.mddescription only

All twelve gates in full: the rule, the documented failure it prevents, where it is enforced, and the check that decides pass or fail.

This file is woven through with material from the two real reckonings, so it stays described rather than shown.

scope-probe-method.mddescription only

The Scope dialogue: the list of shapes a hidden question tends to wear, the construction rules for candidate framings, and worked probe dialogues.

This file is woven through with material from the two real reckonings, so it stays described rather than shown.

staging-and-recon.mddescription only

The staging procedure: the two lanes of evidence, the source-file discipline, and the rules for citing outside research honestly.

This file is woven through with material from the two real reckonings, so it stays described rather than shown.

render-phase.mddescription only

The Render procedure: the design consultation, the clarity checks, and the hard acceptance rules a finished surface must pass.

This file is woven through with material from the two real reckonings, so it stays described rather than shown.

writing-standard.mddescription only

The prose standard every deliverable inherits, with the naming rules the structure-and-naming gate enforces.

This file is woven through with material from the two real reckonings, so it stays described rather than shown.

reckoning-log-template.mddescription only

The schema for the single state file: the ledgers, the phase table, the section for questions still waiting on Ida.

This file is woven through with material from the two real reckonings, so it stays described rather than shown.

What It Is


Some questions don't fit in a work session. They have sat unresolved for months, tangled up with other decisions, and part of what keeps them unresolved is fear of what the answer might be. The skill's shorthand: a question with a ditch under it.

The real founding questions are too personal to publish, which is true of most questions this skill exists for. So here is an invented one, marked as such: a creative director's, written the way Scope would leave it, in first person:

My team now ships work I would have rejected five years ago, and clients love it. Do I retrain my taste, or is holding the old bar the last useful thing I do?

A question like that can be failed in two opposite directions, and the skill is built against both. It can be whitewashed — false reassurance, "your standards haven't slipped" when the honest answer is harder. Or it can collapse into the dreaded binary — the reductive verdict the person fears, delivered with false confidence. Both failure directions are named out loud at the start of every reckoning, and the output has to hold a genuine range of options and one named recommendation. A companion rule keeps the answer honest: confidence sets how firmly a conclusion is stated, never whether it appears. A hard answer the reckoning is only 60% sure of still gets stated, at 60%.

◇ Qualifies
  • held open for months and high-stakes
  • nebulous and entangled with other decisions
  • frozen partly by fear of the answer
  • rich in real evidence — career and identity, practice and business, financial structure, creative legacy
◇ Doesn't
  • a merely stalled project (the skill refuses to run on one)
  • everyday decisions
  • prioritizing what to do next
  • purely emotional or relational questions, which are held out as a known future variant

A full reckoning runs across six or more fresh sessions, and the skill says so before anything starts, so a casual invocation can downshift to a lighter run that stops earlier.

What It Needs to Run


How It Runs


A reckoning is built in passes, each producing a piece the next one builds on.

Scopea conversation, never a document analysis, because the presenting question is usually the wrong one. It ends with the founding question written in Ida's own first person.
Stagethe skill decides what evidence the question needs and goes and gets it: her named source files on one side, outside research on the other, every citation verified in a fresh window before it's trusted. The corpus is over-supplied on purpose, then ranked so the reasoning leans on the strongest material.
Reckonthe paths compared on shared dimensions, a confidence per path with its derivation shown, tensions held open rather than resolved, and one named recommendation. A lighter run stops here.
Forward-project · optionaleach path played forward: what doing it looks like, what would make it fail, and the go/no-go signals to watch for.
Render · optionaleverything rebuilt as one navigable page, described below.

Why Scope is a conversation. The question Ida arrives with is usually a stand-in for the real one — the reckoning that taught this began as a small copy-revision project that turned out to be quietly carrying a much bigger question. So Scope offers her two or three candidate framings at different sizes to react to, because she recognizes and edits a stated framing far more reliably than she generates one from nothing. Get the question wrong and everything after is aimed at the wrong target.

Why every step gets a fresh window. Each pass is drafted in the main conversation, then stress-tested by a fresh session that has never seen the work, then executed by another fresh session working only from the brief. A session grading its own work misses what an independent one catches — a failure documented in the reckoning this skill was built from.

The Gates


The skill writes twelve reasoning rules, called gates, into every pass. Two guard against the two failure directions, whitewash and the false verdict. The other ten are working rules for the reasoning itself:

Gate What it holds
1 · Register, not existence Every conclusion appears at its true confidence; a high bar sets tone only, never suppresses the hard answer
2 · The two fears The bar names whitewash and collapse-into-binary; the output carries a range and a named recommendation
3 · Ground or flag A path is instructive only with a real exemplar or a record episode; naming who to call is a complete answer, not a failure
4 · Symmetric search Look for disconfirming cases and the base rate, not just confirming ones; state the reference class before the particulars adjust it
5 · Evidence calibration Tag each claim by what could settle it — records, a real-world test, or only living it; every confidence carries its derivation
6 · Hold tensions Named tensions stay open; a verdict that depends on one is rendered conditionally, with what would settle it
7 · Cross-phase contradiction A later phase reports where its findings move an earlier read rather than silently deferring to it
8 · Desire, quietly rendered Stated desire is weighed alongside the calibrated read, marked once, never promoted above it
9 · Preserve the position A decision that is Ida's renders as hers to make, marked "your call," never a filled-in default
10 · Name what's beyond reach Blind spots are flagged up front; every named exemplar carries its verification tag
11 · One-glance top layer The surface opens with the founding question, then the answer, then the recommendation, above the detail
12 · Structure and naming A set of distinctions is a framework only if moving a case across its parts changes a decision; names carry their concepts

The gates exist because of an honest problem. The process was worked out on a stronger, slower model that held these disciplines by judgment. Day to day the skill runs on the faster everyday model, which fails in exactly the opposite ways: quick verdicts on thin evidence, more confidence than the evidence supports, drama where calibration should be. The gates put the discipline into the structure, so it doesn't depend on trusting the model. Whether that works is still being tested; in the first two reckonings, it did.

The Output


The output is a designed page rather than a report, because Ida reasons over structure she can see far better than she organizes from a blank field. The final surface opens with the founding question verbatim, then a one-line answer, then the recommendation — the whole reckoning at one glance, detail underneath.

The Render pass doesn't start from a blank page either. The skill keeps six fully built design directions — from a calm retro technical manual to art nouveau, each suited to a different kind of question — all rendering the same invented case, a fictional bookshop weighing its options, so they compare fairly. Ida reacts to real pages in a browser, and the winner gets rebuilt in the reckoning's own terms.

Changes So Far


Jun–Jul 2026 · Extracted from a real run

Ida ran one reckoning by hand before the skill existed. The result held up — she kept returning to the output for weeks, guiding her actions by its recommendations, and it showed her approaches fitted to her priorities and context that she hadn't considered. So the process that produced it was written down and hardened into a repeatable skill with named checks, built bottom-up from a real run rather than designed in the abstract.

Jul 2026 · First codified run

The next reckoning was the first to run under the written version, and it doubled as the skill's validation: it surfaced fifteen refinements that folded back into the files.

Still early. Two reckonings in, the scope stays narrow on purpose, and the cost is real — which is why the skill states the commitment before anything starts.