About tk-pr-media
A skill for handling a bounded upload of local image or selected video evidence to an existing GitHub PR, then updating its body or a selected comment. It emphasizes validating evidence and target authorization, verifying uploaded assets, and reporting a precise outcome.
Key features
- 01Validate and stage local evidence
- 02Upload repository-scoped GitHub attachments
- 03Update a selected PR body or comment
- 04Verify Markdown tokens and uploaded assets
- 05Record attempts in a single receipt
- 06Handle blocked and unverifiable outcomes
Use cases
- 01Attach screenshots to an existing PR
- 02Publish selected video evidence for a PR
- 03Fulfill a required evidence handoff
- 04Verify uploaded media before reporting completion
Install
Copy the command and run it in a terminal.
npx skills add https:/ / github.com/ mtgvim/ tiger-kit --skill tk-pr-mediaAfter installing, type/tk-pr-mediain Claude Code to confirm it loaded.
Bundled files
References and scripts installed along with SKILL.md.
LICENSE.txtagents/openai.yamlreferences/artifact-paths.mdauthentication.mdcdp-fallback.mdgh-attach.mdquestions.mdvideo-upload.mdDetails
From the project READMEGitHub PR Media Upload
Handle explicit local image/video evidence requests for an existing GitHub PR or an exact
tk-pr-open generic manifest handoff with evidence_required: true. Only the parent handoff is an
automatic trigger.
Optional motion supplements
For a selected MP4, read native video attachments; it owns the distinct upload/byte/player checks instead of the image-only extension steps. A valid optional supplement carries media type, run-owned path, criteria/head, duration/viewport, inspected frames and sampling warnings; never require it to impersonate the required image manifest. Upload it only when publication_selected: true or the user explicitly selected that exact artifact/target.
For selected GIF frame assembly, use the reviewed extension with --gif, an explicit .gif name and delay matching the frame sequence. Capture stderr separately without --json at v0.7.0. Preserve all frame-cap/reduction/oversize warnings, verify actual image/gif bytes and multiple decoded frames, and inspect animation when claiming playback. The 5MiB target is not a hard limit; disclose oversize and altered timing. Existing required screenshot pairs remain unchanged.
Scope
Handle one bounded upload: validate and stage local images, upload repository-scoped GitHub attachments, minimally update the PR body or selected comment, and verify the source Markdown and uploaded image objects. The default target is the PR body.
For tk-pr-open, validate the producer-neutral evidence fields and never branch on, whitelist, or require a producing
skill name. If required evidence is missing or invalid, mark the evidence handoff Blocked; do not revert an
already-created tk-pr-open PR.
Use references/gh-attach.md to classify and run the
extension route. Read references/cdp-fallback.md
only after CDP is selected or the extension route cannot use target write
capability. Invoke tk-wizard only for an exact human-only host setup,
authentication, permission, or restart step; preserve this skill's upload target and route
across the handoff. Use tk-browser-verify for browser-controlled runtime verification.
🔴 CHECKPOINT · 🛑 STOP · Upload mutation boundary
Before any upload or PR body/comment update, reverify the explicitly selected target, valid local evidence or producer handoff, authenticated route, and requested write scope; stop as Blocked or Unverifiable when those preconditions cannot be established.
Workflow
- Confirm
origin, the current PR, and the executing repository from regular local image file(s) or a valid producer evidence handoff. Read the existing PR body. - Reuse the existing
## 스크린샷heading. If absent, insert it before the AI footer or append it at the end. Preserve unrelated body content. For manifest artifacts, use the supplied criterion, role, exactdisplay_route, state/region, viewport, and comparison in the caption; never derive them from a filename, OCR, raw URL, or producer identity. Keepvisual-preservationbaseline and after visibly labeled as one pair. For every pair, with or without a manifest, add a title naming the verified route/screen, one expected-behavior sentence, role headers such as수정 전/수정 후, one image row, and a caption row beneath describing each image's observation. Carry supplied evidence; disclose missing route/screen evidence rather than inventing it. Staging filenames are not headers, captions or an adequate description. - Select and execute an eligible route using gh-attach, or read CDP fallback only after CDP is selected. Repository visibility is not a routing signal, and a route that may have created remote state cannot silently fall back to the other route.
- When the selected route needs a human-only host action, pass a bounded handoff to
tk-wizardand perform no upload until its completion signal. Agent-approved reviewed extension installation remains here. Resume the exact repository, target, image set, and route after the wizard completes. - Update only the requested body or comment through the GitHub API or equivalent.
Before
Pass, reread the target and verify every generated image token appears byte-for-byte exactly once, the pair roles/captions remain correctly associated, decodableimage/*bytes from an authenticated fetch for every asset, and reachability of every linked upload object from the upload ref. Browser-session rendering is optional; report whether it was checked in the receipt without making it aPassprerequisite. - Remove only owned staging files on every exit path. Write each deletion with the literal, fully resolved absolute run-owned staging path already recorded in this run; never derive the target from command substitution, a path file read back, or a glob piped through
headortail. If the exact literal path is unknown, list the candidates and stop instead of deleting.
Execution Receipt · Single Evidence Record
Record every upload attempt in the single receipt below. It is not a separate lifecycle
output; it is the sole evidence record used for the ## GitHub media upload result and
approval/verification decisions. Use none or unavailable for absent values; never guess.
