agentleFS
rozputniirozputnii/ai-usage-app/.agents/skills/eco/SKILL.md

eco

Skill0 starsChanged 33 days agoMITView on GitHub

Owner-invoked economy mode. The primary works lean and mostly alone, delegates only where it pays to lean subagents on the cheapest model that keeps the required quality, and keeps a brief ledger and a path list so that no saving costs a missed detail. Use only when the owner types /eco or asks for economy mode.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

12 rules. About 0 are common, 9 are distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈9

    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
  • could be automatic: a setting could enforce it
  1. Distinctive:

    Every call re-sends the whole context.

    line 18
  2. Distinctive:

    Do the work yourself unless a role below pays.

    line 29
  3. Distinctive:

    Put independent tool calls in one message.

    line 32
  4. Distinctive:

    Read only what the task needs: grep with line numbers, read ranges, never re-read a file you hold.

    line 33
  5. Distinctive:

    Keep command output quiet: full output to a log, then the exit code and the first error lines.

    line 36
  6. Distinctive:

    Keep plans, briefs and replies short.

    line 40
  7. Distinctive:

    When a subagent fails the same step twice, or reports something the state contradicts, rerun that step one level up: haiku high, then …

    line 96
  8. Distinctive:

    Never review below the reviewer entries.

    line 102
  9. Distinctive:

    Never run haiku below high effort (owner rule).

    line 105
  10. Very specific:

    Every required review step ran, or is recorded NOT_RUN with the command's own reason.

    line 112
  11. Very specific:

    Add eco: <solo | the delegation chain> to the plan line, and one line to the final report: the delegations made, their models and efforts, …

    line 115
  12. Very specific:

    When checked_at in model costs is more than three months old, or a new model generation has shipped, say so when the mode starts.

    line 120

≈ “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

eco

Owner-invoked mode (decisions R-222, R-225, R-226). It changes who does the work, on which model, and how the details are tracked, never what the rules require: tiers, red to green, reviews, the merge gate and the always-ask list stay as AGENTS.md and CONTRIBUTING.md define them, and they win wherever this text seems to differ. The mode lasts until the session ends or the owner turns it off. The evidence is in model costs.

The mode was measured with the primary on opus at medium effort. The session's effort is the owner's setting; this skill cannot change it.

What costs

  • Every call re-sends the whole context. Cached reads are about 60% of spend, so cost is roughly context size times the number of calls. The primary is 68-85% of a task's cost.
  • A subagent starts cold: about 4-8k tokens as aiu-scout or aiu-runner, 18-55k as Explore or general-purpose, plus the brief and the report. Delegation pays only when it keeps much more than that out of your context, runs on a much cheaper model, or runs in parallel.
  • Rework and a missed detail are the most expensive outcomes. A too-weak model on a write or a review is false economy.

Primary: lean, mostly alone

  • Do the work yourself unless a role below pays. Measured directionally (confounded with the primary's effort): delegating implementation to an opus worker saved little against the normal mode; a lean primary doing it cost about half.
  • Put independent tool calls in one message; fewer calls is the main saving.
  • Read only what the task needs: grep with line numbers, read ranges, never re-read a file you hold. For rules, read the skill for the area and the named sections of CONTRIBUTING.md or verification.md, found with grep.
  • Keep command output quiet: full output to a log, then the exit code and the first error lines. In the inner loop run one test class; run the full suites in the background, not in the inner loop, and wait for the completion notice. The merge gate's run on the merged tree still applies.
  • Keep plans, briefs and replies short. Think hard at decision points, not on routine steps.
  • One task per session; start a fresh session for the next task.

Brief ledger and path list

A cheaper session saves most of its cost by thinking less, and that is where details get lost. Keep both lists on every task, T0 included, as CONTRIBUTING (Brief ledger and path list) defines them, and write them right after the plan line.

