Ultraskill

Ultraskill

Four pathways, one router. The router reads the session and picks; the pathway file it loads carries the full procedure. Nothing in this file mutates anything; the pathway files do, and only after the undo in the shared blocks exists.

Neighbours, so the description above does not claim their work: bare "create a skill" with no improvement asked for is skill-creator's; "improve this" with no creative and no skill in context is /improve's (session-driven repair of tooling).

Step 0: Detect

Five probes, all from the session. Ask the user nothing yet.

0. Verb. The user named a pathway: polish / critique / make this better → polish; heal / fix and prevent / output is broken → heal; evals / benchmark / pass rate / optimize the skill → learn; create and improve / new skill → create.

1. Output. An explicit path argument that is not a skill directory (no SKILL.md at that path); else the most recent generated file in the conversation; else the most recent creative text (copy, script, prompt, page).

2. Producing skill. Named by the user ("heal the carousel skill") or given as a path; else a skill invocation earlier in the conversation that preceded the output; else this path table, the weakest evidence, and it counts only when <root>/<skill>/SKILL.md exists in a skill root (a hit on a skill that is not installed leaves probe 2 unresolved):

   Output pattern	Likely skill
   output/ads/ or a nano-banana command	ad-creative-graphic
   carousel slides, or carousel in the path	carousels (or carousel-post where that is the installed one)
   output/banners/ or banner dimensions	social-media-banner
   output/thumbnails/ or a YouTube thumb	youtube-thumbnail
   beat sheet or kinetic text	kinetic-text-ad
   email copy or a nurture sequence	lgj-email-marketing
   script with timestamps	short-form-script or youtube-script

   A skill root here is either home root or the current project's own .claude/skills. Remember which rung resolved it. This table is the only copy; references/heal.md points here.

3. Evals. <skill>/evals/evals.json exists for the skill probe 2 resolved, and probe 1 found no output. An output in hand with a resolved producing skill is heal even when that skill ships evals; otherwise any skill with evals could never be healed.

4. Create ask. "new skill", "create a skill", "skill for X".

Also recover the original brief (the request that produced the output). Heal and polish need it to regenerate. If it cannot be recovered: when the one question below fires, it is asked inside that question; when it does not, the only question is "What was the original request?" Either way the router asks at most once.

Routing

First match wins:

Two preconditions and one fallback, checked before anything is announced:

• Deny list, mutating pathways only (heal, learn). Never target skill-creator, a skill directory that is a symlink, or a directory under a plugins, vendor or plugin-cache path of either skill root (~/.claude/skills for Claude and ~/.agents/skills for Codex are the two roots; neither is a git repo). Resolve a name against the root this skill was loaded from first, the other root second; an explicit path outside both roots is allowed and the same tests apply to it. A target on the list stops the run: "Not touching <skill>: it is <on the deny list because ...>", naming every reason that applies. Every item on this list is a mechanical test; nothing here is a judgment call.
• The one question. When heal and polish are both supported and probe 2 resolved only by the path table, ask once, verbatim: "Heal <skill> (mutates the skill, then regenerates) or polish the output only (skill untouched)?" and, if the brief could not be recovered, append "And what was the original request?" Unattended (no answer), take polish: it mutates nothing.
• Nothing resolves. "Nothing to work on. Generate something, name a skill, or say what skill to create." Stop.

After the preconditions pass, announce the pathway in one line, then load its reference file and follow it. Do not load the others.

Shared blocks

Every pathway file points here instead of restating these.

Undo, established before the first mutation

Two copies in the scratchpad, taken unconditionally, repo or not:

IT0="<scratchpad>/<skill>-SKILL.md.iter0"     # immutable: the pre-loop file, what "undo" means in any report
ROLL="<scratchpad>/<skill>-SKILL.md.revert"   # rolling: advances on every kept mutation
cp <path-to-SKILL.md> "$IT0"; cp <path-to-SKILL.md> "$ROLL"
test -s "$IT0" && test -s "$ROLL"

If either copy is missing or empty, the loop does not start. A missing undo is a stop, not a warning.

Then decide the per-iteration revert. Git is the revert only when all three hold: git -C "$(dirname <path>)" rev-parse --git-dir succeeds, git ls-files --error-unmatch <path> succeeds, and git status --porcelain -- <path> prints nothing. Otherwise the rolling copy is the revert (cp "$ROLL" <path>), and the report says "undo: file baseline" so the reader knows git revert is not available. Never commit pre-existing changes in someone else's repo on the loop's behalf.

Mutation ladder (full catalogue in references/mutation-strategies.md)

One targeted change per iteration. Place a new rule near the top of its section (primacy); put a negative constraint right after the positive rule it reinforces. Escalate only when the previous level was reverted for the same failure. Bloat guard: past 2x the original byte size with under 10% score gain, the next iteration is a consolidation pass, not a mutation.

When a mutation targets a rule that already existed, do not restate the rule. Write down the argument that beat it, verbatim, and the counter. A rule that has already lost once loses again to the same sentence.

Banned words

references/banned-words.md: one shared list plus a copywriting overlay, U+2014 and U+2013 both banned, and the recorded pointer to the agent instruction file's own list.

Commits

Write the message to a file, then git commit -F <path>. Never -m (backticks are command substitution and the term vanishes from the message), never a heredoc (dies on the first apostrophe). Both fail with exit 0.

Images

sips --resampleHeightWidthMax 1000 "<image>" before any Read of an image. Claude crashes at 2000px with several images in context.

Cost

Keep a running estimate of generation cost. Past 2.00 USD, pause, report the breakdown, stop. Never write a dollar sign followed by a digit in a skill file: it loads as an empty string or as the skill's second argument, and a cost guard written that way read as "0.00" in one session on 2026-09-09. Write the currency after the number.

Autonomy, once a loop begins

Never stop to ask whether to continue, whether to commit or revert, or whether this is a good stopping point; the stop table below is the only stop. Never pause between iterations. On an unexpected error (git, file read, generation), attempt recovery three times, then log it and continue to the next iteration. The human may be away from the keyboard. The one question in Routing is asked before the loop starts, never inside it.

Stop conditions

Status line after each iteration, one line, in the pathway's own shape ([Heal N] ..., [Learn N] ..., Pass N: ...).

Rules that survive any consolidation pass

Marked load-bearing in the pathway files. Do not trim them for length.

• heal V11: a flaw you already noticed and argued away is a defect you declined to file.
• heal V12: dead space in a frame that will be cropped.
• heal T11: an unsourced number is a defect on anything a stranger will read.
• heal T12: register that misses the stated audience.
• "Write down the argument that beat the rule" (Mutation ladder, above).
• The skill roots are not git repos; the file baseline is the undo there (Undo, above).

Files

SKILL.md                              this router
references/polish.md                  output-only critique and regenerate
references/heal.md                    output to skill-gap to mutation loop
references/learn.md                   eval-driven mutation loop
references/create.md                  skill-creator handoff, then learn and heal
references/mutation-strategies.md     L1-L5 in full, selection algorithm, anti-patterns
references/assertion-patterns.md      assertion library per domain (learn)
references/blatant-defect-catalog.md  expanded V/T checklists with false positives (heal)
references/defect-to-skill-mapping.md defect category to SKILL.md gap (heal)
references/banned-words.md            the one list
references/graphic-ad-references.json reference ads per type (polish, graphic ads)
evals/templates/*.evals.json          general, copywriting, graphic-design (learn, create)
scripts/score_report.py               prints a learn workspace summary