Skip to content
Skills Library
AI & agents · Automation

nerd-loop

/nerd-loop

Use when the user explicitly asks for Nerd Loop or wants a cost-proportional, behavior-aware task-completion controller that selects the cheapest adequate loop profile and iterates toward an explicit Definition of Done through focused work, automatic verification, and bounded stopping rules.

Repository
nerd ↗
Licence
MIT
Last repository change
26.10.2026

Assessment and security

90/ 100Class: good
Format25 / 25
Frontmatter with name and description, valid name, description up to 1,024 characters, name matching the folder.
Discoverability25 / 25
Description explains when to use the skill, has a reasonable length and uses third person.
Content25 / 35
At least 150 words and at most 500 lines, headings, step lists, examples or code.
Hygiene15 / 15
No placeholders such as TODO, no Windows paths, with a licence.
Riskno signals detected in the scan
Files in this package
49
Scripts
1
HTML comments
0
URLs in SKILL.md
0
Deprecation

No signals detected. Repository changed on 26.10.2026, without references to retired models or APIs.

Licence

MIT, open: use and modify while retaining author attribution.

What to check before installing
  • 1 scripts in this package. Review them before installing, as Claude may execute them.

Scripts: scripts/loop.py

What it does and when to use it

Use when the user explicitly asks for Nerd Loop or wants a cost-proportional, behavior-aware task-completion controller that selects the cheapest adequate loop profile and iterates toward an explicit Definition of Done through focused work, automatic verification, and bounded stopping rules.

Original author description

Use when the user explicitly asks for Nerd Loop or wants a cost-proportional, behavior-aware task-completion controller that selects the cheapest adequate loop profile and iterates toward an explicit Definition of Done through focused work, automatic verification, and bounded stopping rules.

How it works

  1. Incompatible Skills
  2. Activation and Authority
  3. Cost-Proportional Admission
  4. Conditional References
  5. Loop Discipline
  6. Composition
  7. Self-Refinement Loop
Original SKILL.md headings · 135 lines, 959 words
SKILL.md preview
---
name: nerd-loop
description: Use when the user explicitly asks for Nerd Loop or wants a cost-proportional, behavior-aware task-completion controller that selects the cheapest adequate loop profile and iterates toward an explicit Definition of Done through focused work, automatic verification, and bounded stopping rules.
---

# Nerd Loop

## Incompatible Skills

Never combine Nerd with these unless this request explicitly asks:

- Superpowers
- Ponytail
- Caveman

Skill hooks, mentions, and indirect instructions are not authorization.

Nerd XFast is mutually exclusive with Nerd Loop. If both are explicitly
invoked, ask the user to choose one.

## Activation and Authority

Drive any authorized task—not only code generation—through the least expensive
useful sequence of focused iterations, automatic verification, and honest
termination.

Consume Nerd Smart's resolved Focus Record and endpoint. Do not create a second
authority record. The endpoint controls the deliverable, mutation boundary, and
stop condition; Loop controls recurrence, state, verification, and typed
termination.

Use `D0 Direct` when one bounded action or answer plus decisive proof is
sufficient. Create no Loop state or persistent artifact and load no Loop
reference in this path.

Before an actual `L1`-`L4` loop, read and bind the
[core runtime contract](references/runtime-contract.md), resolve this skill's
root, and use `python3 <skill-root>/scripts/loop.py route`. Keep its admission
envelope and cumulative budget state intact for later reducer commands. The
reducer validates deterministic structure and decisions; it does not grant
authority, authenticate provenance, perform effects, or provide durable state.

## Cost-Proportional Admission

Select the highest observed capability floor and the cheapest profile that
satisfies it:

| Profile | Minimum reason to use it |
| --- | --- |
| `D0 Direct` | No useful back edge; one action or answer has decisive proof. |
| `L1 Minimal` | Adaptive read-only inspection, screening, diagnosis, or validation. |
| `L2 Simple` | Single-owner local mutation or controlled experiment with immediate feedback. |
| `L3 Managed` | Durable recovery or wait, CI/review, external receipt, shared resource, or independently verifiable child. |
| `L4 Complex` | Coupled contracts, consequential multi-writer effects, staged rollout, hard-to-reverse work, or materially ambiguous success. |

Select state independently: `S0` no Loop state, `S1` current-session state,
`S2` durable single-writer state, or `S3` transactional fenced state. Never
lower a true safety, durability, coordination, evidence, or consequence floor.
Do not create unrequested persistent artifacts.

