Slingshot

Slingshot

Small, precise input; outsized result. The operating principle: spend expensive tokens once — on specification, routing, and learning — so every future token is cheaper. Intelligence placed in the prompt substitutes for intelligence in the runtime: a weaker model with a dense spec beats a stronger model with a vague one.

Two layers

1. Always-on layer — a managed rule block injected into CLAUDE.md (and AGENTS.md for other agents) between SLINGSHOT:BEGIN/END markers by the installer. It carries the economy rules and auto-triggers on every request with ~20 lines of standing cost. Source of truth: rules/rule-block.md.
2. Deep layer — this skill: full protocols, loaded only when a move or command actually runs.

An optional SessionStart hook injects the ledgers into every session automatically (see references/manage.md).

Setup gates (non-optional)

Gate	Check	If fail
Installed	The SLINGSHOT:BEGIN marker exists in the active CLAUDE.md (run status if unsure).	Offer install once; proceed either way.
Reference	The matching reference file is loaded when a command runs.	Load it before continuing.
Ledgers	Before DISTILL/ROUTE on significant work, both ledgers were consumed this session.	Run the RECALL move first.

Ledgers (the skill's memory)

Ledger	Path	Survives updates
Generic	~/.claude/slingshot/generic.md — portable techniques, any project	Yes — lives outside the skill folder
Project	.claude/slingshot.md at the project root — this repo's quirks	Yes — lives in the repo; commit it

The skill ships learnings/seed.md; the installer copies it to the generic ledger path only if none exists. Updates replace the skill folder wholesale and never touch ledgers.

Commands

Invoked as /slingshot <command> [target], or auto-triggered per the rule block. Load the reference before executing the command.

Command	Category	Description	Reference
distill [request]	Optimize	Translate vague intent into an executable Spec Block	references/distill.md
route [task]	Optimize	Pick minimal sufficient model tier + delegation prompt	references/route.md
budget	Optimize	Audit and correct this session's context economy	references/budget.md
arsenal	Optimize	Recommend vetted external token-saving tools for the user's workflow	references/arsenal.md
learn	Flywheel	Extract session learnings into the ledgers	references/learn.md
recall [topic]	Flywheel	Consume ledgers; surface entries relevant to a topic	references/recall.md
teach	Flywheel	Repo scan + one interview → seed the project ledger	references/teach.md
status	Manage	Version, install state, rule block, ledger stats	references/manage.md
install / update / uninstall	Manage	Run the installer script	references/manage.md

Bare /slingshot with no command → run status, then suggest the most relevant command for the current situation.

Move summaries

DISTILL — extract GOAL / SCOPE / CONSTRAINTS / NON-GOALS / DONE WHEN from the user's words plus the repo. Infer before asking; at most one batched question. The Spec Block is the payload: it becomes the prompt for whoever executes.

ROUTE — capability gap is compensated by specification density, and routing is two-dimensional: tier × effort (frontier low often beats prior-gen xhigh — down-effort before down-tier). Haiku for mechanical pattern-following with exact examples; Sonnet for implementation from a full spec; Opus/Fable for ambiguity, architecture, and verification. The compound pattern: strong model distills → cheap model executes → fresh-context script or strong model verifies. One failure → fix the spec; two → escalate; never loop.

BUDGET — control what enters the window, what stays, and what always loads. Files are durable memory; the window is scratch.

LEARN / RECALL — after significant tasks, distill 0–2 one-line learnings into the right ledger; before significant tasks, consume them. Capped, deduped, falsified entries deleted on sight.

Anti-goals

• Never trade correctness for tokens. A wrong answer is the most expensive output there is.
• No ceremony on trivial tasks. Distilling "fix this typo" burns more than it saves.
• No visible bureaucracy. Apply the moves silently; surface them only when the user asks about cost or model choice.