| name | dojo |
| description | This skill should be used when the user wants to create, improve, test, or evaluate a Claude Code skill โ new skills, skill edits, trigger tuning, or verifying a skill actually changes agent behavior. Responds to "dojo", "new skill", "write a skill", "turn this into a skill", "test this skill", "skill evals", "why isn't my skill triggering", "improve this skill description", or any request to build or harden a SKILL.md. |
Dojo โ Skills Earn Their Place
Overview
A skill is a claim about future agent behavior. Untested claims ship as
liabilities: self-generated, unverified skills measurably make agents
worse (SkillsBench โ1.3pp), while the same authorship under scored
rollouts and validation gates lifts frozen-model accuracy by ~20pp
(SkillOpt). It was never agents-writing-skills that failed โ it was
writing without verification. The dojo is where the verification happens:
no skill ships on vibes.
Dojo covers the full lifecycle: create, edit, test, evaluate, package.
For small edits, run only the kata that the change touches (a description
tweak needs kata 6โ7, not a full baseline) โ say which kata you're
skipping and why.
The Measurability Rule
Automated optimization needs a scalar grader; most skill output quality
has none. So dojo never pretend-measures quality. Pass criteria are:
- Observable process checks โ y/n facts about what the agent did
("scoped search to a discovered venue before deep-reading: y/n")
- Exact-match trigger tests โ which skill did the router pick
Holistic output quality stays human/agent judgment, applied openly as
judgment. If a skill someday has a real scalar grader, point SkillOpt at
it instead of the dojo.
Tiers โ Match Rigor to Skill Type
Full TDD on a reference doc is theater. Classify at intake:
| Tier | What it is | Example | Rigor |
|---|
| Discipline | Rules agents rationalize around | verification gates, TDD | Full RED-GREEN-REFACTOR + adversarial pressure variants |
| Technique | Multi-step orchestration | hivemind, suno-pack | Baseline-fail test + skilled-walkthrough test |
| Reference | Facts, flags, recipes | searxng | Correctness review + trigger eval |
Every tier gets the trigger eval (kata 6). Only discipline skills get
adversarial variants (time pressure, sunk cost, authority pressure).
The Seven Kata
-
Intake. What skill, which tier, new or edit. Design the test
scenarios NOW โ one per archetype the skill claims to handle, each
with pre-written y/n pass criteria. Persist the battery immediately
to ~/.limitless/dojo/<repo-slug>/<skill>/<skill>-scenarios.md โ
and during kata 2โ6, append every subagent prompt VERBATIM as
actually sent, plus each run's result line. Dojo artifacts are
skill-owned runtime output: keep them under ~/.limitless/dojo/ by
default so target repos do not fill with raw training debris. Curated
docs/examples may be copied into the repo only when the user asks.
Hold 1โ2 scenarios back: never used during iteration, run once at
the end (kata 5). If a transcript or session sparked the skill, mine
it for scenario material. See references/pressure-testing.md for
scenario and criteria design.
-
Baseline (RED). Before writing anything: fresh subagents run the
training scenarios without the skill. Score the criteria, document
the exact failure modes โ they are the skill's curriculum. A skill
that doesn't answer an observed failure is decoration. Skip only when
failure is already concretely evidenced; record that evidence instead.
-
Write (GREEN). Author the SKILL.md targeting the observed
failures โ walk the criteria list and check each one is answered by an
explicit instruction. House rules: under ~350 lines, progressive
disclosure into references/, description in pushy limitless style
("This skill should be used whenโฆ Responds to 'X', 'Y'โฆ") โ trigger
reliability beats minimalism.
-
Pressure-test (REFACTOR). Fresh subagents, same training
scenarios, skill content present. Score every criterion. On failure:
bounded edits โ one targeted add/delete/replace per loophole, then
re-run that scenario, so you know which change fixed what. Log edits
that didn't survive in the dojo record's rejected-fix table; future
sessions must not retry dead ends. Exit when all criteria pass or a
criterion is consciously demoted to Known Limitations with rationale.
-
Graduation. Run the held-out scenarios once, no edits in between.
Passing trained scenarios only proves you wrote to the test; the
holdout is the real signal. Fail โ fix with bounded edits, then design
NEW holdouts before re-attempting. Burned holdouts are never reused.
-
Trigger eval. Build a matrix of 10โ15 realistic prompts โ
should-trigger positives plus near-miss negatives that belong to other
installed skills or to none. Present the full installed-skill
description list to a routing-judge subagent, exact-match score, run
twice. Every positive must hit; a negative landing on the new skill is
a collision โ fix the description, re-run. See
references/trigger-evals.md.
-
Package & record. Plugin layout, manifest/README updates, reload,
smoke-invoke. Preserve the run outputs: copy every test
workspace (baselines, pressure, holdouts) into
~/.limitless/dojo/<repo-slug>/<skill>/<skill>-runs/<run-name>/
with a short README table โ the produced files ARE the readable
evidence. Leak-check anything copied from scratch space before it
enters the long-lived dojo folder. Then write the dojo record to
~/.limitless/dojo/<repo-slug>/<skill>/<skill>-record.md, linking
the scenarios file and the runs dir:
baseline findings, loopholes closed, rejected fixes, graduation
result, trigger score, known limitations. Every skill carries its belt
rank. See references/packaging.md for the checklist and template.
Subagent Rules
- Fresh subagent per test run โ never reuse one that has seen the skill
or the discussion.
- Subagents lack this conversation's context: every prompt carries a
self-contained scenario, environment notes, and tool availability.
- Pass criteria are written down before the run โ scoring after the
fact invites rationalization.
- Tell test subagents their final message is raw working data (steps
taken + result), not a user-facing summary.
When NOT to Use Dojo
- One-line config or doc tweaks that aren't skills.
- Writing non-skill documents (plans, specs, articles).
- Skills with a genuine scalar grader and eval corpus โ that's SkillOpt
territory, not hand-run kata.
References
references/pressure-testing.md โ scenario design, pass-criteria
patterns, baseline/pressure subagent prompt templates, adversarial
variants, the bounded-edit rule. Read at kata 1, 2, and 4.
references/trigger-evals.md โ matrix design, routing-judge template,
scoring, collision handling. Read at kata 6.
references/packaging.md โ limitless plugin conventions, ship
checklist, dojo record template. Read at kata 7.