Skip to content

tmusser/ai-engineering-skills / skills/eco-mode

eco-mode

Verified

Keep a small, budgeted working context for low-cost model execution without weakening task contracts, required sources, or verification.

Eco Mode

Purpose

Make bounded agent work economical through source selection, fresh-context discipline, and a measurable preflight. This is an opt-in working-set policy, not a second workflow, model switch, or replacement for verification.

When to use

Use when the operator invokes eco mode, asks to minimize context costs, or runs a bounded task on a fast inexpensive model. Especially useful for maintenance, well-scoped coding, extraction, and small-model delegated work.

Do not force it on ambiguous research, incident response, high-stakes decisions, or tasks whose required evidence cannot fit. Escalate when necessary.

Inputs

  • Objective, current source of truth, scope and acceptance/proof targets
  • Only the selected relevant code/docs and durable state files
  • Optional provider/model choice outside this skill
  • Local selected-content budget, default 60,000 UTF-8 bytes
  • Optional, explicitly named required source files and literal proof anchors

Workflow

  1. Resolve the route first, or honor an already explicit route. Do not append ceremony to a sufficient workflow.
  2. Load only the current authority-bearing task sources, relevant file sections, and next proof target. Search and expand when a named risk warrants it.
  3. Optionally use existing context hydration for Markdown: python scripts/aes.py context "bounded task" --strict > /tmp/eco.md. Inspect the packet status and required-file coverage before relying on it.
  4. Check the actual selected text files before sending them: python scripts/aes.py eco /tmp/eco.md, or pipe the selected prompt to python scripts/aes.py eco -. The default combined ceiling is 60,000 bytes; REVIEW_REQUIRED means prune redundancy or explicitly expand.
  5. For tasks with a named context contract, use the optional sufficiency check: python scripts/aes.py eco SPEC.md VERIFY.md --require-source SPEC.md --require-source VERIFY.md --require-proof "targeted pytest command". The preflight checks actual selected files or source-verified context-packet excerpts. Metadata mentions alone do not establish representation. If a requirement is missing, REVIEW_REQUIRED means add the minimum necessary source or escalate; do not silently waive it. When a particular section governs the next action, add --require-heading "SPEC.md::Acceptance criteria" so a metadata mention or unrelated excerpt cannot stand in for the section. If the operator needs recorded verification evidence rather than a named test target, opt into --proof-mode recorded with a selected VERIFY.md. It matches exact command-backed PASS records with provenance, but does not independently run tests or validate current Git workspace freshness.
  6. Execute one slice with existing scope controls and verification. Avoid redundant tool-output envelopes and needless refetches. Reuse anchors.
  7. Preserve required instructions, exact file paths, uncertainties, acceptance criteria, failed tests, compatibility checks, and verification evidence. Never abbreviate a failing proof into apparent success.
  8. On scope drift, missing evidence, or budget pressure, retrieve the smallest missing authoritative source. Use a trusted handoff into a fresh context, or switch to a more capable model when required. Never silently truncate.
  9. Stop after required proof. Measure actual provider token usage and spend where available; the byte checker cannot estimate billed usage.

Model guidance

  • Claude Haiku 5.5 (claude-haiku-5-5) has lower API pricing for prompts up to 100k tokens than for prompts above that threshold.
  • GPT-6 Luna (gpt-6-luna) supports reasoning.effort: xhigh through the Responses API. Deeper reasoning may spend more billed output tokens.
  • Both models support very large contexts; this mode optimizes cost and decision saliency, not merely whether a prompt fits in the window.
  • Provider limits, identifiers, tokenizer behavior, and pricing change. Check provider documentation before production routing.

Outputs

No extra ledger by default. On an explicit status request, output only the selected sources, preflight verdict, preserved proof target, and next action.

Sufficiency checks are opt-in. Source representation means the complete source was selected as a file, or a packet contains source-verified excerpts; it does not prove that the right sections were selected. Proof targets are literal text anchors, not test execution, semantic correctness, or a guarantee of completeness. An optional recorded-proof mode checks matching schema-v1 PASS command records selected from the live verification source, but cannot attest execution or semantic adequacy. Packet WARN/FAIL, stale excerpts, missing sources, and missing anchors require review. Never trust metadata-only mentions.

The preflight's PASS applies to selected UTF-8 bytes only, not full runtime history, tool schemas, provider tokens, or pricing. It never removes context already injected into a live session.

Stop conditions

  • Required sources are represented or the escalation is explicit.
  • Context preflight was inspected and verification remains intact.
  • No redundant summary cycle was created solely for token savings.

Anti-patterns

  • Assuming a 1M context window means all tokens are cheaply priced.
  • Treating 60,000 bytes as 60,000 model tokens or a guaranteed cost tier.
  • Loading all installed skills or the full repository by default.
  • Dropping risk, permission constraints, or verification to fit a budget.
  • Pretending a summary erased previously supplied context.
  • Silently truncating required files, reclassifying unknowns as irrelevant, or keeping a cheap model on a task beyond its reliable capabilities.
slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 196kUpdated Claude CodeCodexVerifiedInstall