3xa-harness
01Plugin Claude Code
Discipline skills for agents doing long-running real work: freeze the scope, gate what only a human can judge, log the decisions, close out honestly.
Plugin Claude Code
Discipline skills for agents doing long-running real work: freeze the scope, gate what only a human can judge, log the decisions, close out honestly.
Plugin Claude Code
The five-skill discipline loop for long-running agent work: freeze a task's scope into a workorder, gate sensory output behind human review, log decisions so they survive a context reset, hand off state when work changes hands, and close out with evidence instead of claims.
Plugin Claude Code
Two runnable checks over a project's own records: list the numbers stated as fact with no source a reader could check, and report the records and code that decayed quietly — each finding with the file, the line, and the edit that closes it.
Skill Claude CodeCodex
List the lines in a project's records that state a number as fact with nothing a reader could check it against, then handle each one by adding the source, marking it [UNVERIFIED], or deleting the claim. Use when writing or reviewing a decision log, workorder, closeout, or spec that a later session will trust; when a…
Skill Claude CodeCodex
Check a project's own records and code for the decay that accumulates without anyone noticing — a log past its split threshold, pointers that no longer resolve, corrections piling up on one cause, debt signals in source, documents nothing references any more — and report each finding with the file, the line, and the…
Plugin Claude Code
Six-stage diagnosis order for hard errors: reproduce, minimise, hypothesise, instrument, fix, regression-test — one stage's output is the next stage's entry ticket.
Skill Claude CodeCodex
Diagnose a hard error in six ordered stages — reproduce, minimise, hypothesise, instrument, fix, regression-test — where each stage produces the thing the next stage needs before it may start. Use on a crash, traceback, silent wrong answer, zero-byte or truncated output, intermittent flake, or a check that started…
Skill Claude CodeCodex
Keep one append-only log where decisions, rejections, and corrections survive a context reset — each entry carrying a fixed index line a later session can search, and pointing at where the full detail lives instead of copying it. Read it first when picking up a project you have no memory of; write to it when recording…
Skill Claude CodeCodex
Write one snapshot of the current state, organized by what the next reader needs rather than by what this session did, so a person or session with zero memory of the work can read it once and keep going — every open thread's status, every already-spent asset and how to start it, every cross-thread dependency, every…
Skill Claude CodeCodex
Close a task with a report a reviewer can trust — every acceptance line answered with evidence or a recorded blocking reason, human corrections credited to whoever made them, refuted hypotheses kept, background processes proven stopped, and the chat message copied verbatim from the report. Use when reporting work as…
Skill Claude CodeCodex
Hand images, audio, video, or any artifact only a human eye or ear can judge to a human, in one batch, and hold there until an identified reviewer records the approved batch scope — rejections made by dragging a file into a rejects folder. Use when producing renders, generated art, TTS or voice output, edited video…
Skill Claude CodeCodex
Freeze a task's scope into a written order before building, then run it to done in one pass — build, self-verify, self-fix, re-verify — closing with a passing result or a two-strike failure trail. Use when a task spans more than a few steps, when scope is drifting mid-build, when a request is too vague to start from…