Repository: <owner>/<repo>
PR / Target: <pr-number> / body | existing comment <comment-id>
Source images: <absolute path list>
Route: gh-attach | CDP
Entry: <exact reviewed command | browser route>
Generated Markdown: <asset refs | none>
Remote ref: <refs/uploads/issues/<pr-number> | none | unknown>
Verification: image-token byte match and single occurrence: <Pass | Fail | unavailable>; image bytes/type: <Pass | Fail | unavailable>; upload ref reachability: <Pass | Fail | unavailable>; rendering: <checked | not performed | unavailable>
Changed: <body/comment changed | unchanged | unknown>
Cleanup: <owned staging path removed | failed | not applicable>
Status: Pass | Fail | Pending | Blocked | Unverifiable
Evidence Manifest Handoff
When evidence_required: true, require all of the following:
verification_status: Pass, a non-empty criterion, evidence kind, comparison result, and limitations.- Every artifact has a role, non-empty image, absolute path, run-owned evidence directory, state/region, and viewport.
- Every image is marked
inspected: trueand has an exact origin-freedisplay_routeor an explicit safe omission limitation. Never reconstruct or mask a missing route. visual-preservationcontains both a baseline and after artifact from its verified pair.
Reject arbitrary screenshots, missing paths, non-Pass results, uninspected artifacts, incomplete preservation pairs,
and artifacts not tied to the current run. If required evidence is missing or invalid, return Blocked before upload.
Prohibitions
- Do not create or merge a PR, change reviewers, publish a release, or insert a comment unless the user explicitly selected that comment target.
- Do not bypass
tk-browser-verifyor use Orca or screen control as an automatic browser fallback. - Do not overwrite a pre-existing draft or click comment, close-with-comment, or any other submit button.
- Do not choose a route based on public/private visibility or claim success from only a placeholder, Markdown presence, fixed delay, or unrendered API response.
- Do not log or return a signed URL, JWT, or raw secret-bearing query string. This does not prohibit a verified-safe,
origin-free
display_route; preserve its pathname, query, and fragment exactly as supplied by the manifest. - Redact only secret-bearing values (signed parameters, JWTs, tokens) in displayed output; preserve rendering parameters
such as
raw=true. Keep each generated image Markdown token byte-for-byte, including its alt, URL and rendering parameters; surrounding table layout may be recomposed with captions beneath the images. Never rebuild links from redacted or summarized output. Do not copy secret-bearing originals into logs, receipts, or evidence artifacts.
Failure Handling
Return Pending while waiting for an absent-extension or unreviewed-upstream choice or a
human-only tk-wizard completion signal. Return Blocked for unknown provenance, a draft, or a rejected
installation without a selected CDP route. Return Unverifiable if authentication,
image-byte validation, or upload-ref reachability cannot be verified; rendered-page
evidence is optional. Return Fail for upload, remote ref
verification, or cleanup errors. State whether the selected body/comment changed
and whether an upload ref may remain.
For tk-pr-open, preserve the separate PR operation result and return the
evidence state as uploaded or blocked. While required evidence remains
blocked, the parent cannot claim full completion.
Result
Start the terminal response with ## GitHub media upload. Include ## Uploaded,
## Verification, and ## Cleanup when applicable. End the result-owning section with exactly
one Status: Pass|Fail|Pending|Blocked|Unverifiable line. Expose asset URLs only when safe and
redact signed parameters.
Artifact Paths
Artifacts require task authority. Before writes/checkout/transport/ignore setup, read artifact paths for Git exclusion/path/ownership checks. Reuse identical in-context content; hash-check uncertainty and reread changes/loss; always recheck safety. Use .tigerkit/ unless a destination is explicit. Chat-only: skip file/ignore setup and this reference. This grants no unrelated mutation/publication authority. Prefer fresh directories and supported trash for authorized non-secret cleanup; target only recorded literal resolved absolute run-owned paths, never cd plus glob/relative deletion.
Reader-facing Result Summary
Return a brief result with the outcome, actual verification, material limitations and direct deliverable links. Distinguish plans/skipped checks from executed work; retain outstanding manual checks and real cleanup, blocker or human-action details. Never claim unverified completion or cleanup. Preserve the owner's required artifact, publication preview, approval boundary, machine return, exact status/footer and final question block; keep nested returns compact and omit empty sections.
Output Notation
Use ASCII numbering (1. Item, (1) Item) and - Item, not enclosed/keycap numerals.
Show complete absolute clickable file paths, including examples, without ellipses.
Preserve exact code, commands, identifiers, URLs, quotations and verified UI labels
unless their edit is authorized.
For chat comparisons with multiple shared attributes, use narrow GFM pipe tables (about four columns, header, short cells); otherwise use lists. No box-drawing, fenced or empty tables. Owner-specific previews take precedence; publication bodies and file artifacts retain their own table formats.
In rendered Markdown prose, use hyphens for ranges and escape literal tildes as \~;
preserve exact code, quotations and identifiers.
For authorized temporary input files, follow Artifact Paths before creation and consumption; a prefilled template is not completed input. This grants no write authority.
User Questions
Investigate facts; reuse authorization. For user-owned decisions read
question rounds; otherwise proceed.
Plain chat, even one approval: standalone ---, then
❓ **Q<n> · <title>**: <question> and ➡️ <recommendation> per item.
Ask only settled-prerequisite questions; separate approval.
Questions come last; put status, risks and plans before them.
Skill Feedback
For a reusable incident, retain minimal non-secret evidence in the current conversation
and suggest tk-retro; do not write it to persistent memory unless the user asks.
Invocation, feedback artifacts, installed-skill edits and publication require separate
authorization through the owning workflow; this pointer grants none.