template-loop-compress

Always run this skill in an Opus subagent

Whoever triggers this skill delegates the actual work to a subagent (Agent tool) with model: opus. Never compress inline in the model currently running — Sonnet, Haiku, or any other loop worker that happens to hit this file hands the job off. This mirrors .loopboard/LOOP.md Rule 14, which routes New-task grooming to a subagent of the task's groomer: model instead of running it in the main loop.

The subagent does everything below: read the file, compress, run the self-check, and — only on a pass — write the file and land it on a task/** branch (a PR is optional, Rule 7).

Why this file is different from ordinary docs

media/template-loop.md is not documentation. It is the upstream source of the standing instructions autonomous loop workers execute. src/controller.ts:322 reads it, and store.syncTemplates / store.autoHeal (src/store.ts) push its marked sections into every user's live .loopboard/LOOP.md via syncMarkedSections (src/sync.ts) on every Sync and on activation auto-heal. A compression of this file is not local doc drift — it rewrites the instructions every running loop reads next. "No information loss" is a correctness requirement, not a style nicety: a dropped negation, a softened "never", a lost scope qualifier, or a renumbered rule is a behavioural regression in every loop that reads the result.

Untouchable, verbatim

Never alter, reorder, renumber, or drop any of the following. If a proposed compression would touch one of these, leave that span byte-identical and compress only the prose around it.

1. The six sync marker comments, in order and unmodified: <!-- loopboard:sync:loop-intro:begin/end -->, task-index-format, task-file-format, done-index, rules, automation. A missing or reordered begin/end pair breaks extractBlock (src/sync.ts) and forces a full-overwrite sync path instead of a marked-section merge — both mutate live user files.
2. The ## Automation heading and the first fenced code block inside it. src/loop.ts matches /^##\s+Automation\b/i and requires a fenced block in that slice to build the ~200-char bootstrap prompt that spawns every loop terminal. That block's text must also stay apostrophe-free (it rides as single-quoted shell argv).
3. Rule numbering 1–16 (the count today — preserve whatever the INPUT has; never renumber or invent a rule to reach a particular range). CLAUDE.md, TODO.md entries, task files, and LOOP.md itself cross-reference rules by number ("Rule 16", "Rule 9"). Renumbering or merging rules breaks every reference silently.
4. The task-index-format and task-file-format fenced code blocks. Field names, their fixed order, the phase vocabulary, and the heading names are the contract parser.ts / writer.ts / taskfile.ts implement.
5. Every normative clause. MUST / NEVER / ONLY statements, exact counts ("at most ONE", "up to 3"), date-field semantics, and literal paths/filenames.

Code fences, inline code spans, and URLs anywhere in the file are byte-identical in output — never reflowed, reworded, or re-cased.

Compression technique

Terseness rules adapted from the caveman skill's "full" intensity tier, applied to the redundant prose only — the untouchable list above outranks every rule here.

Drop:

• Articles (a/an/the).
• Filler: just, really, basically, actually, simply.
• Pleasantries and hedging.
• Restated explanations that say the same thing twice; separate sentences stating one constraint merge into one.
• Wordy transitions and throat-clearing.

Allowed:

• Sentence fragments.
• Short synonyms — "big" not "extensive", "fix" not "implement a solution for".
• Standard well-known tech acronyms (DB, API, HTTP, PR, CLI).

Never:

• Invent new abbreviations (cfg, impl, req, res, fn). The tokenizer splits them the same as the full word: zero tokens saved, reader still decodes. Full word is cheaper AND clearer.
• Use causal arrows (→) — own token, saves nothing.
• Add decorative tables or emoji.
• Narrate tool calls or add meta-commentary.
• Add a word to sound terse. Compression only; style never grows output. No inserted pronoun or copula to fake broken grammar. Keep the correct verb form when it costs the same.

Never remove, soften, or reorder:

• A negation ("never", "not", "no", "except") — a flipped meaning is worse than any token saved.
• A scope qualifier ("only", "exactly", "at most", "regardless of").
• A count, unit, date-field name, literal path, filename, or identifier — exact.
• A technical term — exact.
• An example that a rule or piece of code depends on for meaning.

Compressing "never promote a New task yourself" into "never promote New" is fine; dropping "never" is not.

Writing the result: the self-check gates the write

The self-check below is a SOFT, pre-commit-style step — advisory, run before committing, enforced by no hook and no CI — but it is a HARD gate on this skill's own write.

Pass: write the compressed text to media/template-loop.md and land it through the normal repo flow — fetch main, branch off main, commit and push to a task/** branch. No chat-level approval needed first. .loopboard/LOOP.md Rule 7 forbids committing to main and makes the PR OPTIONAL: open one when review or a CI artifact comment is wanted, otherwise the branch is the delivery.

Fail: write nothing. Report the failing check(s) and stop. No partial write, no "mostly passing" candidate, no backup-and-overwrite.

The discipline is non-negotiable because of propagation: this file's marked sections are pushed into every user's live .loopboard/LOOP.md by store.syncTemplates / store.autoHeal (src/sync.ts). A bad compression is an unattended edit of other people's agent instructions. State this plainly rather than assuming it away: the PR used to be mandatory precisely because it was the human review point for that exposure. With Rule 7 relaxed, the self-check is the only automated gate before the branch, and human review shifts to whenever the branch is reviewed or merged to main. What still holds it: the self-check remains a hard write gate (fail → write nothing), and marked sections reach OTHER workspaces only after a release plus a manual Sync (activation auto-heal is create-only).

Self-check before writing

Before writing anything, verify against the input:

• All six marker comments present, unmodified, and in the same order.
• The ## Automation heading and its first fenced block are intact, byte-identical, and still apostrophe-free.
• Rules are still numbered exactly as the INPUT numbers them (1–16 today), with no gaps, merges, or reorders. A mismatch means fix the OUTPUT — never add a rule to reach the number here.
• The task-index-format and task-file-format fenced blocks are byte-identical to the input.
• Every code fence, inline code span, and URL in the input still appears, byte-identical, in the output.
• Spot-check a sample of MUST/NEVER/ONLY clauses and counts from the input against the output — each must still be present with the same force.
• Report a bullet-count / line-count delta as a sanity signal (a large unexplained drop in bullet count is a smell, per the caveman-compress project's own regression history — see below).

If any check fails, do not write the file and do not commit anything — report the failing check(s) and stop rather than paper over the gap. Structural validation (fences/headings/markers preserved) is not the same as semantic validation (meaning preserved); this self-check must reason about meaning, not just shape.

Provenance

The compression technique above is adapted from the caveman skill's "full" intensity tier (https://github.com/JuliusBrussee/caveman) — its drop-articles/filler/hedging rules, fragments, short synonyms, no-invented-abbreviations reasoning, and no-causal-arrows rule. Adaptation, not a vendored dependency: only the terseness rules carry over, layered under this skill's own untouchable list, which always wins. Caveman's ultra and wenyan tiers, and its persistence / language / boundary machinery, are out of scope — that is a chat-response style skill, this is a document compressor.

The validation discipline is informed by a read-only audit of the caveman-compress project (https://github.com/JuliusBrussee/caveman/tree/main/skills/caveman-compress): its lessons on preserving fences/headings/paths, and its recorded false-pass (structural checks passing while prose was reordered) and dropped-path regression are the reason the self-check above insists on semantic checks, not just structural ones. No code, script, or dependency from either project is vendored, installed, or executed — both are technique references only. This skill is markdown instructions, run by the Opus subagent, matching this repo's Docker-only, zero-runtime-dependency rules.