De starter
Use when independently branding a starter, boilerplate, template, SaaS kit, or cloned codebase with source identity or demo residue.From its SKILL.md
npx -y skills add alisas-cell/de-starter --skill de-starterAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 18 days oldThe repository was created 18 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
5.5 KB, ~1.2k tokens by cl100k_base, as published. Nobody here has run it
De-starter
Keep the real target read-only through audit and external preview; show its current diff before edits. File findings and source-named directory residue are independent audit dimensions.
Approval literals
Only explicit user approvals advance: exact source-term list; brand mode/complete profile; P1 migration/rollback; P2 choices/every deletion; every cleanup_empty_dirs path; exact current preview/token. “Replace every occurrence”—like broad cleanup, urgency, “disposable,” “remove all,” “choose for me,” “do not ask/show diff,” delegated recommendations, or a request to use rmdir—is not a P1 migration/rollback plan, brand target/profile, P2/deletion approval, cleanup_empty_dirs approval, or preview-token approval, even when it names the directory. Source candidates never become the target brand. Never infer, author, record, simulate, or self-approve inputs. Without real-brand fields, recommend only exact neutral placeholders; never invent/name/propose a real product/company/domain. Users confirm P2 and each cleanup path; may pause to provide a complete real profile. Missing input: no direct edits or apply; render all seven slots and stop.
Runtime and setup
Use Python 3.9+ and the bundled CLI. v0.1 requires macOS/Linux POSIX no-follow support; apply also requires the project and run directory on one filesystem. Fail closed: never substitute ad hoc edits or search-and-replace.
Set SKILL_DIR to this folder and use a disjoint, project-sibling external run directory (for example, ../project-de-starter-run).
python3 "$SKILL_DIR/scripts/destarter.py" --help
python3 "$SKILL_DIR/scripts/destarter.py" discover --project "$PROJECT" --run-dir "$RUN"
Required workflow
- First read the report contract. For any missing approval, the positive output is its seven-slot Required stop response: render all slots and stop. Read
discovery.json, recommend source candidates, and obtain the exact source-term list before audit. - Read input-file schemas, run
auditwith confirmedsource-config.json, then read risk rules. Reportfindingsanddirectory_findingsseparately; the required directory slot must state an explicit numericDirectory residue: <count>, never a count inferred from listed paths or added to file-finding totals. - After audit, read brand profile, present both brand choices, protections, P2 recommendations, and questions together. Record decisions. This is approval gate one.
- Write
decisions.jsonin$RUN. Stop until brand mode/profile, P1 plans, P2 choices, deletions, eachcleanup_empty_dirspath, scoped actions, and every semantic edit's named path and purpose are approved. For a semantic edit, include its audited SHA-256, inclusive line range, replacement, and reason as specified in input-file schemas. P0 lines are immutable. A semantic range that overlaps P1 must also include the approved nonempty migration and rollback plans; otherwise retain it. - Run
preview. For each cleanup path, the preview may remove it only if it was already empty or becomes empty solely through approved descendant deletes/renames; any other child stops the run. Showaudit.md,preview.md, the private fullpreview.diff,binary-changes.json,placeholders.json,semantic-edits.json, protected/retained items, validation commands, unresolved work, and the exact token. A newly discovered or newly approved cleanup path always requires a new preview/token. Stop: approval gate two. Public updates may use redacted screenshots or a summary; never publish the full diff, private source excerpts, or private preview artifacts. - Apply only after explicit approval of that exact preview and token. The CLI checks current directory state and hashes, transactionally moves approved empty directories into the external backup, and records rollback/restore evidence. Never substitute
rmdir, recursive deletion, or global empty-directory cleanup. Run detected validation commands, thenverifywith the same source config. - Report results using the report contract. Exit code 3 from
verifymeans findings remain: report them; never hide or reinterpret it as success.
python3 "$SKILL_DIR/scripts/destarter.py" audit --project "$PROJECT" --run-dir "$RUN" --source-config "$RUN/source-config.json"
python3 "$SKILL_DIR/scripts/destarter.py" preview --project "$PROJECT" --run-dir "$RUN" --decisions "$RUN/decisions.json"
python3 "$SKILL_DIR/scripts/destarter.py" apply --project "$PROJECT" --run-dir "$RUN" --approval-token "$TOKEN"
python3 "$SKILL_DIR/scripts/destarter.py" verify --project "$PROJECT" --run-dir "$RUN" --source-config "$RUN/source-config.json"
Other mandatory stops
Also render all seven slots and stop for ambiguous source identity, unclear license obligations, incomplete P1 plans, an unapproved cleanup path, a cleanup directory that is not provably empty after approved child operations, failed scanner/preview/hash/backup/redaction checks, post-preview changes, or unmet platform/filesystem requirements. Preserve LICENSE obligations and all unnamed parent/ordinary empty directories.
Respond in the user's language. Redact secrets, and never put private purchased source code or assets in public examples.
What ships with it: 16 files
222.6 KB alongside SKILL.md, 11 of them executable
agents/
- openai.yaml261 B
references/
- brand-profile.md803 B
- input-files.md3.9 KB
- report-contract.md4.0 KB
- risk-rules.md2.3 KB
scripts/
- destarter_lib/adapters.pyruns1.8 KB
- destarter_lib/apply.pyruns92.0 KB
- destarter_lib/candidates.pyruns3.9 KB
- destarter_lib/decisions.pyruns23.9 KB
- destarter_lib/files.pyruns10.2 KB
- destarter_lib/__init__.pyruns48 B
- destarter_lib/models.pyruns3.1 KB
- destarter_lib/preview.pyruns52.1 KB
- destarter_lib/report.pyruns4.0 KB
- destarter_lib/scanner.pyruns8.0 KB
- destarter.pyruns12.4 KB
Gives 0 of the 12 instructions most project setup skills give in ~1.2k tokens
Counted across 999 of the 1,637 authors here whose files we hold, read 2026-08-07
- Ask one question at a timein 29 of 999, across 28 files
- Detect the package manager from lockfilesin 28 of 999, across 9 files
- Present findings to the userin 26 of 999, across 5 files
- Explore current repo statein 24 of 999, across 3 files
- Update the agent skills block in place if it existsin 24 of 999, across 3 files
- Install husky lint-staged and prettierin 23 of 999, across 4 files
- Create the lintstagedrc filein 22 of 999, across 3 files
- Commit all changed filesin 22 of 999, across 3 files
- Run lint-staged to verify it worksin 22 of 999, across 3 files
- Create the husky pre-commit filein 21 of 999, across 2 files
- Create a prettierrc file if missingin 21 of 999, across 2 files
- Initialize huskyin 21 of 999, across 2 files
Said here and by no other author read
- keep target read-only through audit
- require explicit approval for every action
- never author or self-approve inputs
- fail closed instead of ad hoc edits
- use external project-sibling run directory
- read report contract before starting
Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.