Applying the Dojo Protocol
The ten rules are injected once at session start. The full text lives in one place: ${CLAUDE_PLUGIN_ROOT}/PROTOCOL.md. Read it if it is not in your context. This skill does not repeat it; it tells you which rule to reach for and how. Rules are cited by number.
The injection reaches the main thread only. Subagents never see it, so a brief to a subagent has to carry the rules it needs (see the delegate skill). The four role agents in this plugin already carry theirs.
Role to model
| Role |
Agent |
Model |
Use it to |
| scout |
dojo-protocol:scout |
haiku |
look things up: find, read a slice, cite path:line |
| builder |
dojo-protocol:builder |
sonnet |
make a briefed change in files it owns and run the check |
| reviewer |
dojo-protocol:reviewer |
sonnet |
read a change with fresh eyes, find real defects or say none |
| judge |
dojo-protocol:judge |
opus |
try to refute claims and re-run the checks |
Plain subagent types and workflow steps have no pinned model. Pass model on every dispatch (rule 1). An omitted model inherits whatever the session runs on, which is usually your top tier.
Tell, then action
| If you notice... |
Do this |
Rule |
| a subagent or workflow step about to start |
set model and say why |
1 |
| an edit to code you have not read this session |
search, read the slice, or send a scout |
2 |
| a large file, a log, a whole directory |
grep, then read excerpts |
3 |
| a fix of a few lines in one file |
do it yourself, no agent |
3 |
| three or more files, or two or more agents |
write the shared decisions to a file first |
4 |
| you are about to write "done" or "tests pass" |
run the check now and show it |
5 |
| a fan-out summary |
give dispatched, returned and failed counts |
6 |
| you are about to patch |
name the causal chain and the smallest experiment |
7 |
| it fails every time, at a boundary |
check env, settings and credentials first |
8 |
| a probe came back empty |
feed it a known positive before believing it |
9 |
| push, deploy, publish, delete, send |
list what leaves the machine, ask, wait for yes |
10 |
Which skill to open next
- Unread code to change:
scout-first.
- Work that spans files or agents:
contract, then delegate for each brief.
- A change or a claim to evaluate:
refute.
Overrides
./DOJO.md in the project root replaces the injected text for that project.
DOJO_PROTOCOL_OFF=1 turns the injection off. DOJO_OFF=1 turns off every Dojo plugin.
The protocol is injected text. Nothing enforces it; the agents' tools lists are the only hard limits in this plugin.
1---2name: protocol3description: "Use when starting work or dispatching agents: which of the ten rules applies, and which model fits each role."4model: inherit5category: dispatch-coordinate6metadata:7 source: https://skillmd.com/skills/dojogenesis/protocol8---910# Applying the Dojo Protocol1112The ten rules are injected once at session start. The full text lives in one place: `${CLAUDE_PLUGIN_ROOT}/PROTOCOL.md`. Read it if it is not in your context. This skill does not repeat it; it tells you which rule to reach for and how. Rules are cited by number.1314The injection reaches the main thread only. Subagents never see it, so a brief to a subagent has to carry the rules it needs (see the `delegate` skill). The four role agents in this plugin already carry theirs.1516## Role to model1718| Role | Agent | Model | Use it to |19|---|---|---|---|20| scout | `dojo-protocol:scout` | haiku | look things up: find, read a slice, cite `path:line` |21| builder | `dojo-protocol:builder` | sonnet | make a briefed change in files it owns and run the check |22| reviewer | `dojo-protocol:reviewer` | sonnet | read a change with fresh eyes, find real defects or say none |23| judge | `dojo-protocol:judge` | opus | try to refute claims and re-run the checks |2425Plain subagent types and workflow steps have no pinned model. Pass `model` on every dispatch (rule 1). An omitted model inherits whatever the session runs on, which is usually your top tier.2627## Tell, then action2829| If you notice... | Do this | Rule |30|---|---|---|31| a subagent or workflow step about to start | set `model` and say why | 1 |32| an edit to code you have not read this session | search, read the slice, or send a scout | 2 |33| a large file, a log, a whole directory | grep, then read excerpts | 3 |34| a fix of a few lines in one file | do it yourself, no agent | 3 |35| three or more files, or two or more agents | write the shared decisions to a file first | 4 |36| you are about to write "done" or "tests pass" | run the check now and show it | 5 |37| a fan-out summary | give dispatched, returned and failed counts | 6 |38| you are about to patch | name the causal chain and the smallest experiment | 7 |39| it fails every time, at a boundary | check env, settings and credentials first | 8 |40| a probe came back empty | feed it a known positive before believing it | 9 |41| push, deploy, publish, delete, send | list what leaves the machine, ask, wait for yes | 10 |4243## Which skill to open next4445- Unread code to change: `scout-first`.46- Work that spans files or agents: `contract`, then `delegate` for each brief.47- A change or a claim to evaluate: `refute`.4849## Overrides5051- `./DOJO.md` in the project root replaces the injected text for that project.52- `DOJO_PROTOCOL_OFF=1` turns the injection off. `DOJO_OFF=1` turns off every Dojo plugin.5354The protocol is injected text. Nothing enforces it; the agents' tools lists are the only hard limits in this plugin.