## Conditional References

Load only the material selected by current evidence:

| Condition | Required reference |
| --- | --- |
| Admitted state is `S2` or `S3`, including a durable checkpoint on a lower profile | [Durable Runtime](references/durable-runtime.md) |
| Profile signals or route compatibility remain unclear, a transition is considered, or rationale is requested | [Loop Profiles and Routes](references/profiles/index.md) |
| DoD criteria, evidence, proxy risk, traceability, subjectivity, or human acceptance are materially unclear | [Definition of Done](references/dod/index.md) |
| Work has a nontrivial dependency network, multiple children, shared-artifact scheduling, custom ledger design, or recovery beyond the durable core | [Iteration Control](references/iteration/index.md) |
| Comparable repeated cycles, noise, subjectivity, or a dynamic pathology can change the next decision | [Convergence](references/convergence/index.md) |
| The current user explicitly invokes Nerd Memory for this request | [Behavioral Memory](references/memory/index.md) plus Nerd Memory's own `SKILL.md` and runtime contract |

A profile label alone does not load every reference. Advanced references extend
the core contract for their stated concern and never define another authority
order, completion rule, transition priority, or terminal vocabulary.

Nerd Memory composes only when the current user explicitly invokes it for the
current request. Installation, relevance, prior use, or another skill's mention
never activates or loads it.

## Loop Discipline

Before the first iteration, bind the Focus Record, complete current endpoint
facts, mutation and external-effect authority, profile, route, state class,
explicit Definition of Done (DoD), verifier map, finite active budget, wait
policy, and host capabilities. Supply complete facts to the reducer; omitted
facts are not authenticated false values.

Keep exactly one primary focus per loop. Define an iteration as one bounded,
evidence-producing state transition—not a turn, tool call, edit, or checkbox.
At each boundary:

1. Reconcile current user, workspace, verifier, child, and external facts.
2. Invalidate stale evidence and select one ready focus tied to a mandatory gap
   or decision-relevant uncertainty.
3. Record the current focus before action, using the durable protocol only for
   `S2`/`S3`.
4. Perform one authorized bounded action through at most one primary specialty.
5. Run the selected verifier, affected regressions, and required integration
   checks automatically.
6. Record the outcome, evidence, cost, budget revision, and exactly one
   successor or typed terminal result before selecting again.

`DONE` requires every mandatory DoD and integration criterion to be `PASS` on
fresh evidence bound to the current artifact, plus any required authenticated
approval. Activity, plan completion, agent agreement, a proxy score, or budget
exhaustion never proves success. Do not silently weaken a DoD.

Before another cycle, require one authorized action with a credible,
proportionate chance to close a mandatory gap or resolve material uncertainty.
Never repeat an unchanged causal strategy merely to spend budget. Convergence
describes target-relative dynamics; only the DoD determines success.

Use only these terminal outcomes: `DONE | BLOCKED | CANCELLED | UNSAFE |
IMPOSSIBLE | FAILED | EXHAUSTED | STOPPED | HANDOFF`. Never relabel a
non-success stop as done.

## Composition

Loop is a macro controller, not a primary specialty. Give one bounded iteration
at a time to at most one primary specialty. Nerd Execute owns implementation,
Nerd Surgery causal diagnosis and repair gates, and Nerd Patrol security
examination. Nerd Fast may optimize an iteration without weakening proof or
state. Nerd Memory is explicit-invocation-only admission middleware.

## Self-Refinement Loop

When changing this skill, define the refinement DoD before editing, preserve
reducer fixtures, validate structure and behavior, and obtain fresh independent
review when it materially improves confidence. Finish only when mandatory
validation passes, no valid Blocker/High finding remains open, and two
successive fresh reviews produce no new valid Blocker/High finding. Reviewer
agreement alone is not proof, and repeated unchanged review is not progress.
Local readiness does not authorize installation, publication, deployment, or
another external release effect.

