agentleFS
kurofunekurofune/claudit/.agents/skills/consult-council/SKILL.md

consult-council

Skill4 starsChanged 3 days agoMITView on GitHub

Ask two outside models from two different families (Astra through Codex, Fable through Claude Code; the provider profile can change the pair) what is wrong with a plan and what they would do instead, through `djinn consult`, then answer in plain words: a one-sentence bottom line, each concern explained simply, and what we will do, with the concern table last as details. Use when the operator asks, or suggest it before spending real money or making an irreversible design choice; never by reflex. Triggers on: /consult-council, consult the council, consult the gods, ask Astra and Fable, second opinion.

The file

We split this skill into the rules it gives an agent

8 rules. About 0 are common, 5 are distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 3

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Before anything else: run djinn skills check --here in the repo you are working in.

    line 9
  2. Distinctive:

    Before you write the brief, list every issue you already know of in the plan or code under consultation: an open review finding, a failing …

    line 29
  3. Distinctive:

    Leave an issue open only for a reason you can state: the fix needs an operator decision, or the fix is the very question you are asking the …

    line 35
  4. Distinctive:

    Exempt: when the operator is discarding the current plan or implementation to explore fresh ideas, fix nothing in the part being replaced …

    line 41
  5. Very specific:

    Write the brief to a temp file (e.g.

    line 47
  6. Distinctive:

    Never put your own conclusion, preference or confidence in the brief.

    line 55
  7. Very specific:

    When the plan is a filed bead, skip the file: --bead <id> uses the bead's title, description, acceptance criteria and design as the brief …

    line 58
  8. Distinctive:

    Check every concern against the code before you rule on it: a model's claim is a hypothesis, not evidence.

    line 84

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md
<!-- djinn-compat-check: begin -->

Before anything else: run djinn skills check --here in the repo you are working in. If it exits non-zero, stop: relay its output to the user verbatim — it names the fix — and run no other step of this skill. On exit 0, continue.

<!-- djinn-compat-check: end -->

/consult-council — ask the outside models to attack a plan

Two models from two families read the same brief, each in its own headless session in this repo, in parallel; neither sees the other's answer. Each says what is wrong with the plan and what it would do instead. You weigh both and say what you are changing.

When: the operator asks ("consult the council", "consult the gods", "ask Astra and Fable", "second opinion"), or you suggest it before spending real money or making a design choice that is hard to undo. Never run it by reflex.

1. Fix what you already know

A council run costs two models several minutes each. Do not spend it on problems you already know about: the models find them again and bury the new concerns under old ones.

Before you write the brief, list every issue you already know of in the plan or code under consultation: an open review finding, a failing check, a problem you found yourself. Fix each one now, then run again the check that showed it (the failing test, the build, the review) and see it pass. Do not wait for the council to come back first.

Leave an issue open only for a reason you can state: the fix needs an operator decision, or the fix is the very question you are asking the council. Each issue you leave open goes into the brief (§2) with its reason.

This step starts no new search. With no known issue, go straight to the brief.

Exempt: when the operator is discarding the current plan or implementation to explore fresh ideas, fix nothing in the part being replaced and do not run its checks again. Consult at once.

2. Write the brief

Write the brief to a temp file (e.g. $TMPDIR/consult-brief.md). It holds:

  • Goal: what the plan is for, in one or two sentences.
  • Constraints: what must stay true (DECISIONS.md rulings, budgets, deadlines, interfaces).
  • File pointers: path:line for every file the models should read. Point; do not paste.
  • Known open issues: one line for each known issue you left open (§1): the issue, its path:line or check name, and why it stays open. These are facts; the council spends no effort finding them again. Omit this bullet when no issue is open.
  • The plan, framed as a proposal to attack: "Here is the proposed plan. Find what is wrong with it."

Never put your own conclusion, preference or confidence in the brief. A model handed the answer tends to confirm it.

When the plan is a filed bead, skip the file: --bead <id> uses the bead's title, description, acceptance criteria and design as the brief and posts each answer on the bead as a djinn-consult <model>: comment. A fix to the plan is then a rewrite of those fields before you consult. Put the known open issues in the bead's design under a Known open issues: line (read the design first: bd update --design replaces the whole field), so --bead still carries them and still posts the answers.

3. Run djinn consult

djinn consult --brief "$TMPDIR/consult-brief.md" --json
# or: djinn consult --bead <id> --json

Options: --only astra runs one model; --astra medium --fable high sets each model's effort (default high). Each model may take several minutes; run the command in the background when you have other work.

It prints one JSON object per model: name, adapter, model, effort, verdict (agree, disagree or abstain), wrong (a list), instead, reason. An abstain carries why in reason (a failed run or an unreadable answer); say so in the answer rather than hiding it.

4. Answer with the bottom line first

Check every concern against the code before you rule on it: a model's claim is a hypothesis, not evidence. Then answer in this order, the table last.

Bottom line: one sentence stating the council's overall verdict and what changes in the plan (or that nothing changes, and why).

What the council said: a numbered list, one item per distinct concern, merged across both models. Explain each in one or two plain sentences, as to someone new to the code (explain like I'm five), and name which model raised it (Astra, Fable, or both). No table before this list.

What we will do: a numbered list. Each item names the concrete change to the plan — what file, step, budget or interface changes, and how. A bare "adopt", "reject" or "adapt" with no stated change is not an item. A rejected concern gets an item only to say what stays as-is and the file:line or fact that rejects it. "Both agree, proceeding" with nothing changed is not an answer: when nothing changes, say which concern you rejected and the evidence that rejects it.

Details: the concern table, under its own Details heading, after everything above:

ConcernAstraFableMy call
<one concern><its view, or "—"><its view, or "—"><the concrete change, or the file:line or fact that rejects it>

One row per concern from the numbered list. Name each column for the model that actually answered. The my-call column never holds a bare adopt, reject or adapt: it names the change made, or the evidence for rejecting.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.