| name | moredakka |
| description | Use when you need a hard, bounded, multi-model improvement loop over the current problem surface. Best for prompts like "what should I do next here?", "review this branch brutally", "turn this diff into a minimal patch plan", or "give me the next 3 commits for this mess". Strongest today for repo/code work. Do not use for trivial one-file edits, non-software tasks, or broad research detached from the current local work surface. |
What this skill is for
This skill turns the current local problem surface into a disciplined multi-role analysis loop.
It should stay glued to the strongest current work surface, which today is usually repo/code context:
- current directory
- git branch
- changed files
- diff
- nearby docs like
README.md, AGENTS.md, PLAN.md, TODO.md, SPEC.md
- repo-local skill manifests if relevant
Then it should apply a bounded role loop:
- planner
- implementer
- breaker
- minimalist
- synthesizer
The output must be an operating report, not a brainstorm.
When to use it
Use this skill when the user is effectively asking for one of these:
- tell me what I am actually working on and what to do next
- review the current branch or diff hard
- turn a messy work surface into a plan of operations
- extract a minimal patch path
- get a stronger next-step sequence than one single model pass would give
When not to use it
Do not use this skill when:
- the task is a trivial direct edit
- there is no meaningful local problem context
- the user wants broad open-ended research rather than action on the current local surface
- the user needs actual patch application more than planning or review
Required behavior
-
Stay local first.
Read the current work surface before making recommendations.
-
Prefer diffs over file dumps.
The delta matters more than the entire repo.
-
Use bounded loops.
Usually 2 rounds is enough. Stop when novelty collapses.
-
Assign distinct roles.
Do not ask every model the same thing.
-
Use typed outputs.
Preserve clear fields: problems, actions, tests, risks, edits, commit plan, disagreements.
-
Emit one recommended path.
The user should get a concrete next move.
-
Never write files outside .moredakka/. No exceptions.
Any artifact moredakka produces — reports, packs, role outputs, scratch
notes, JSON dumps, run logs, caches, anything — MUST land under
<repo-root>/.moredakka/ (or a subpath of it). Do not create or write
into fixtures/, research/, notes/, tmp/, out/, reports/,
plans/, or any other top-level directory. If .moredakka/ does not
exist, create it. If the user names a different output location,
refuse and point them at .moredakka/. The repo's .gitignore is
expected to ignore .moredakka/ — never bypass that by writing
elsewhere "so it gets committed."
If the CLI is installed
Run the CLI from the repo root or current module root:
moredakka here
moredakka plan --objective "<objective>"
moredakka review --base-ref main
moredakka patch --objective "<objective>"
moredakka loop --rounds 3
If you want the raw local context packet:
moredakka pack --mode plan
If you want saved reports:
moredakka review --base-ref main --write-prefix .moredakka/latest
--write-prefix MUST point inside .moredakka/. Acceptable:
.moredakka/latest, .moredakka/reports/<name>, .moredakka/runs/<id>.
Never use a prefix like research/..., notes/..., out/..., or any
absolute path outside the repo's .moredakka/ directory. If you find
yourself wanting to save somewhere else, do not — print the report to
stdout and let the user decide where (if anywhere) to copy it.
If you are in the source repo and have not installed the package yet, the repo-local wrapper works too:
bin/moredakka here
If the CLI is not installed
Emulate the same contract manually:
- Inspect current branch, status, recent commits, and working diff.
- Read nearby docs.
- Infer the immediate objective.
- Run the equivalent role loop in a bounded way.
- Return the same report sections the CLI would produce.
The same filesystem rule applies in emulation mode: by default, return
the report inline in the chat and write nothing to disk. If — and only
if — the user explicitly asks for a saved artifact, write it under
.moredakka/ (e.g. .moredakka/reports/<slug>.md). Do not invent
sibling directories like research/, fixtures/, notes/, plans/,
or out/ to hold moredakka-style outputs. Those namespaces belong to
the project, not to this skill.
Report contract
Your final report should contain these sections:
- inferred objective
- one-line take
- selected path
- top problems
- next actions
- commit plan
- tests
- edit targets
- major risks
- disagreements
- stop conditions
- open questions
- confidence
Quality bar
The report must be:
- operational
- falsifiable
- tied to files and commands when possible
- explicit about uncertainty
- minimal enough to act on now
Read references/moredakka-contract.md if you need a compact contract summary.