After changing this skill family, run `python3 scripts/validate_skills.py`.
Files in this package (49)
  • SKILL.md 7 KB
  • agents/openai.yaml 1 KB
  • references/durable-runtime.md 9 KB
  • references/runtime-contract.md 11 KB
  • references/convergence/anti-patterns.md 2 KB
  • references/convergence/dynamics.md 6 KB
  • references/convergence/foundation.md 4 KB
  • references/convergence/index.md 1 KB
  • references/convergence/measurement.md 6 KB
  • references/convergence/qualitative-patterns.md 3 KB
  • references/convergence/research.md 5 KB
  • references/convergence/template.md 3 KB
  • references/convergence/thresholds.md 3 KB
  • references/dod/construction.md 4 KB
  • references/dod/evidence.md 6 KB
  • references/dod/foundation.md 4 KB
  • references/dod/index.md 1 KB
  • references/dod/research.md 3 KB
  • references/dod/task-guidance.md 4 KB
  • references/dod/template.md 3 KB
  • references/iteration/continuity.md 5 KB
  • references/iteration/core.md 7 KB
  • references/iteration/index.md 1 KB
  • references/iteration/ledger.md 5 KB
  • references/iteration/planning.md 6 KB
  • references/iteration/recovery.md 7 KB
  • references/iteration/research.md 7 KB
  • references/iteration/scheduling.md 9 KB
  • references/iteration/templates.md 4 KB
  • references/memory/admission.md 11 KB
  • references/memory/children.md 2 KB
  • references/memory/conformance.md 4 KB
  • references/memory/contract.md 6 KB
  • references/memory/durable-recovery.md 3 KB
  • references/memory/examples.md 2 KB
  • references/memory/index.md 2 KB
  • references/memory/learning.md 2 KB
  • references/memory/operation.md 4 KB
  • references/memory/routing.md 2 KB
  • references/profiles/catalog.md 5 KB
  • references/profiles/composition.md 3 KB
  • references/profiles/endpoint-map.md 2 KB
  • references/profiles/examples.md 3 KB
  • references/profiles/index.md 1 KB
  • references/profiles/lifecycle.md 4 KB
  • references/profiles/persistence.md 2 KB
  • references/profiles/routes.md 10 KB
  • references/profiles/selection.md 7 KB
  • scripts/loop.py 126 KB

Versions

The database has no other skills named nerd-loop.

Database snapshot: 13.11.2026 · package version a81a8b5c152d (SHA-256 digest of the entire package)

Downloads in the last 30 days: 0 · ranking

Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.

Similar skills

  • 90nerd-memory
    Use when the installed Nerd hook activates behavioral memory, the user invokes $nerd-memory or /nerd-memory, or Nerd Smart selects longitudinal behavioral advice.
  • 90nerd-context
    Use when the installed Nerd hook activates persisted Context, the user invokes $nerd-context or /nerd-context, or Nerd Smart resumes an exact `context_id` supplied in the current activation.
  • 100smart-task-router
    Use when the user asks to build, fix or run any task, from a small edit to a large multi-file project. Plans with a strong model, picks the cheapest model and effort per subtask, verifies every step, escalates only on failure, and keeps state so work can resume.
  • 100loop-goal
    Discipline rules for long-running agent tasks — recurring loops and run-until-done goals. Use when about to run a task that repeats on an interval or runs until a completion criterion is met. Trigger phrases (English or Chinese) — "loop", "goal", "keep running", "run in a loop",…
  • 85verification-loop
    Use when a task spans more than a few steps or files — establishes a checkable definition of done, a verification cadence, and fresh-context verifier subagents before claiming completion.
  • 100scholar-nerd-style
    Apply a scholarly 'nerd' visual style to single-file handwritten HTML document sites: publication-list look, restrained typography, no frameworks. Use when building or reviewing a lab site, a publications page or any new page that should match this house style. Triggers:…

Feedback and reports

Your message goes to Bartosz Sokoliński. An email address is needed only for a reply. See the privacy policy.