| name | using-slipstream |
| description | Use whenever the user says "use slipstream", "use superpowers", or starts any task in a project that has slipstream. A hard, always-on discipline that forces scoped reads over whole-file reads to save tokens, recalls memory before acting, and records what is durable every turn so memory grows constantly. |
| slipstream | {"category":"context","requires":[],"tags":["discipline","memory","token-efficiency"]} |
Overview
slipstream only saves tokens and builds memory if the agent actually uses it. The dashboard showing "0% optimised" means whole files were read instead of scoped slices: a failure, not a neutral state. This skill is a rigid discipline, not a suggestion. When the user invokes slipstream, follow every step. Do not read a whole file when a symbol or a line range will do. Do not finish a turn without recording what is durable.
Steps
- Recall first, always. Before reading any code, pull relevant durable facts and the last session's summary (
/slipstream:recall). Never start from zero on something a past session already decided.
- Read the map, never the whole file. Refresh and read
.claude/slipstream/map.md (/slipstream:map), then pull exactly what you need with sp_symbol (one symbol) or sp_lines (one range). Reading an entire file when a slice would do is a defect; the optimisation tally must not stay at 0%.
- Search scoped, not broad. Use
sp_search over the map and sp_search_memory over past work before grepping the tree.
- State the task in one line and make the smallest change that satisfies it. No unrelated refactors.
- Record durable outcomes every turn. When the turn produces a decision, convention, gotcha or credential location, save it with
/slipstream:remember. Memory should grow constantly, not only when asked.
- Coordinate with other tabs. If the session-start context lists other open sessions, build on their work; do not duplicate or undo it.
Verify
Confirm three things before claiming the turn is done: you recalled memory before the first read, you used scoped reads (sp_symbol/sp_lines) rather than whole-file reads so the optimisation tally moved above 0%, and any durable outcome is saved (npx slipstream memory list --root . shows it). If you read a whole file when a slice would have worked, say so and correct the habit.
