The problem with most AI skills
A generic AI consultancy gets a two-paragraph brief, writes a prompt, ships it in 48 hours. The skill works on the example the brief described. It fails on the 40 real cases the brief didn’t mention — the matter type the associate wasn’t thinking about, the legacy file format only three people still use, the approval loop nobody documented.
Adoption dies within a month. The team goes back to doing it manually. We do the opposite.
Our method — Observe, Co-design, Ship
1. Observe
We shadow the people who will use the skill for a half-day each. Not interview — shadow. We sit with 2–3 users per function, across seniority levels, and we take notes on every decision point, every system touched, every judgment call, every shortcut.
2. Co-design
We run a one-hour co-design session with the user and the Tinkso engineer who will build the skill — together. By the end of the session, the skill spec is theirs as much as ours.
3. Ship
We build and test against 5–10 real cases from the user’s actual work in the last 60 days. Not synthetic data, not the happy path — the ones that break.
Three skills we’ve shipped — and what we heard that wasn’t in the brief
Renewal Brief Generator (legal services)
Pulls contracts from SharePoint, extracts renewal terms, cross-references account owner in Salesforce, drafts outreach email.
What the brief said: “Automate the quarterly renewal outreach — it takes associates 6 hours.”
What we heard during shadow sessions: three of the six hours weren’t writing the outreach — they were finding the last outreach. We redesigned the skill to lead with a relevant-context pull, not a blank draft.
Outcome: 6 hours → 45 minutes.
Deal Prep Packet (B2B SaaS)
Pulls account history, support tickets, product usage, decision-maker LinkedIn into a one-page brief.
What the brief said: “Automate pre-call research — reps spend 20 min per call.”
What we heard: reps weren’t struggling with the research — they were struggling with which product usage metric mattered for which account type. We built branching logic that keyed off account type.
Outcome: 20 min per call → 75 seconds.
Variance Commentary (finance)
Runs month-end variance: actuals from NetSuite, compares to budget, drafts board-ready commentary.
What the brief said: “Automate variance commentary — takes the analyst a full day.”
What we heard: the commentary wasn’t the slow part. The slow part was explaining variances the analyst didn’t understand until she’d called three operators. We added a ranked “chase list” with the operator’s name pre-attached.
Outcome: full day → 90 minutes.
Why work-first design wins on adoption
- Users adopt skills that feel native. When the output matches how the user thinks about the work, they use it.
- The skill you don’t build matters as much as the ones you do. Discovery kills bad ideas before they become tech debt.
- Co-design produces internal champions automatically.
How we build skills — the engagement
Stage 1 — Shadow (days 1–3)
Half-day observation sessions. Deliverable: Workflow Map.
Stage 2 — Co-design (day 4)
One-hour session with user + Tinkso engineer. Skill spec drafted live.
Stage 3 — Build and test (days 5–8)
Implementation, testing against real past cases, iteration with 1–2 pilot users.
Stage 4 — Ship (days 9–10)
Publish to your Claude Cowork workspace. User-facing docs. 2-minute demo video.
Pricing
| Skill complexity | Typical range |
|---|---|
| Single-system, clear inputs/outputs | Low four figures |
| Multi-system, structured output | Low-to-mid four figures |
| Complex orchestration, heavy editing logic | Mid-to-upper four figures |
Bundles: 5+ skills scoped together are discounted ~15%. Included in our standard implementation engagement. Full pricing context →
What makes our skills different
- Workflow-first design. We observe, co-design, then build.
- Typed inputs and outputs. Every skill has a schema.
- MCP-native. Skills route through our MCP connectors.
- Co-designed with users. Automatic champions by launch.
- Maintained. Included in 60-day post-launch support.
FAQ
Tell us how your team actually works.
Bring us your three most-repeated workflows. We'll shadow the people running them, co-design the skills with them, and ship.