AI Sprint: Build Your Own AI Skills - Day 1 (Solve It Three Times Before You Build the Skill, Prompt / Skill / Project Rule, the Video-Restoration Walkthrough, and the Two-Skill Voice-Then-Deslop Pipeline)

Recap: a skill is a recipe, a set of instructions. The confusion is whether every process deserves one. His rule of thumb: a one-off task stays a prompt; something repeated more than three times within a month becomes a skill; something repeated at long intervals - quarterly planning, half-yearly reviews - becomes a project; once a month is borderline, your call. The exception that overrides the interval rule is memory: your main Claude or ChatGPT memory mixes work, curriculum design, family and 'why has my spouse stopped talking to me', so when he wants a focused, uncorrupted memory - his weekly 'self review' project, where he coaches himself because senior roles stop getting coaching - he uses a project, because project memory is separate from the global chat memory. Skills can be created inside a project (scoped to it) or outside. Everything said applies to both Codex and Claude - 'they might be one week here and there'; Gemini has not caught up to skills yet.

The undocumented rule: replicate the workflow at least three times before you make it a skill. Do the task step by step with prompts and feedback until two things are true - the goal has clarity (you know what you want the skill to do) and the output satisfies you (if not, you still owe it feedback). Only then convert. The common mistake is the reverse: 'if you want to solve a problem, you build a skill to solve the problem'. He is explicit that he did not build a video-restoration skill and then run it; he restored the video by conversation, liked the result, and only then asked for the process. Keep the skill to the common 80% and prompt the customisations rather than bloating it with if-this-then-that branches. Saving is trivial - 'save this as a skill'; naming is yours or the model's; auto-invocation is unreliable when you have many skills, so he says 'use this skill and do X'. And the best token-optimisation technique is clarity: wandering and re-brainstorming is where tokens go. Side rule: for the version-two and version-three of a skill you still need to have gone through the process yourself, even if an interview-style skill creator can draft version one.

The footage: his grandparents' 50th anniversary (1994) on a CD that grew fungus; cleaned, copied, grainy, audio damaged, with a stray Backstreet Boys recording in the last five minutes from the original cassette. The process, in a ChatGPT project on the $100 plan: (1) 'take a look at the folder, check the files, what can be done' - on GPT-6 Luna, because reading files is cheap and he wanted to stay inside the five-hour session limit; the model reported damaged audio packets, noise, colour smearing. (2) His father spotted a letter in the video; frame-finding and reconstruction got Astra extra-high - 'intense work, higher model; light work, lighter model'; the letter stayed unreadable (camera resolution). (3) The meat: compare a 20-second representative section under conservative clean-up with optional upscaling before deciding on the full set - cheap experiments, then 'apply this to the whole video'. Perplexity found GitHub restoration repos; a neural model ('RestoreX' as heard) did well but crashed a 16 GB Mac, so the final approach was balanced neighbouring-frame noise clean-up, only in-scene colour repair, no over-enhancement ('original people with noise is better than restored people whose face is distorted'), gentle 2x enlargement, straight cuts, high-quality MP4 - run overnight. (4) Trim the unrelated tail; (5) audio plan: keep the sacred chanting untouched, restore original songs where found, Tyagaraja kritis as instrumental fallback, audition clips, a written soundtrack plan. The trick on cost: don't have the AI do the work, have it write the code (OpenCV) that does the work - code running costs no tokens, model choice only changes the code's quality. Finally 'give the entire process from scratch so I can restore other videos' produced a markdown workflow (source condition, stabilisation, frame clean-up, defect repair, trim and join, music restore) -> 'build a skill out of this'.

Pet peeve: Claude and ChatGPT prose 'takes the soul out of writing', obvious in LinkedIn posts and comments. Step 1, don't start from scratch on a universal problem - search skills.sh for existing skills (no-ai-slop, a humanizer; check stars and whether it targets prose or websites), paste the repos into Codex on GPT-6.1 Sol ('building a skill is the one activity where I always use Fable or Astra - build once, use a hundred times'), and ask it to list the AI patterns each repo removes, then combine into one exhaustive list. AI slop is repetition: millions writing in the model's house style, so identify patterns and subtract. Step 2, the mega pattern skill alone produced bad output - roadblock. Step 3, realise two problems were tangled: AI writing and his writing; 'it is easier to make writing sound like me than to make it sound human'. Add voice by samples: first LinkedIn posts (Vaibhav's earlier method) - off, because LinkedIn indexes on hook-body-closure; then email and Slack through connectors, where his authentic working voice lives - better, not great. Step 4, the brainwave after asking the model why it still felt artificial: subtraction-then-addition caused a sync problem; write in his voice first, then remove AI patterns. Step 5, because the two operations kept confusing each other, split into two skills - a voice skill (trained on Slack and email, 'dilip-voice' as heard) and a no-AI-pattern skill (merged from others plus patterns he noticed) - run in sequence: voice, then de-slop. Current state: the posts he publishes, including the one about this process, are written this way and read ~80% non-AI; still work in progress (transcripts of his teaching are a suggested next voice source). If you have no samples, write ten by hand. The whole thing took two weeks of 20-30-minute tinkering sessions - 'continuous refinement is where the gold lies'; there is no framework, only 'this is not working, this is working', with AI as the workhorse and you as the judgment.

When you create a skill through the skill creator (itself a skill that appears when you ask to create skills), evals are generated automatically: the same task is run with and without the skill and an LLM-as-judge compares; a considerable difference means the skill earns its place, none means it is useless - you can ask 'run the evals and show me the output' or give guidelines for how to judge. Two kinds of skill: capability skills cover what the model cannot yet do (filling a formatted PDF) and have a short shelf life as models improve; workflow skills encode your process and last. To build either you need a thorough understanding of the process - version one can come from an interview with the model, versions two and three need you to have done it. Skills built on ChatGPT carry over to the new platforms; Codex and Claude share the same concepts. And no, skills do not replace MCP: MCP is a connector (a tool), a skill is a recipe; each has its job. Use other people's skills, but as a base - 'their skill only solves one level of a problem'.