Author: Nils W. Holmager
Real work, fully auditable, in one file. work is a standalone, target-agnostic improvement-reasoning skill: point an agent at anything it can reason about — code, documents, plans, letters, music — and it examines, improves, and records material decisions and their outcomes in an auditable trail. The whole skill is one readable markdown file: work/SKILL.md. No required installer, tooling, or dependency — by design, permanently.
It is built on the three Principles of Earned Autonomy:
- Operator's Intent — the agent is given a destination, not a route.
- Observable Autonomy — material decisions and predictions are recorded before action, with outcomes appended afterward; an observer can reconstruct what was done and why from the trail alone. The final Markdown cannot prove wall-clock write order, so stronger timing guarantees require independent capture.
- Convergence Is Silence — finding nothing actionable is a valid outcome; ceremony is never manufactured to look thorough.
Delegated work fails when context disappears between sessions or the model drifts from the human purpose. Literal prompt compliance is insufficient: fabricated supporting comments can satisfy a request for comments while defeating a journalist's real need for authentic evidence. Work externalizes Destination, current state, decisions, outcomes, and learning so practical context loss and cognitive drift can be prevented, detected, and recovered from at the workflow level. It does not claim perfect access to hidden model cognition.
Its parent, the full skills suite, delivers the same discipline as five separate skills (intent, improve, destination, orient, trail) — and proved the capability, but priced itself out of daily use: five skill-loads and full ceremony on every run, even a one-line fix. work is the consolidation — one skill-load, one loop, with Improve-grade decision reasoning whenever a real choice exists, while target evidence and durable narration scale to domain risk and stakes:
- Intent — folded into the loop's own first step, narrated only when the interpretation is genuinely uncertain.
- Improve — the base loop: build a risk-sized target map, form a falsifiable model, derive the lenses reality warrants, compare plausible routes, challenge structure, decide, act, and check human legibility. Standard and Full share these operations; only evidence and recording depth vary.
- Destination — a condensed first-contact process that asks one short question, escalating to a deeper multi-question interview when the stakes call for it. The interview ends on operator recognition rather than checklist completion. Later corrections, changed emphasis, new connections, or repeated near-misses can trigger one sourced hunch-question before contradiction or plateau; the agent detects the gap, while the operator alone settles what Destination becomes.
- Orient — one open-world operation: inspect the target before reading inherited scores, scan historical realizations and reversals, re-derive measurements from the current Destination, compare the old and new maps, then rate with cited evidence, active operational rules, and an advisory todo. It runs whenever Orientation is invalid, the operator asks, evidence challenges the current view, or a small backstop catches what judgment missed.
- Trail — always happens, every run, with a variable-depth entry format (Micro / Standard / Full) so routine work costs little and real decisions still get full reasoning.
The skill leads the workflow, not the operator: it announces what it's doing and why, and names the one decision that belongs to the operator at each moment. probe (contamination / pattern-matching detection) is intentionally excluded from this consolidation — a settled decision, not a placeholder.
Work remains complete and useful on its own. For domains that warrant independent traffic evidence, it also knows the optional llm-harness-proxy companion. The intended integration stores private proxy-authored session JSONL under target-local .acm/sessions/, where Work can inspect only the sessions relevant to risk, contradiction, provenance, audit, or a challenged self-report. The proxy currently writes .harness/sessions/; migration to ACM is future proxy work, not a capability claimed by this release.
Standalone Work provides an agent-authored semantic Trail. Harness-backed Work adds verified, independently captured protocol evidence; the two records stay separate and disagreements remain visible. Provider reasoning may be absent, provider-specific, incomplete, or not causally faithful, and streaming can release output before final persistence. The harness does not provide operator intent, authorization, or proof of reasoning correctness.
Work may offer a companion install or update when assurance is requested or compatibility is stale or unknown. It never silently installs, updates, starts, or routes traffic through the proxy; those environment changes require operator consent.
v3.14.1 — the operator's confirmed daily-usage skill, authored by Nils W. Holmager and dogfooded hard: the skill has been pointed at itself and at external targets (a Python code repo where it made a test-verified source change, and editorial/content repos), with every run — including its own design decisions — recorded in this repo's trail. Destination is trusted but lossy human compression: operator recognition confirms the current articulation, while sourced hunches can surface meaningful lag when the operator learns, changes emphasis, or makes a new connection during the work. Hunches remain questions until the operator confirms, corrects, or rejects them. Independent models converging to bounded silence is strong evidence against the current confirmed Destination, not proof that the operator's richer latent direction is complete. Work prescribes reasoning operations, auditability, and operator authority rather than a solution doctrine; target principles come from Destination and domain evidence. Self-runs improve reasoning because Work's reasoning architecture is the selected target in this repo, not because arbitrary targets must optimize reasoning or follow this repo's preferred route. Every decision-bearing run preserves Improve-grade judgment and inherits active lessons from a small Work current orientation block at the top of Orientation; Orient replaces that block while Trail preserves historical reasoning, and risk-triggered escalation recovers older realizations and reversals directly from Trail without generated learning/history files. Domain risk controls target evidence and Trail tier controls recording depth. When an external target is used as a probe for another target, Work records the enclosing mission in the evaluation Trail, restates it in the probe Trail, and returns once the registered evidence is obtained or further cost becomes disproportionate; local Orientation cannot silently turn the probe into the destination. Trail's two-stage record is mandatory and makes retrospective reconstruction a visible breach, while final Markdown alone is explicitly not claimed to prove wall-clock timing. Orient's rubric and todo remain advisory maps. Human legibility is part of auditability. Orientation records its exact Destination basis; a changed basis invalidates the old rubric and makes fresh Orient mandatory. A second design bar applies alongside the operator's own use: a stranger should be able to pick up work/SKILL.md and use it without the author explaining anything — jargon is defined inline, the file must be self-contained, and no sibling skill or hidden context may be required.
Point an agent at work/SKILL.md — copy the work/ folder into wherever your agent loads skills from (such as ~/.copilot/skills/), or just paste the file into a prompt. There is no installer, and there won't be one — by design. The whole point is that it's just a skill: one readable file, no build step, no script to trust between reading it and using it.