| name | optimize-context |
| description | Use when the user asks to reduce per-turn context or token load, trim kernels, skills, or connectors, or propagate a generic context optimization to the starter. |
optimize-context — cut per-turn context cost, propagate to the starter
The dominant token cost in long sessions is input that reloads every turn, not output prose. This skill is the playbook for finding and cutting it, plus porting the generic wins to ryanportfolio/Harness-Firmware (formerly AI-Firmware, before that claude-starter) so every future repo inherits them.
First principle: compression is lossless when you relocate, not delete. Always-loaded files become thin hooks; full detail moves to a subfile (.claude/reference/, a MEMORY subfile) read on demand. Accuracy/rules never trade against brevity — only fluff and duplication die.
The lever catalog (ranked by payoff)
1. MCP connectors — biggest single win, NOT scriptable
Each connected MCP server injects its tool-name list AND any instruction block into every turn. claude mcp list shows them. Disconnect ones this project never uses.
- Account/marketplace connectors (claude.ai → Settings → Connectors) are account-level —
claude mcp remove can't touch them (they re-sync); the user toggles them in the UI. Reversible (reconnect anytime).
- This is the user's action, not a file edit. Identify the dead weight, name it, let them disconnect. Flag if a disconnect removes a capability the project sometimes needs (e.g. Context7 live docs → say so when you'd have used it).
2. Skill visibility — skillOverrides (project/bundled skills ONLY)
The harness injects every visible skill's name + description every turn. Hide unused project or bundled skills in committed .claude/settings.json:
"skillOverrides": { "lab": "off" }
- Values (schemastore-verified):
on | name-only | user-invocable-only | off. off = gone from Claude's context AND the / picker.
- Keys are BARE skill names (the project skill's directory name).
- PLUGIN SKILLS ARE NOT AFFECTED — hard limitation, schema-explicit. Any key targeting a plugin skill (colon-prefixed or bare) is a silent no-op; 17 such keys sat in this template as dead config until 2026-07. Trim plugin skills at their source instead:
- Account-synced skills plugins (
@inline, from claude.ai) → disable the individual skills or the sync in claude.ai settings. Per-account, not a repo file.
- Marketplace plugins →
claude plugin disable <name> (all projects) or uninstall. No per-project plugin disable exists.
off takes effect next session. Scope precedence: settings.local.json > project settings.json > ~/.claude/settings.json.
3. Kernel (CLAUDE.md) + index (MEMORY.md) compression
- Caveman-ultra the prose, keep every rule exact.
- Thin hooks, detail in subfiles. A fat index/kernel line taxes every turn; the subfile is free until recalled. Push PR-history / mechanism / proof to
.claude/reference/<topic>.md (or a MEMORY subfile); leave a one-line hook + pointer. Before pointing a hook at a subfile, confirm the detail is actually there — relocate, don't dangle.
- Don't restate what the harness already injects every turn: the available-skills list, the environment block (OS/shell/cwd/git/model), tool-doc behavior. Cut them — keep only the project's value-add (non-obvious implication, version pins not in env, the project-specific rule).
- Remove untrue / stale / irrelevant rules outright (with the user's confirmation when it's a behavioral rule).
4. Verify the cut
- Measure first, then again after:
bash .claude/scripts/context-weight.sh prints the file-measurable per-turn weight (kernel + global CLAUDE.md + every skill's injected description) with per-skill breakdown. Run it before touching anything so the saving is a real delta, not a guess.
- Byte counts before/after (
wc -c) for kernel/index edits.
- For skill disables / MCP disconnects: confirm next session the items are gone (current session's context is fixed).
- Never claim a per-turn saving you didn't measure.
Gotchas (burned in)
- Durable vs ephemeral carrier: a rule that looks duplicated by a SessionStart-hook message is NOT safe to cut — hook output is a one-time early message (droppable at compaction); CLAUDE.md re-injects every turn, so the kernel is the durable home (e.g. the caveman default stays in CLAUDE.md).
skillOverrides is a silent no-op for plugin skills — no error, the skill just keeps loading. A skillUsage entry like superpowers:writing-skills proves the skill RAN, not that an override key in that form works. Verify a disable by checking the skills list in a NEXT session, never by the write succeeding.- Don't trust an agent's schema claim un-verified — confirm enum values / property names against the real schema (
schemastore.org/claude-code-settings.json) or the running config.
Propagation to claude-starter
Only the generic, portable wins go up — run them through sync-starter Direction B (it owns the genericize + apply-at-~/code/Harness-Firmware + PR mechanics):
- Portable:
skillOverrides disabling broadly-unused bundled skills (lean default; a project that needs docx/pdf re-enables it), CLAUDE.md structural conventions (caveman default, thin-hooks, don't-restate-injected-content), a leaner skill file (e.g. the trimmed caveman).
- NOT portable: account-level MCP connector disconnects (per-account, not a repo file), project-specific reference/memory content, anything named after the current project.
- Genericize first: strip project names/paths/URLs. If it can't be genericized, it stays local.
Anti-patterns
- Deleting detail instead of relocating it to a subfile.
- Cutting a rule that must persist all session (durable-carrier trap).
- Bare-naming a
skillOverride that has a project dupe.
- Disabling a bundled skill a future project might need without leaving it re-enableable (the template default is lean; document the re-enable).
- Pushing project-flavored content to the template — genericize or leave it.
- Running this on every session — it's occasional, user-triggered maintenance.