Roles

Delegate only to these agents, not to general-purpose or Explore directly: those start at 18-55k tokens against 4-8k. When one of these agents is not loaded, use a general-purpose agent given its contract, as CLAUDE.md says. Pass the model and effort of the matching entry in eco models on every call (CONTRIBUTING, Subagent models).

WorkAgentEntry in eco models
Locate code or answer "where and what" across many files, when much work remainsaiu-scoutaiu-scout; aiu-scout/why when it must explain why
Full check runs (restore, build, suites, validator, diff check, smokes)aiu-runneraiu-runner
Work with a clear brief: a large, separable implementation, a small, fully specified mechanical one, or the diagnosis of an unexplained failureaiu-implementeraiu-implementer; aiu-implementer/small for the small brief; aiu-implementer/diagnose for the diagnosis
Every fresh-context review the tier requires (per-task, whole-feature, T3 focused; a T0 diff stays the primary's review)aiu-revieweraiu-reviewer, T3 included; aiu-reviewer/security when the diff touches a security-lifecycle area (a precaution, not measured)

Outside a parallel run, the implementer works in the task's worktree on the task branch and stops after its report (implementer contract, step 4). The primary then runs the review, numbering, records and merge under CONTRIBUTING, Git flow. In a parallel run the contract applies unchanged.

A brief is self-contained: exact paths, values, acceptance items, commands, the rule excerpts the worker needs, which files to read, and the ledger lines and paths of its part; it asks for the contract's short report. Excerpts never replace the reads a contract requires.

Review brief

The review a tier requires is the whole set in CONTRIBUTING, Risk tiers, never a part of it. The review brief names the spec, the acceptance items, the commit range, the checks to rerun, the brief ledger and the path list, and asks the reviewer to check each ledger line and each path against the diff. For T1 and up it tells the reviewer to run local /code-review on the integrated diff where available (reviewer contract); for T3 it also asks for the blocking focused review with the convergence-review skill and /security-review where available. A step counts as NOT_RUN only with the reason the command itself gave. Do not run /code-review yourself as well, unless the reviewer reports it unavailable or not run; then run it yourself on the integrated diff. Per-task reviews in a parallel run do not replace it.

Evidence

A delegated report is a claim. Confirm every write by observed state (tests, diff, build). A reported PASS counts at the merge gate only after you confirm it from the log's last lines (the exit=<code> line the runner contract adds, and the totals where the check prints them). On a failure, read only the log's failing part. A runner brief names the worktree, the commands with their check IDs and the log directory; the runner contract holds the rest.

Escalation

  • NEEDS_CONTEXT calls for a better brief, and an environmental BLOCKED (a locked desktop, a missing restore) for fixing the cause, not a bigger model.
  • When a subagent fails the same step twice, or reports something the state contradicts, rerun that step one level up: haiku high, then sonnet high, then opus medium, then opus high. Do not retry the same level unchanged. When opus high fails too, take the step over or report it BLOCKED.
  • A failing check found by a runner goes to you or the aiu-implementer/diagnose entry for diagnosis (systematic-debugging); the runner never fixes anything.
  • Never review below the reviewer entries.
  • An unavailable model is not a failure: use the entry's next model (CONTRIBUTING, Subagent models).
  • Never run haiku below high effort (owner rule).
  • Haiku costs five times more once a prompt passes 100k tokens: keep haiku briefs small and have it read excerpts, never whole large logs or files.

Before the final report

  • Tick the brief ledger and the path list.
  • Every required review step ran, or is recorded NOT_RUN with the command's own reason.
  • Your final message is the report the task asks for, complete in itself, also when a late reminder or notice arrives after the work is done.
  • Add eco: <solo | the delegation chain> to the plan line, and one line to the final report: the delegations made, their models and efforts, and any rework.

Data freshness

When checked_at in model costs is more than three months old, or a new model generation has shipped, say so when the mode starts. Refreshing the data and the roles is its own verified change, never part of the task at hand.

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.